first commit

2026-03-11 16:13:59 +08:00 · 2026-03-11 16:13:59 +08:00 · 048860986f
commit 048860986f
220 changed files with 33672 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@ -0,0 +1,73 @@
 # Python
 __pycache__/
 *.py[cod]
 *.pyo
 *.pyd
 .Python
 *.egg-info/
 dist/
 build/
 .eggs/
 # 虛擬環境
 venv/
 env/
 .env/
 # 模型權重與二進位檔
 *.pt
 *.pth
 *.onnx
 *.bie
 *.nef
 *.engine
 *.tflite
 *.pb
 # 訓練輸出（很大，不放 git）
 runs/
 work_dirs/
 # 資料集圖片（不放 git，改用 DVC 或外部儲存）
 data4/
 data50/
 test14data/
 teachabledata/
 numberocr/
 # ONNX 輸出目錄
 onnx/
 # NumPy 暫存
 npy/
 # 暫放區
 暫放區/
 # 權重目錄（除了下載腳本）
 weights/*.pt
 weights/*.pth
 # Claude Code 設定
 .claude/
 # Jupyter Notebook 輸出
 .ipynb_checkpoints/
 # 系統檔
 .DS_Store
 Thumbs.db
 desktop.ini
 # IDE
 .vscode/
 .idea/
 *.swp
 *.swo
 # 日誌
 *.log
 wandb/
 # 空的佔位檔
 python
--- a/52
+++ b/52
@ -0,0 +1,52 @@
 # Start FROM Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
 FROM nvcr.io/nvidia/pytorch:20.10-py3
 # Install dependencies
 RUN pip install --upgrade pip
 # COPY requirements.txt .
 # RUN pip install -r requirements.txt
 RUN pip install gsutil
 # Create working directory
 RUN mkdir -p /usr/src/app
 WORKDIR /usr/src/app
 # Copy contents
 COPY . /usr/src/app
 # Copy weights
 #RUN python3 -c "from models import *; \
 #attempt_download('weights/yolov5s.pt'); \
 #attempt_download('weights/yolov5m.pt'); \
 #attempt_download('weights/yolov5l.pt')"
 # ---------------------------------------------------  Extras Below  ---------------------------------------------------
 # Build and Push
 # t=ultralytics/yolov5:latest && sudo docker build -t $t . && sudo docker push $t
 # for v in {300..303}; do t=ultralytics/coco:v$v && sudo docker build -t $t . && sudo docker push $t; done
 # Pull and Run
 # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host $t
 # Pull and Run with local directory access
 # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all -v "$(pwd)"/coco:/usr/src/coco $t
 # Kill all
 # sudo docker kill $(sudo docker ps -q)
 # Kill all image-based
 # sudo docker kill $(sudo docker ps -a -q --filter ancestor=ultralytics/yolov5:latest)
 # Bash into running container
 # sudo docker container exec -it ba65811811ab bash
 # Bash into stopped container
 # sudo docker commit 092b16b25c5b usr/resume && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco --entrypoint=sh usr/resume
 # Send weights to GCP
 # python -c "from utils.general import *; strip_optimizer('runs/train/exp0_*/weights/best.pt', 'tmp.pt')" && gsutil cp tmp.pt gs://*.pt
 # Clean up
 # docker system prune -a --volumes
--- a/674
+++ b/674
@ -0,0 +1,674 @@
 GNU GENERAL PUBLIC LICENSE
                       Version 3, 29 June 2007
 Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
 Everyone is permitted to copy and distribute verbatim copies
 of this license document, but changing it is not allowed.
                            Preamble
  The GNU General Public License is a free, copyleft license for
 software and other kinds of works.
  The licenses for most software and other practical works are designed
 to take away your freedom to share and change the works.  By contrast,
 the GNU General Public License is intended to guarantee your freedom to
 share and change all versions of a program--to make sure it remains free
 software for all its users.  We, the Free Software Foundation, use the
 GNU General Public License for most of our software; it applies also to
 any other work released this way by its authors.  You can apply it to
 your programs, too.
  When we speak of free software, we are referring to freedom, not
 price.  Our General Public Licenses are designed to make sure that you
 have the freedom to distribute copies of free software (and charge for
 them if you wish), that you receive source code or can get it if you
 want it, that you can change the software or use pieces of it in new
 free programs, and that you know you can do these things.
  To protect your rights, we need to prevent others from denying you
 these rights or asking you to surrender the rights.  Therefore, you have
 certain responsibilities if you distribute copies of the software, or if
 you modify it: responsibilities to respect the freedom of others.
  For example, if you distribute copies of such a program, whether
 gratis or for a fee, you must pass on to the recipients the same
 freedoms that you received.  You must make sure that they, too, receive
 or can get the source code.  And you must show them these terms so they
 know their rights.
  Developers that use the GNU GPL protect your rights with two steps:
 (1) assert copyright on the software, and (2) offer you this License
 giving you legal permission to copy, distribute and/or modify it.
  For the developers' and authors' protection, the GPL clearly explains
 that there is no warranty for this free software.  For both users' and
 authors' sake, the GPL requires that modified versions be marked as
 changed, so that their problems will not be attributed erroneously to
 authors of previous versions.
  Some devices are designed to deny users access to install or run
 modified versions of the software inside them, although the manufacturer
 can do so.  This is fundamentally incompatible with the aim of
 protecting users' freedom to change the software.  The systematic
 pattern of such abuse occurs in the area of products for individuals to
 use, which is precisely where it is most unacceptable.  Therefore, we
 have designed this version of the GPL to prohibit the practice for those
 products.  If such problems arise substantially in other domains, we
 stand ready to extend this provision to those domains in future versions
 of the GPL, as needed to protect the freedom of users.
  Finally, every program is threatened constantly by software patents.
 States should not allow patents to restrict development and use of
 software on general-purpose computers, but in those that do, we wish to
 avoid the special danger that patents applied to a free program could
 make it effectively proprietary.  To prevent this, the GPL assures that
 patents cannot be used to render the program non-free.
  The precise terms and conditions for copying, distribution and
 modification follow.
                       TERMS AND CONDITIONS
  0. Definitions.
  "This License" refers to version 3 of the GNU General Public License.
  "Copyright" also means copyright-like laws that apply to other kinds of
 works, such as semiconductor masks.
  "The Program" refers to any copyrightable work licensed under this
 License.  Each licensee is addressed as "you".  "Licensees" and
 "recipients" may be individuals or organizations.
  To "modify" a work means to copy from or adapt all or part of the work
 in a fashion requiring copyright permission, other than the making of an
 exact copy.  The resulting work is called a "modified version" of the
 earlier work or a work "based on" the earlier work.
  A "covered work" means either the unmodified Program or a work based
 on the Program.
  To "propagate" a work means to do anything with it that, without
 permission, would make you directly or secondarily liable for
 infringement under applicable copyright law, except executing it on a
 computer or modifying a private copy.  Propagation includes copying,
 distribution (with or without modification), making available to the
 public, and in some countries other activities as well.
  To "convey" a work means any kind of propagation that enables other
 parties to make or receive copies.  Mere interaction with a user through
 a computer network, with no transfer of a copy, is not conveying.
  An interactive user interface displays "Appropriate Legal Notices"
 to the extent that it includes a convenient and prominently visible
 feature that (1) displays an appropriate copyright notice, and (2)
 tells the user that there is no warranty for the work (except to the
 extent that warranties are provided), that licensees may convey the
 work under this License, and how to view a copy of this License.  If
 the interface presents a list of user commands or options, such as a
 menu, a prominent item in the list meets this criterion.
  1. Source Code.
  The "source code" for a work means the preferred form of the work
 for making modifications to it.  "Object code" means any non-source
 form of a work.
  A "Standard Interface" means an interface that either is an official
 standard defined by a recognized standards body, or, in the case of
 interfaces specified for a particular programming language, one that
 is widely used among developers working in that language.
  The "System Libraries" of an executable work include anything, other
 than the work as a whole, that (a) is included in the normal form of
 packaging a Major Component, but which is not part of that Major
 Component, and (b) serves only to enable use of the work with that
 Major Component, or to implement a Standard Interface for which an
 implementation is available to the public in source code form.  A
 "Major Component", in this context, means a major essential component
 (kernel, window system, and so on) of the specific operating system
 (if any) on which the executable work runs, or a compiler used to
 produce the work, or an object code interpreter used to run it.
  The "Corresponding Source" for a work in object code form means all
 the source code needed to generate, install, and (for an executable
 work) run the object code and to modify the work, including scripts to
 control those activities.  However, it does not include the work's
 System Libraries, or general-purpose tools or generally available free
 programs which are used unmodified in performing those activities but
 which are not part of the work.  For example, Corresponding Source
 includes interface definition files associated with source files for
 the work, and the source code for shared libraries and dynamically
 linked subprograms that the work is specifically designed to require,
 such as by intimate data communication or control flow between those
 subprograms and other parts of the work.
  The Corresponding Source need not include anything that users
 can regenerate automatically from other parts of the Corresponding
 Source.
  The Corresponding Source for a work in source code form is that
 same work.
  2. Basic Permissions.
  All rights granted under this License are granted for the term of
 copyright on the Program, and are irrevocable provided the stated
 conditions are met.  This License explicitly affirms your unlimited
 permission to run the unmodified Program.  The output from running a
 covered work is covered by this License only if the output, given its
 content, constitutes a covered work.  This License acknowledges your
 rights of fair use or other equivalent, as provided by copyright law.
  You may make, run and propagate covered works that you do not
 convey, without conditions so long as your license otherwise remains
 in force.  You may convey covered works to others for the sole purpose
 of having them make modifications exclusively for you, or provide you
 with facilities for running those works, provided that you comply with
 the terms of this License in conveying all material for which you do
 not control copyright.  Those thus making or running the covered works
 for you must do so exclusively on your behalf, under your direction
 and control, on terms that prohibit them from making any copies of
 your copyrighted material outside their relationship with you.
  Conveying under any other circumstances is permitted solely under
 the conditions stated below.  Sublicensing is not allowed; section 10
 makes it unnecessary.
  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
  No covered work shall be deemed part of an effective technological
 measure under any applicable law fulfilling obligations under article
 11 of the WIPO copyright treaty adopted on 20 December 1996, or
 similar laws prohibiting or restricting circumvention of such
 measures.
  When you convey a covered work, you waive any legal power to forbid
 circumvention of technological measures to the extent such circumvention
 is effected by exercising rights under this License with respect to
 the covered work, and you disclaim any intention to limit operation or
 modification of the work as a means of enforcing, against the work's
 users, your or third parties' legal rights to forbid circumvention of
 technological measures.
  4. Conveying Verbatim Copies.
  You may convey verbatim copies of the Program's source code as you
 receive it, in any medium, provided that you conspicuously and
 appropriately publish on each copy an appropriate copyright notice;
 keep intact all notices stating that this License and any
 non-permissive terms added in accord with section 7 apply to the code;
 keep intact all notices of the absence of any warranty; and give all
 recipients a copy of this License along with the Program.
  You may charge any price or no price for each copy that you convey,
 and you may offer support or warranty protection for a fee.
  5. Conveying Modified Source Versions.
  You may convey a work based on the Program, or the modifications to
 produce it from the Program, in the form of source code under the
 terms of section 4, provided that you also meet all of these conditions:
    a) The work must carry prominent notices stating that you modified
    it, and giving a relevant date.
    b) The work must carry prominent notices stating that it is
    released under this License and any conditions added under section
    7.  This requirement modifies the requirement in section 4 to
    "keep intact all notices".
    c) You must license the entire work, as a whole, under this
    License to anyone who comes into possession of a copy.  This
    License will therefore apply, along with any applicable section 7
    additional terms, to the whole of the work, and all its parts,
    regardless of how they are packaged.  This License gives no
    permission to license the work in any other way, but it does not
    invalidate such permission if you have separately received it.
    d) If the work has interactive user interfaces, each must display
    Appropriate Legal Notices; however, if the Program has interactive
    interfaces that do not display Appropriate Legal Notices, your
    work need not make them do so.
  A compilation of a covered work with other separate and independent
 works, which are not by their nature extensions of the covered work,
 and which are not combined with it such as to form a larger program,
 in or on a volume of a storage or distribution medium, is called an
 "aggregate" if the compilation and its resulting copyright are not
 used to limit the access or legal rights of the compilation's users
 beyond what the individual works permit.  Inclusion of a covered work
 in an aggregate does not cause this License to apply to the other
 parts of the aggregate.
  6. Conveying Non-Source Forms.
  You may convey a covered work in object code form under the terms
 of sections 4 and 5, provided that you also convey the
 machine-readable Corresponding Source under the terms of this License,
 in one of these ways:
    a) Convey the object code in, or embodied in, a physical product
    (including a physical distribution medium), accompanied by the
    Corresponding Source fixed on a durable physical medium
    customarily used for software interchange.
    b) Convey the object code in, or embodied in, a physical product
    (including a physical distribution medium), accompanied by a
    written offer, valid for at least three years and valid for as
    long as you offer spare parts or customer support for that product
    model, to give anyone who possesses the object code either (1) a
    copy of the Corresponding Source for all the software in the
    product that is covered by this License, on a durable physical
    medium customarily used for software interchange, for a price no
    more than your reasonable cost of physically performing this
    conveying of source, or (2) access to copy the
    Corresponding Source from a network server at no charge.
    c) Convey individual copies of the object code with a copy of the
    written offer to provide the Corresponding Source.  This
    alternative is allowed only occasionally and noncommercially, and
    only if you received the object code with such an offer, in accord
    with subsection 6b.
    d) Convey the object code by offering access from a designated
    place (gratis or for a charge), and offer equivalent access to the
    Corresponding Source in the same way through the same place at no
    further charge.  You need not require recipients to copy the
    Corresponding Source along with the object code.  If the place to
    copy the object code is a network server, the Corresponding Source
    may be on a different server (operated by you or a third party)
    that supports equivalent copying facilities, provided you maintain
    clear directions next to the object code saying where to find the
    Corresponding Source.  Regardless of what server hosts the
    Corresponding Source, you remain obligated to ensure that it is
    available for as long as needed to satisfy these requirements.
    e) Convey the object code using peer-to-peer transmission, provided
    you inform other peers where the object code and Corresponding
    Source of the work are being offered to the general public at no
    charge under subsection 6d.
  A separable portion of the object code, whose source code is excluded
 from the Corresponding Source as a System Library, need not be
 included in conveying the object code work.
  A "User Product" is either (1) a "consumer product", which means any
 tangible personal property which is normally used for personal, family,
 or household purposes, or (2) anything designed or sold for incorporation
 into a dwelling.  In determining whether a product is a consumer product,
 doubtful cases shall be resolved in favor of coverage.  For a particular
 product received by a particular user, "normally used" refers to a
 typical or common use of that class of product, regardless of the status
 of the particular user or of the way in which the particular user
 actually uses, or expects or is expected to use, the product.  A product
 is a consumer product regardless of whether the product has substantial
 commercial, industrial or non-consumer uses, unless such uses represent
 the only significant mode of use of the product.
  "Installation Information" for a User Product means any methods,
 procedures, authorization keys, or other information required to install
 and execute modified versions of a covered work in that User Product from
 a modified version of its Corresponding Source.  The information must
 suffice to ensure that the continued functioning of the modified object
 code is in no case prevented or interfered with solely because
 modification has been made.
  If you convey an object code work under this section in, or with, or
 specifically for use in, a User Product, and the conveying occurs as
 part of a transaction in which the right of possession and use of the
 User Product is transferred to the recipient in perpetuity or for a
 fixed term (regardless of how the transaction is characterized), the
 Corresponding Source conveyed under this section must be accompanied
 by the Installation Information.  But this requirement does not apply
 if neither you nor any third party retains the ability to install
 modified object code on the User Product (for example, the work has
 been installed in ROM).
  The requirement to provide Installation Information does not include a
 requirement to continue to provide support service, warranty, or updates
 for a work that has been modified or installed by the recipient, or for
 the User Product in which it has been modified or installed.  Access to a
 network may be denied when the modification itself materially and
 adversely affects the operation of the network or violates the rules and
 protocols for communication across the network.
  Corresponding Source conveyed, and Installation Information provided,
 in accord with this section must be in a format that is publicly
 documented (and with an implementation available to the public in
 source code form), and must require no special password or key for
 unpacking, reading or copying.
  7. Additional Terms.
  "Additional permissions" are terms that supplement the terms of this
 License by making exceptions from one or more of its conditions.
 Additional permissions that are applicable to the entire Program shall
 be treated as though they were included in this License, to the extent
 that they are valid under applicable law.  If additional permissions
 apply only to part of the Program, that part may be used separately
 under those permissions, but the entire Program remains governed by
 this License without regard to the additional permissions.
  When you convey a copy of a covered work, you may at your option
 remove any additional permissions from that copy, or from any part of
 it.  (Additional permissions may be written to require their own
 removal in certain cases when you modify the work.)  You may place
 additional permissions on material, added by you to a covered work,
 for which you have or can give appropriate copyright permission.
  Notwithstanding any other provision of this License, for material you
 add to a covered work, you may (if authorized by the copyright holders of
 that material) supplement the terms of this License with terms:
    a) Disclaiming warranty or limiting liability differently from the
    terms of sections 15 and 16 of this License; or
    b) Requiring preservation of specified reasonable legal notices or
    author attributions in that material or in the Appropriate Legal
    Notices displayed by works containing it; or
    c) Prohibiting misrepresentation of the origin of that material, or
    requiring that modified versions of such material be marked in
    reasonable ways as different from the original version; or
    d) Limiting the use for publicity purposes of names of licensors or
    authors of the material; or
    e) Declining to grant rights under trademark law for use of some
    trade names, trademarks, or service marks; or
    f) Requiring indemnification of licensors and authors of that
    material by anyone who conveys the material (or modified versions of
    it) with contractual assumptions of liability to the recipient, for
    any liability that these contractual assumptions directly impose on
    those licensors and authors.
  All other non-permissive additional terms are considered "further
 restrictions" within the meaning of section 10.  If the Program as you
 received it, or any part of it, contains a notice stating that it is
 governed by this License along with a term that is a further
 restriction, you may remove that term.  If a license document contains
 a further restriction but permits relicensing or conveying under this
 License, you may add to a covered work material governed by the terms
 of that license document, provided that the further restriction does
 not survive such relicensing or conveying.
  If you add terms to a covered work in accord with this section, you
 must place, in the relevant source files, a statement of the
 additional terms that apply to those files, or a notice indicating
 where to find the applicable terms.
  Additional terms, permissive or non-permissive, may be stated in the
 form of a separately written license, or stated as exceptions;
 the above requirements apply either way.
  8. Termination.
  You may not propagate or modify a covered work except as expressly
 provided under this License.  Any attempt otherwise to propagate or
 modify it is void, and will automatically terminate your rights under
 this License (including any patent licenses granted under the third
 paragraph of section 11).
  However, if you cease all violation of this License, then your
 license from a particular copyright holder is reinstated (a)
 provisionally, unless and until the copyright holder explicitly and
 finally terminates your license, and (b) permanently, if the copyright
 holder fails to notify you of the violation by some reasonable means
 prior to 60 days after the cessation.
  Moreover, your license from a particular copyright holder is
 reinstated permanently if the copyright holder notifies you of the
 violation by some reasonable means, this is the first time you have
 received notice of violation of this License (for any work) from that
 copyright holder, and you cure the violation prior to 30 days after
 your receipt of the notice.
  Termination of your rights under this section does not terminate the
 licenses of parties who have received copies or rights from you under
 this License.  If your rights have been terminated and not permanently
 reinstated, you do not qualify to receive new licenses for the same
 material under section 10.
  9. Acceptance Not Required for Having Copies.
  You are not required to accept this License in order to receive or
 run a copy of the Program.  Ancillary propagation of a covered work
 occurring solely as a consequence of using peer-to-peer transmission
 to receive a copy likewise does not require acceptance.  However,
 nothing other than this License grants you permission to propagate or
 modify any covered work.  These actions infringe copyright if you do
 not accept this License.  Therefore, by modifying or propagating a
 covered work, you indicate your acceptance of this License to do so.
  10. Automatic Licensing of Downstream Recipients.
  Each time you convey a covered work, the recipient automatically
 receives a license from the original licensors, to run, modify and
 propagate that work, subject to this License.  You are not responsible
 for enforcing compliance by third parties with this License.
  An "entity transaction" is a transaction transferring control of an
 organization, or substantially all assets of one, or subdividing an
 organization, or merging organizations.  If propagation of a covered
 work results from an entity transaction, each party to that
 transaction who receives a copy of the work also receives whatever
 licenses to the work the party's predecessor in interest had or could
 give under the previous paragraph, plus a right to possession of the
 Corresponding Source of the work from the predecessor in interest, if
 the predecessor has it or can get it with reasonable efforts.
  You may not impose any further restrictions on the exercise of the
 rights granted or affirmed under this License.  For example, you may
 not impose a license fee, royalty, or other charge for exercise of
 rights granted under this License, and you may not initiate litigation
 (including a cross-claim or counterclaim in a lawsuit) alleging that
 any patent claim is infringed by making, using, selling, offering for
 sale, or importing the Program or any portion of it.
  11. Patents.
  A "contributor" is a copyright holder who authorizes use under this
 License of the Program or a work on which the Program is based.  The
 work thus licensed is called the contributor's "contributor version".
  A contributor's "essential patent claims" are all patent claims
 owned or controlled by the contributor, whether already acquired or
 hereafter acquired, that would be infringed by some manner, permitted
 by this License, of making, using, or selling its contributor version,
 but do not include claims that would be infringed only as a
 consequence of further modification of the contributor version.  For
 purposes of this definition, "control" includes the right to grant
 patent sublicenses in a manner consistent with the requirements of
 this License.
  Each contributor grants you a non-exclusive, worldwide, royalty-free
 patent license under the contributor's essential patent claims, to
 make, use, sell, offer for sale, import and otherwise run, modify and
 propagate the contents of its contributor version.
  In the following three paragraphs, a "patent license" is any express
 agreement or commitment, however denominated, not to enforce a patent
 (such as an express permission to practice a patent or covenant not to
 sue for patent infringement).  To "grant" such a patent license to a
 party means to make such an agreement or commitment not to enforce a
 patent against the party.
  If you convey a covered work, knowingly relying on a patent license,
 and the Corresponding Source of the work is not available for anyone
 to copy, free of charge and under the terms of this License, through a
 publicly available network server or other readily accessible means,
 then you must either (1) cause the Corresponding Source to be so
 available, or (2) arrange to deprive yourself of the benefit of the
 patent license for this particular work, or (3) arrange, in a manner
 consistent with the requirements of this License, to extend the patent
 license to downstream recipients.  "Knowingly relying" means you have
 actual knowledge that, but for the patent license, your conveying the
 covered work in a country, or your recipient's use of the covered work
 in a country, would infringe one or more identifiable patents in that
 country that you have reason to believe are valid.
  If, pursuant to or in connection with a single transaction or
 arrangement, you convey, or propagate by procuring conveyance of, a
 covered work, and grant a patent license to some of the parties
 receiving the covered work authorizing them to use, propagate, modify
 or convey a specific copy of the covered work, then the patent license
 you grant is automatically extended to all recipients of the covered
 work and works based on it.
  A patent license is "discriminatory" if it does not include within
 the scope of its coverage, prohibits the exercise of, or is
 conditioned on the non-exercise of one or more of the rights that are
 specifically granted under this License.  You may not convey a covered
 work if you are a party to an arrangement with a third party that is
 in the business of distributing software, under which you make payment
 to the third party based on the extent of your activity of conveying
 the work, and under which the third party grants, to any of the
 parties who would receive the covered work from you, a discriminatory
 patent license (a) in connection with copies of the covered work
 conveyed by you (or copies made from those copies), or (b) primarily
 for and in connection with specific products or compilations that
 contain the covered work, unless you entered into that arrangement,
 or that patent license was granted, prior to 28 March 2007.
  Nothing in this License shall be construed as excluding or limiting
 any implied license or other defenses to infringement that may
 otherwise be available to you under applicable patent law.
  12. No Surrender of Others' Freedom.
  If conditions are imposed on you (whether by court order, agreement or
 otherwise) that contradict the conditions of this License, they do not
 excuse you from the conditions of this License.  If you cannot convey a
 covered work so as to satisfy simultaneously your obligations under this
 License and any other pertinent obligations, then as a consequence you may
 not convey it at all.  For example, if you agree to terms that obligate you
 to collect a royalty for further conveying from those to whom you convey
 the Program, the only way you could satisfy both those terms and this
 License would be to refrain entirely from conveying the Program.
  13. Use with the GNU Affero General Public License.
  Notwithstanding any other provision of this License, you have
 permission to link or combine any covered work with a work licensed
 under version 3 of the GNU Affero General Public License into a single
 combined work, and to convey the resulting work.  The terms of this
 License will continue to apply to the part which is the covered work,
 but the special requirements of the GNU Affero General Public License,
 section 13, concerning interaction through a network will apply to the
 combination as such.
  14. Revised Versions of this License.
  The Free Software Foundation may publish revised and/or new versions of
 the GNU General Public License from time to time.  Such new versions will
 be similar in spirit to the present version, but may differ in detail to
 address new problems or concerns.
  Each version is given a distinguishing version number.  If the
 Program specifies that a certain numbered version of the GNU General
 Public License "or any later version" applies to it, you have the
 option of following the terms and conditions either of that numbered
 version or of any later version published by the Free Software
 Foundation.  If the Program does not specify a version number of the
 GNU General Public License, you may choose any version ever published
 by the Free Software Foundation.
  If the Program specifies that a proxy can decide which future
 versions of the GNU General Public License can be used, that proxy's
 public statement of acceptance of a version permanently authorizes you
 to choose that version for the Program.
  Later license versions may give you additional or different
 permissions.  However, no additional obligations are imposed on any
 author or copyright holder as a result of your choosing to follow a
 later version.
  15. Disclaimer of Warranty.
  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
 APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
 HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
 OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
 THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
 IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
 ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
  16. Limitation of Liability.
  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
 WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
 THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
 GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
 USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
 DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
 PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
 EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
 SUCH DAMAGES.
  17. Interpretation of Sections 15 and 16.
  If the disclaimer of warranty and limitation of liability provided
 above cannot be given local legal effect according to their terms,
 reviewing courts shall apply local law that most closely approximates
 an absolute waiver of all civil liability in connection with the
 Program, unless a warranty or assumption of liability accompanies a
 copy of the Program in return for a fee.
                     END OF TERMS AND CONDITIONS
            How to Apply These Terms to Your New Programs
  If you develop a new program, and you want it to be of the greatest
 possible use to the public, the best way to achieve this is to make it
 free software which everyone can redistribute and change under these terms.
  To do so, attach the following notices to the program.  It is safest
 to attach them to the start of each source file to most effectively
 state the exclusion of warranty; and each file should have at least
 the "copyright" line and a pointer to where the full notice is found.
    <one line to give the program's name and a brief idea of what it does.>
    Copyright (C) <year>  <name of author>
    This program is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation, either version 3 of the License, or
    (at your option) any later version.
    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.
    You should have received a copy of the GNU General Public License
    along with this program.  If not, see <http://www.gnu.org/licenses/>.
 Also add information on how to contact you by electronic and paper mail.
  If the program does terminal interaction, make it output a short
 notice like this when it starts in an interactive mode:
    <program>  Copyright (C) <year>  <name of author>
    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
    This is free software, and you are welcome to redistribute it
    under certain conditions; type `show c' for details.
 The hypothetical commands `show w' and `show c' should show the appropriate
 parts of the General Public License.  Of course, your program's commands
 might be different; for a GUI interface, you would use an "about box".
  You should also get your employer (if you work as a programmer) or school,
 if any, to sign a "copyright disclaimer" for the program, if necessary.
 For more information on this, and how to apply and follow the GNU GPL, see
 <http://www.gnu.org/licenses/>.
  The GNU General Public License does not permit incorporating your program
 into proprietary programs.  If your program is a subroutine library, you
 may consider it more useful to permit linking proprietary applications with
 the library.  If this is what you want to do, use the GNU Lesser General
 Public License instead of this License.  But first, please read
 <http://www.gnu.org/philosophy/why-not-lgpl.html>.
--- a/README.md
+++ b/README.md
@ -0,0 +1,270 @@
 <h1 align="center">  Object Detection </h1>
 Object Detection task with YOLOv5 model.
 This document contains the explanations of arguments of each script.
 You can find the tutorial document for finetuning a pretrained model on COCO128 dataset under the `tutorial` folder, `tutorial/README.md`. 
 The ipython notebook tutorial is also prepared under the `tutorial` folder as `tutorial/tutorial.ipynb`. You may upload and run this ipython notebook on Google colab.
 # Prerequisites
 - Python 3.8 or above
 # Installation
 ```bash
 $ pip install -U pip
 $ pip install -r requirements.txt
 ```
 # Dataset & Preparation
 The image data, annotations and dataset.yaml are required. 
 ## MS COCO
 Our traning script accepts MS COCO dataset. You may download the dataset using the following link:
 - Download [2017 MS COCO Dataset](https://cocodataset.org/#download) 
 ## Custom Datasets
 You can also train the model on a custom dataset. 
 ### Annotations Format
 After using a tool like [CVAT](https://github.com/openvinotoolkit/cvat), [makesense.ai](https://www.makesense.ai) or [Labelbox](https://labelbox.com) to label your images, export your labels to YOLO format, with one `*.txt` file per image (if no objects in image, no `*.txt` file is required). The `*.txt` file specifications are:
 - One row per object
 - Each row is `class x_center y_center width height` format.
 - Box coordinates must be in normalized xywh format (from 0 - 1). If your boxes are in pixels, divide `x_center` and `width` by image `width`, and `y_center` and `height` by image height.
 - Class numbers are zero-indexed (start from 0).
 <div align="center">
 <img src="./tutorial/screenshots/readme_img.jpg" width="50%" /> 
 </div>
 The label file corresponding to the above image contains 2 persons (class 0) and a tie (class 27):
 <div align="center">
 <img src="./tutorial/screenshots/readme_img2.png" width="40%" /> 
 </div>
 ###  Directory Organization
 Your own datasets are expected to have the following structure. We assume `/dataset` is next to the `/yolov5` directory. YOLOv5 locates labels automatically for each image by replacing the last instance of `/images/` in each image path with `/labels/`.
 ```bash
 - Dataset name
    -- images
        -- train
            --- img001.jpg
            --- ...
        -- val
            --- img002.jpg
            --- ...
    -- labels
        -- train
            --- img001.txt
            --- ...
        -- val
            --- img002.txt
            --- ...
 - yolov5
 - generate_npy
 - exporting
 ```
 ###  dataset.yaml
 The yaml file for COCO dataset has been prepared in `./data/coco.yaml`. For custom dataset, you need to prepare the yaml file and save it under `./data/`. The yaml file is expected to have the following format:
 ```bash
 # train and val datasets (image directory or *.txt file with image paths)  
 train: ./datasets/images/train/  
 val: ./datasets/images/val/  
 # number of classes  
 nc: number of classes  
 # class names  
 names: list of class names
 ```
 # Train
 For training on MS COCO, execute commands in the folder `yolov5`:
 ```shell
 CUDA_VISIBLE_DEVICES='0' python train.py --data coco.yaml --cfg yolov5s-noupsample.yaml --weights '' --batch-size 64 
 ```
 `CUDA_VISIBLE_DEVICES='0'` indicates the gpu ids.
 `--data` the yaml file. (located under `./data/`)
 `--cfg` the model configuration. (located under `./model/`) (`yolov5s-noupsample.yaml` for 520, `yolov5s.yaml` for 720)
 `--hyp` the path to hyperparameters file. (located under `./data/`)
 `--weights` the path to pretained model weights. ('' if train from scratch)
 `--epochs` the number of epochs to train. (Default: 300)
 `--batch-size` batch size. (Default: 16)
 `--img-size` the input size of the model. (Default: (640, 640))
 `--workers` the maximum number of dataloader workers. (Default: 8)
 By default, the trained models are saved under `./runs/train/`.
 ## Generating .npy for different model input
 We can generating `.npy` for different model input by using `yolov5_generate_npy.py`. Execute commands in the folder `generate_npy`:
 ```shell
 python yolov5_generate_npy.py --input-h 640 --input-w 640 
 ```
 `--input-h` the input height. (Default: 640)
 `--input-w` the input width. (Default: 640)
 We could get `*.npy`
 # Configure the paths yaml file
 You are expected to create a yaml file which stores all the paths related to the trained models. This yaml file will be used in the following sections. You can check and modify the `pretrained_paths_520.yaml` and `pretrained_paths_720.yaml` under `/yolov5/data/`. The yaml file is expected to contain the following information:
 ```shell
 grid_dir: path_to_npy_file_directory
 grid20_path: path_to_grid20_npy_file
 grid40_path: path_to_grid40_npy_file
 grid80_path: path_to_grid80_npy_file
 yolov5_dir: path_to_yolov5_directory
 path: path_to_pretrained_yolov5_model_weights_pt_file
 yaml_path: path_to_the_model_configuration_yaml_file
 pt_path: path_to_export_yolov5_model_weights_kneron_supported_file
 onnx_export_file: path_to_export_yolov5_onnx_model_file
 input_w: model_input_weight
 input_h: model_input_height
 nc: number_of_classes
 names: list_of_class_names
 ```
 # Save and Convert to ONNX
 This section will introduce how to save the trained model for pytorch1.4 supported format and convert to ONNX. 
 ## Exporting ONNX model in the PyTorch 1.7 environment
 We can convert the model to onnx by using `yolov5_export.py`. Execute commands in the folder `yolov5`:
 ```shell
 python ../exporting/yolov5_export.py --data path_to_pretrained_path_yaml_file
 ```
 `--data` the path to pretrained model paths yaml file (Default: ../yolov5/data/pretrained_paths_520.yaml)
 We could get onnx model. 
 ## Converting onnx by tool chain
 Pull the latest [ONNX converter](https://github.com/kneron/ONNX_Convertor/tree/master/optimizer_scripts) from github. You may read the latest document from Github for converting ONNX model. Execute commands in the folder `ONNX_Convertor/optimizer_scripts`:
 (reference: https://github.com/kneron/ONNX_Convertor/tree/master/optimizer_scripts)
 ```shell
 python -m onnxsim input_onnx_model output_onnx_model
 python pytorch2onnx.py input.pth output.onnx
 ```
 We could get converted onnx model.
 # Inference
 Before model inference, we assume that the model has been converted to onnx model as in the previous section (even if only inference pth model). Create a yaml file containing the path information. For model inference on a single image, execute commands in the folder `yolov5`:
 ```shell
 python inference.py --data path_to_pretrained_path_yaml_file --img-path path_to_image --save-path path_to_saved_image
 ```
 `--img-path` the path to the image.
 `--save-path` the path to draw and save the image with bbox.
 `--data` the path to pretrained model paths yaml file. (Default: data/pretrained_paths_520.yaml)
 `--conf_thres` the score threshold of bounding boxes. (Default: 0.3)
 `--iou_thres` the iou threshold for NMS. (Default: 0.3) 
 `--onnx` whether is onnx model inference. 
 You could find preprocessing and postprocessing processes under the folder `exporting/yolov5/`. 
 # Evaluation 
 ## Evaluation Metric
 We will use mean Average Precision (mAP) for evaluation. You can find the script for computing mAP in `test.py`.
 `mAP`: mAP is the average of Average Precision (AP). AP summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold, with the increase in recall from the previous threshold used as the weight:
 <img src="https://latex.codecogs.com/svg.image?AP&space;=&space;\sum_n&space;(R_n-R_{n-1})P_n&space;" title="AP = \sum_n (R_n-R_{n-1})P_n " />
 where <img src="https://latex.codecogs.com/svg.image?R_n" title="R_n" />  and <img src="https://latex.codecogs.com/svg.image?P_n" title="P_n" /> are the precision and recall at the nth threshold. The mAP compares the ground-truth bounding box to the detected box and returns a score. The higher the score, the more accurate the model is in its detections.
 ## Evaluation on a Dataset
 For evaluating the trained model on dataset:
 ```shell
 python test.py --weights path_to_pth_model_weight --data path_to_data_yaml_file
 ```
 `--weights` The path to pretrained model weight. (Defalut: best.pt)
 `--data` The path to data yaml file. (Default: data/coco128.yaml)
 `--img-size` Input shape of the model (Default: (640, 640))
 `--conf-thres` Object confidence threshold. (Default: 0.001)
 `--device` Cuda device, i.e. 0 or 0,1,2,3 or cpu. (Default: cpu)
 `--verbose` Whether report mAP by class.
 ## End-to-End Evaluation
 If you would like to perform an end-to-end test with an image dataset, you can use `inference_e2e.py` under the directory `yolov5` to obtain the prediction results.
 You have to prepare an initial parameter yaml file for the inference runner. You may check `utils/init_params.yaml` for the format.
 ```shell
 python inference_e2e.py --img-path path_to_dataset_folder --params path_to_init_params_file --save-path path_to_save_json_file
 ```
 `--img-path` Path to the dataset directory
 `--params` Path to initial parameter yaml file for the inference runner
 `--save-path` Path to save the prediction to a json file
 `--gpu` GPU id  (-1 if cpu) (Default: -1)
 The predictions will be saved into a json file that has the following structure:
 ```bash
 [
    {'img_path':image_path_1
    'bbox': [[l,t,w,h,score,class_id], [l,t,w,h,score,class_id]]
    },
    {'img_path':image_path_2
    'bbox': [[l,t,w,h,score,class_id], [l,t,w,h,score,class_id]]
    },
    ...
 ]
 ```
 # Model
 Backbone | Input Size |  FPS on 520 | FPS on 720  | Model Size | mAP
 --- | --- |:---:|:---:|:---:|:---:
 [YOLOv5s (no upsample)](https://github.com/kneron/Model_Zoo/tree/main/detection/yolov5/yolov5s-noupsample) | 640x640 | 4.91429 | - | 13.1M | 40.4%
 [YOLOv5s (with upsample)](https://github.com/kneron/Model_Zoo/tree/main/detection/yolov5/yolov5s) | 640x640 | - | 24.4114 | 14.6M | 50.9%
 [YOLOv5s (no upsample)](https://github.com/kneron/Model_Zoo/tree/main/detection/yolov5/yolov5s-noupsample) is the yolov5s model backbone without upsampling operation, since 520 hardware does not support upsampling operation.
--- a/data/annapretrained_paths_520.yaml
+++ b/data/annapretrained_paths_520.yaml
@ -0,0 +1,17 @@
 grid_dir: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/
 grid20_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/20_640x640.npy
 grid40_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/40_640x640.npy
 grid80_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/80_640x640.npy
 yolov5_dir: C:/Users/rd_de/kneronyolov5/yolov5
 path: C:/Users/rd_de/kneronyolov5/yolov5/runs/train/exp59/weights/best.pt
 yaml_path: C:/Users/rd_de/kneronyolov5/yolov5/models/yolov5s-noupsample.yaml
 pt_path: C:/Users/rd_de/kneronyolov5/yolov5/weights/for520best.pt
 onnx_export_file: C:/Users/rd_de/kneronyolov5/yolov5/runs/train/exp59/weights/best.onnx
 input_w: 640
 input_h: 640
 nc: 1
 names: ['License Plate']
--- a/data/coco.yaml
+++ b/data/coco.yaml
@ -0,0 +1,33 @@
 # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
 path: ../datasets/coco  # dataset root dir
 train: /home/ziyan/Dataset/COCO/coco/images/train2017/  # 118287 images
 val: /home/ziyan/Dataset/COCO/coco/images/val2017/  # 5000 images
 # number of classes
 nc: 80
 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']
 # Download script/URL (optional)
 download: |
  from utils.general import download, Path
  # Download labels
  segments = False  # segment or box labels
  dir = Path(yaml['path'])  # dataset root dir
  url = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/'
  urls = [url + ('coco2017labels-segments.zip' if segments else 'coco2017labels.zip')]  # labels
  download(urls, dir=dir.parent)
  # Download data
  urls = ['http://images.cocodataset.org/zips/train2017.zip',  # 19G, 118k images
          'http://images.cocodataset.org/zips/val2017.zip',  # 1G, 5k images
          'http://images.cocodataset.org/zips/test2017.zip']  # 7G, 41k images (optional)
  download(urls, dir=dir / 'images', threads=3)
--- a/data/coco128.yaml
+++ b/data/coco128.yaml
@ -0,0 +1,28 @@
 # COCO 2017 dataset http://cocodataset.org - first 128 training images
 # Train command: python train.py --data coco128.yaml
 # Default dataset location is next to /yolov5:
 #   /parent_folder
 #     /coco128
 #     /yolov5
 # download command/URL (optional)
 download: https://github.com/ultralytics/yolov5/releases/download/v1.0/coco128.zip
 # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
 train: ../coco128/images/train2017/  # 128 images
 val: ../coco128/images/train2017/  # 128 images
 # number of classes
 nc: 80
 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']
--- a/data/custom.yaml
+++ b/data/custom.yaml
@ -0,0 +1,9 @@
 # train and val data as directory: path/labels/
 train: ../image_data/images/train/ 
 val: ../image_data/images/val/  
 # number of classes
 nc: 3
 # class names
 names: ['air conditioner', 'dog', 'fence']
--- a/data/hyp.finetune.yaml
+++ b/data/hyp.finetune.yaml
@ -0,0 +1,38 @@
 # Hyperparameters for VOC finetuning
 # python train.py --batch 64 --weights yolov5m.pt --data voc.yaml --img 512 --epochs 50
 # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials
 # Hyperparameter Evolution Results
 # Generations: 306
 #                   P         R     mAP.5 mAP.5:.95       box       obj       cls
 # Metrics:        0.6     0.936     0.896     0.684    0.0115   0.00805   0.00146
 lr0: 0.0032
 lrf: 0.12
 momentum: 0.843
 weight_decay: 0.00036
 warmup_epochs: 2.0
 warmup_momentum: 0.5
 warmup_bias_lr: 0.05
 box: 0.0296
 cls: 0.243
 cls_pw: 0.631
 obj: 0.301
 obj_pw: 0.911
 iou_t: 0.2
 anchor_t: 2.91
 # anchors: 3.63
 fl_gamma: 0.0
 hsv_h: 0.0138
 hsv_s: 0.664
 hsv_v: 0.464
 degrees: 0.373
 translate: 0.245
 scale: 0.898
 shear: 0.602
 perspective: 0.0
 flipud: 0.00856
 fliplr: 0.5
 mosaic: 1.0 #0.0
 mixup: 0.243
--- a/data/hyp.scratch.yaml
+++ b/data/hyp.scratch.yaml
@ -0,0 +1,33 @@
 # Hyperparameters for COCO training from scratch
 # python train.py --batch 40 --cfg yolov5m.yaml --weights '' --data coco.yaml --img 640 --epochs 300
 # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials
 lr0: 0.01  # initial learning rate (SGD=1E-2, Adam=1E-3)
 lrf: 0.2  # final OneCycleLR learning rate (lr0 * lrf)
 momentum: 0.937  # SGD momentum/Adam beta1
 weight_decay: 0.0005  # optimizer weight decay 5e-4
 warmup_epochs: 3.0  # warmup epochs (fractions ok)
 warmup_momentum: 0.8  # warmup initial momentum
 warmup_bias_lr: 0.1  # warmup initial bias lr
 box: 0.05  # box loss gain
 cls: 0.5  # cls loss gain
 cls_pw: 1.0  # cls BCELoss positive_weight
 obj: 1.0  # obj loss gain (scale with pixels)
 obj_pw: 1.0  # obj BCELoss positive_weight
 iou_t: 0.20  # IoU training threshold
 anchor_t: 4.0  # anchor-multiple threshold
 # anchors: 3  # anchors per output layer (0 to ignore)
 fl_gamma: 0.0  # focal loss gamma (efficientDet default gamma=1.5)
 hsv_h: 0.015  # image HSV-Hue augmentation (fraction)
 hsv_s: 0.7  # image HSV-Saturation augmentation (fraction)
 hsv_v: 0.4  # image HSV-Value augmentation (fraction)
 degrees: 0.0  # image rotation (+/- deg)
 translate: 0.1  # image translation (+/- fraction)
 scale: 0.5  # image scale (+/- gain)
 shear: 0.0  # image shear (+/- deg)
 perspective: 0.0  # image perspective (+/- fraction), range 0-0.001
 flipud: 0.0  # image flip up-down (probability)
 fliplr: 0.5  # image flip left-right (probability)
 mosaic: 1.0  # image mosaic (probability)
 mixup: 0.0  # image mixup (probability)
--- a/data/mepretrained_paths_520.yaml
+++ b/data/mepretrained_paths_520.yaml
@ -0,0 +1,21 @@
 grid_dir: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/
 grid20_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/20_640x640.npy
 grid40_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/40_640x640.npy
 grid80_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/80_640x640.npy
 yolov5_dir: C:/Users/rd_de/kneronyolov5/yolov5
 path: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp24/weights/best.pt
 yaml_path: C:/Users/rd_de/golfaceyolov5/yolov5/models/yolov5s-noupsample.yaml
 pt_path: C:/Users/rd_de/golfaceyolov5/yolov5/weights/for520best.pt
 onnx_export_file: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp24/weights/best.onnx
 input_w: 640
 input_h: 640
 nc: 4
 names: ['100', '1000', '50', '500']
 #nc: 6
 #names: ['Break circuit', 'bulge', 'foreign object', 'scratch', 'short circuit', 'white spot']
--- a/data/mepretrained_paths_630.yaml
+++ b/data/mepretrained_paths_630.yaml
@ -0,0 +1,17 @@
 grid_dir: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/
 grid20_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/20_640x640.npy
 grid40_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/40_640x640.npy
 grid80_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/80_640x640.npy
 yolov5_dir: C:/Users/rd_de/kneronyolov5/yolov5
 path: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp16/weights/best.pt
 yaml_path: C:/Users/rd_de/golfaceyolov5/yolov5/models/yolov5s.yaml
 pt_path: C:/Users/rd_de/golfaceyolov5/yolov5/weights/for720best.pt
 onnx_export_file: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp16/weights/best.onnx
 input_w: 640
 input_h: 640
 nc: 4
 names: ['car', 'greenery', 'person', 'tree']
--- a/data/mepretrained_paths_630class2.yaml
+++ b/data/mepretrained_paths_630class2.yaml
@ -0,0 +1,17 @@
 grid_dir: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/
 grid20_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/20_640x640.npy
 grid40_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/40_640x640.npy
 grid80_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/80_640x640.npy
 yolov5_dir: C:/Users/rd_de/kneronyolov5/yolov5
 path: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp13/weights/best.pt
 yaml_path: C:/Users/rd_de/golfaceyolov5/yolov5/models/yolov5s.yaml
 pt_path: C:/Users/rd_de/golfaceyolov5/yolov5/weights/for720best.pt
 onnx_export_file: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp13/weights/best.onnx
 input_w: 640
 input_h: 640
 nc: 2
 names: ['bunker', 'pond']
--- a/data/mepretrained_paths_630class8.yaml
+++ b/data/mepretrained_paths_630class8.yaml
@ -0,0 +1,16 @@
 grid_dir: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/
 grid20_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/20_640x640.npy
 grid40_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/40_640x640.npy
 grid80_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/80_640x640.npy
 yolov5_dir: C:/Users/rd_de/kneronyolov5/yolov5
 path: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp29/weights/best.pt
 yaml_path: C:/Users/rd_de/golfaceyolov5/yolov5/models/yolov5s.yaml
 pt_path: C:/Users/rd_de/golfaceyolov5/yolov5/weights/for720best.pt
 onnx_export_file: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp29/weights/best.onnx
 nc: 8
 names: ['bunker', 'car', 'grass', 'greenery', 'person', 'pond', 'road', 'tree']
 input_w: 640
 input_h: 640
--- a/data/mepretrained_paths_720.yaml
+++ b/data/mepretrained_paths_720.yaml
@ -0,0 +1,17 @@
 grid_dir: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/
 grid20_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/20_640x640.npy
 grid40_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/40_640x640.npy
 grid80_path: C:/Users/rd_de/kneronyolov5/ai_training/detection/yolov5/generate_npy/80_640x640.npy
 yolov5_dir: C:/Users/rd_de/kneronyolov5/yolov5
 path: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp19/weights/best.pt
 yaml_path: C:/Users/rd_de/golfaceyolov5/yolov5/models/yolov5s.yaml
 pt_path: C:/Users/rd_de/golfaceyolov5/yolov5/weights/for720best.pt
 onnx_export_file: C:/Users/rd_de/golfaceyolov5/yolov5/runs/train/exp19/weights/best.onnx
 input_w: 640
 input_h: 640
 nc: 4
 names: ['car', 'greenery', 'person', 'tree']
--- a/data/model_paths_520_coco128.yaml
+++ b/data/model_paths_520_coco128.yaml
@ -0,0 +1,25 @@
 grid_dir: ../generate_npy/
 grid20_path: ../generate_npy/20_640x640.npy
 grid40_path: ../generate_npy/40_640x640.npy
 grid80_path: ../generate_npy/80_640x640.npy
 yolov5_dir: ./
 path: ./runs/train/exp/weights/best.pt
 yaml_path: ./models/yolov5s-noupsample.yaml
 pt_path: ./yolov5s-noupsample-coco128.pt   # pytorch 1.4
 onnx_export_file: ./yolov5s-noupsample-coco128.onnx
 input_w: 640
 input_h: 640
 # number of classes
 nc: 80
 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']
--- a/data/pretrained_paths_520.yaml
+++ b/data/pretrained_paths_520.yaml
@ -0,0 +1,25 @@
 grid_dir: ../generate_npy/
 grid20_path: ../generate_npy/20_640x640.npy
 grid40_path: ../generate_npy/40_640x640.npy
 grid80_path: ../generate_npy/80_640x640.npy
 yolov5_dir: ./
 path: ./best.pt
 yaml_path: ./models/yolov5s-noupsample.yaml
 pt_path: ./yolov5s-noupsample.pt   # pytorch 1.4
 onnx_export_file: ./yolov5s-noupsample.onnx
 input_w: 640
 input_h: 640
 # number of classes
 nc: 80
 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']
--- a/data/pretrained_paths_720.yaml
+++ b/data/pretrained_paths_720.yaml
@ -0,0 +1,25 @@
 grid_dir: ../generate_npy/
 grid20_path: ../generate_npy/20_640x640.npy
 grid40_path: ../generate_npy/40_640x640.npy
 grid80_path: ../generate_npy/80_640x640.npy
 yolov5_dir: ./
 path: ./best.pt
 yaml_path: ./models/yolov5s.yaml
 pt_path: ./yolov5s.pt   # pytorch 1.4
 onnx_export_file: ./yolov5s.onnx
 input_w: 640
 input_h: 640
 # number of classes
 nc: 80
 # class names
 names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light',
        'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
        'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
        'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard',
        'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',
        'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
        'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
        'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear',
        'hair drier', 'toothbrush']
--- a/data/scripts/get_coco.sh
+++ b/data/scripts/get_coco.sh
@ -0,0 +1,24 @@
 #!/bin/bash
 # COCO 2017 dataset http://cocodataset.org
 # Download command: bash data/scripts/get_coco.sh
 # Train command: python train.py --data coco.yaml
 # Default dataset location is next to /yolov5:
 #   /parent_folder
 #     /coco
 #     /yolov5
 # Download/unzip labels
 d='../' # unzip directory
 url=https://github.com/ultralytics/yolov5/releases/download/v1.0/
 f='coco2017labels.zip'                                                                 # 68 MB
 echo 'Downloading' $url$f ' ...' && curl -L $url$f -o $f && unzip -q $f -d $d && rm $f # download, unzip, remove
 # Download/unzip images
 d='../coco/images' # unzip directory
 url=http://images.cocodataset.org/zips/
 f1='train2017.zip' # 19G, 118k images
 f2='val2017.zip'   # 1G, 5k images
 f3='test2017.zip'  # 7G, 41k images (optional)
 for f in $f1 $f2; do
  echo 'Downloading' $url$f ' ...' && curl -L $url$f -o $f && unzip -q $f -d $d && rm $f # download, unzip, remove
 done
--- a/data/scripts/get_voc.sh
+++ b/data/scripts/get_voc.sh
@ -0,0 +1,137 @@
 #!/bin/bash
 # PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC/
 # Download command: bash data/scripts/get_voc.sh
 # Train command: python train.py --data voc.yaml
 # Default dataset location is next to /yolov5:
 #   /parent_folder
 #     /VOC
 #     /yolov5
 start=$(date +%s)
 mkdir -p ../tmp
 cd ../tmp/
 # Download/unzip images and labels
 d='.' # unzip directory
 url=https://github.com/ultralytics/yolov5/releases/download/v1.0/
 f1=VOCtrainval_06-Nov-2007.zip # 446MB, 5012 images
 f2=VOCtest_06-Nov-2007.zip     # 438MB, 4953 images
 f3=VOCtrainval_11-May-2012.zip # 1.95GB, 17126 images
 for f in $f1 $f2 $f3; do
  echo 'Downloading' $url$f ' ...' && curl -L $url$f -o $f && unzip -q $f -d $d && rm $f # download, unzip, remove
 done
 end=$(date +%s)
 runtime=$((end - start))
 echo "Completed in" $runtime "seconds"
 echo "Splitting dataset..."
 python3 - "$@" <<END
 import xml.etree.ElementTree as ET
 import pickle
 import os
 from os import listdir, getcwd
 from os.path import join
 sets=[('2012', 'train'), ('2012', 'val'), ('2007', 'train'), ('2007', 'val'), ('2007', 'test')]
 classes = ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]
 def convert(size, box):
    dw = 1./(size[0])
    dh = 1./(size[1])
    x = (box[0] + box[1])/2.0 - 1
    y = (box[2] + box[3])/2.0 - 1
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x*dw
    w = w*dw
    y = y*dh
    h = h*dh
    return (x,y,w,h)
 def convert_annotation(year, image_id):
    in_file = open('VOCdevkit/VOC%s/Annotations/%s.xml'%(year, image_id))
    out_file = open('VOCdevkit/VOC%s/labels/%s.txt'%(year, image_id), 'w')
    tree=ET.parse(in_file)
    root = tree.getroot()
    size = root.find('size')
    w = int(size.find('width').text)
    h = int(size.find('height').text)
    for obj in root.iter('object'):
        difficult = obj.find('difficult').text
        cls = obj.find('name').text
        if cls not in classes or int(difficult)==1:
            continue
        cls_id = classes.index(cls)
        xmlbox = obj.find('bndbox')
        b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
        bb = convert((w,h), b)
        out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
 wd = getcwd()
 for year, image_set in sets:
    if not os.path.exists('VOCdevkit/VOC%s/labels/'%(year)):
        os.makedirs('VOCdevkit/VOC%s/labels/'%(year))
    image_ids = open('VOCdevkit/VOC%s/ImageSets/Main/%s.txt'%(year, image_set)).read().strip().split()
    list_file = open('%s_%s.txt'%(year, image_set), 'w')
    for image_id in image_ids:
        list_file.write('%s/VOCdevkit/VOC%s/JPEGImages/%s.jpg\n'%(wd, year, image_id))
        convert_annotation(year, image_id)
    list_file.close()
 END
 cat 2007_train.txt 2007_val.txt 2012_train.txt 2012_val.txt >train.txt
 cat 2007_train.txt 2007_val.txt 2007_test.txt 2012_train.txt 2012_val.txt >train.all.txt
 python3 - "$@" <<END
 import shutil
 import os
 os.system('mkdir ../VOC/')
 os.system('mkdir ../VOC/images')
 os.system('mkdir ../VOC/images/train')
 os.system('mkdir ../VOC/images/val')
 os.system('mkdir ../VOC/labels')
 os.system('mkdir ../VOC/labels/train')
 os.system('mkdir ../VOC/labels/val')
 import os
 print(os.path.exists('../tmp/train.txt'))
 f = open('../tmp/train.txt', 'r')
 lines = f.readlines()
 for line in lines:
    line = "/".join(line.split('/')[-5:]).strip()
    if (os.path.exists("../" + line)):
        os.system("cp ../"+ line + " ../VOC/images/train")
    line = line.replace('JPEGImages', 'labels')
    line = line.replace('jpg', 'txt')
    if (os.path.exists("../" + line)):
        os.system("cp ../"+ line + " ../VOC/labels/train")
 print(os.path.exists('../tmp/2007_test.txt'))
 f = open('../tmp/2007_test.txt', 'r')
 lines = f.readlines()
 for line in lines:
    line = "/".join(line.split('/')[-5:]).strip()
    if (os.path.exists("../" + line)):
        os.system("cp ../"+ line + " ../VOC/images/val")
    line = line.replace('JPEGImages', 'labels')
    line = line.replace('jpg', 'txt')
    if (os.path.exists("../" + line)):
        os.system("cp ../"+ line + " ../VOC/labels/val")
 END
 rm -rf ../tmp # remove temporary directory
 echo "VOC download done."
--- a/datasets/test/images/765706588_040624_mp4-0001_jpg.rf.e5e135b768bfdd6fc7672c6f72939863.jpg
+++ b/datasets/test/images/765706588_040624_mp4-0001_jpg.rf.e5e135b768bfdd6fc7672c6f72939863.jpg
--- a/datasets/test/images/frame_00124_jpg.rf.a7b6f14f61fe01b4978fadaaa0408002.jpg
+++ b/datasets/test/images/frame_00124_jpg.rf.a7b6f14f61fe01b4978fadaaa0408002.jpg
--- a/datasets/test/images/frame_00235_jpg.rf.4247e072bd0c90c04ca47f90af72087c.jpg
+++ b/datasets/test/images/frame_00235_jpg.rf.4247e072bd0c90c04ca47f90af72087c.jpg
--- a/datasets/test/images/frame_00237_jpg.rf.9c6a3860b1e0af420c128366dc45f729.jpg
+++ b/datasets/test/images/frame_00237_jpg.rf.9c6a3860b1e0af420c128366dc45f729.jpg
--- a/datasets/test/images/frame_01255_jpg.rf.8a40b37a20ad9853f0fc94596c1ea3da.jpg
+++ b/datasets/test/images/frame_01255_jpg.rf.8a40b37a20ad9853f0fc94596c1ea3da.jpg
--- a/datasets/test/images/frame_01263_jpg.rf.3335d5da5dd3f10aabaf65f8c1dbb3cb.jpg
+++ b/datasets/test/images/frame_01263_jpg.rf.3335d5da5dd3f10aabaf65f8c1dbb3cb.jpg
--- a/datasets/test/images/frame_01270_jpg.rf.cc45cf417809fcf171fb970ed14c1c4c.jpg
+++ b/datasets/test/images/frame_01270_jpg.rf.cc45cf417809fcf171fb970ed14c1c4c.jpg
--- a/datasets/test/images/frame_01399_jpg.rf.fb73708d42132080535669f0bb37e876.jpg
+++ b/datasets/test/images/frame_01399_jpg.rf.fb73708d42132080535669f0bb37e876.jpg
--- a/datasets/test/images/images_0013_jpg.rf.868fc3ea1dc01309dd13f328c984b257.jpg
+++ b/datasets/test/images/images_0013_jpg.rf.868fc3ea1dc01309dd13f328c984b257.jpg
--- a/datasets/test/images/images_0017_jpg.rf.39ec10af4161c51c195cb10a0c9db28e.jpg
+++ b/datasets/test/images/images_0017_jpg.rf.39ec10af4161c51c195cb10a0c9db28e.jpg
--- a/datasets/test/labels/765706588_040624_mp4-0001_jpg.rf.e5e135b768bfdd6fc7672c6f72939863.txt
+++ b/datasets/test/labels/765706588_040624_mp4-0001_jpg.rf.e5e135b768bfdd6fc7672c6f72939863.txt
@ -0,0 +1,3 @@
 0 0.3703125 0.66875 0.1203125 0.0109375
 1 0.6375 0.73046875 0.15625 0.14140625
 1 0.41875 0.76875 0.28828125 0.0765625
--- a/datasets/test/labels/frame_00124_jpg.rf.a7b6f14f61fe01b4978fadaaa0408002.txt
+++ b/datasets/test/labels/frame_00124_jpg.rf.a7b6f14f61fe01b4978fadaaa0408002.txt
@ -0,0 +1,2 @@
 0 0.771875 0.50859375 0.03203125 0.0703125
 0 0.71484375 0.46171875 0.05 0.0515625
--- a/datasets/test/labels/frame_00235_jpg.rf.4247e072bd0c90c04ca47f90af72087c.txt
+++ b/datasets/test/labels/frame_00235_jpg.rf.4247e072bd0c90c04ca47f90af72087c.txt
@ -0,0 +1,3 @@
 0 0.02265625 0.95078125 0.0453125 0.0953125
 0 0.09375 0.7875 0.0328125 0.07421875
 0 0.15703125 0.68046875 0.04921875 0.090625
--- a/datasets/test/labels/frame_00237_jpg.rf.9c6a3860b1e0af420c128366dc45f729.txt
+++ b/datasets/test/labels/frame_00237_jpg.rf.9c6a3860b1e0af420c128366dc45f729.txt
@ -0,0 +1,2 @@
 0 0.0203125 0.92109375 0.040625 0.1578125
 0 0.1046875 0.73046875 0.0609375 0.07421875
--- a/datasets/test/labels/frame_01255_jpg.rf.8a40b37a20ad9853f0fc94596c1ea3da.txt
+++ b/datasets/test/labels/frame_01255_jpg.rf.8a40b37a20ad9853f0fc94596c1ea3da.txt
@ -0,0 +1,2 @@
 1 0.590625 0.2671875 0.0546875 0.03125
 1 0.6546875 0.28359375 0.06875 0.03046875
--- a/datasets/test/labels/frame_01263_jpg.rf.3335d5da5dd3f10aabaf65f8c1dbb3cb.txt
+++ b/datasets/test/labels/frame_01263_jpg.rf.3335d5da5dd3f10aabaf65f8c1dbb3cb.txt
@ -0,0 +1,2 @@
 1 0.67265625 0.31640625 0.1453125 0.09765625
 1 0.81484375 0.38359375 0.12421875 0.11015625
--- a/datasets/test/labels/frame_01270_jpg.rf.cc45cf417809fcf171fb970ed14c1c4c.txt
+++ b/datasets/test/labels/frame_01270_jpg.rf.cc45cf417809fcf171fb970ed14c1c4c.txt
@ -0,0 +1,3 @@
 1 0.8921875 0.42734375 0.0265625 0.03671875
 1 0.94609375 0.47578125 0.02890625 0.05703125
 1 0.98671875 0.53046875 0.025 0.05546875
--- a/datasets/test/labels/frame_01399_jpg.rf.fb73708d42132080535669f0bb37e876.txt
+++ b/datasets/test/labels/frame_01399_jpg.rf.fb73708d42132080535669f0bb37e876.txt
@ -0,0 +1,3 @@
 1 0.80078125 0.45234375 0.115625 0.07890625
 1 0.93671875 0.53359375 0.1265625 0.1671875
 1 0.678125 0.41171875 0.03125 0.0265625
--- a/datasets/test/labels/images_0013_jpg.rf.868fc3ea1dc01309dd13f328c984b257.txt
+++ b/datasets/test/labels/images_0013_jpg.rf.868fc3ea1dc01309dd13f328c984b257.txt
@ -0,0 +1 @@
 1 0.6609375 0.4421875 0.03046875 0.02265625
--- a/datasets/test/labels/images_0017_jpg.rf.39ec10af4161c51c195cb10a0c9db28e.txt
+++ b/datasets/test/labels/images_0017_jpg.rf.39ec10af4161c51c195cb10a0c9db28e.txt
@ -0,0 +1 @@
 1 0.5234375 0.390625 0.0328125 0.01640625
--- a/detect.py
+++ b/detect.py
@ -0,0 +1,172 @@
 import argparse
 import time
 from pathlib import Path
 import cv2
 import torch
 import torch.backends.cudnn as cudnn
 from numpy import random
 from models.experimental import attempt_load
 from utils.datasets import LoadStreams, LoadImages
 from utils.general import check_img_size, non_max_suppression, apply_classifier, scale_coords, xyxy2xywh, \
    strip_optimizer, set_logging, increment_path
 from utils.plots import plot_one_box
 from utils.torch_utils import select_device, load_classifier, time_synchronized
 def detect(save_img=False):
    source, weights, view_img, save_txt, imgsz = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size
    webcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith(
        ('rtsp://', 'rtmp://', 'http://'))
    # Directories
    save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))  # increment run
    (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
    # Initialize
    set_logging()
    device = select_device(opt.device)
    half = device.type != 'cpu'  # half precision only supported on CUDA
    # Load model
    model = attempt_load(weights, map_location=device)  # load FP32 model
    imgsz = check_img_size(imgsz, s=model.stride.max())  # check img_size
    if half:
        model.half()  # to FP16
    # Second-stage classifier
    classify = False
    if classify:
        modelc = load_classifier(name='resnet101', n=2)  # initialize
        modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model']).to(device).eval()
    # Set Dataloader
    vid_path, vid_writer = None, None
    if webcam:
        view_img = True
        cudnn.benchmark = True  # set True to speed up constant image size inference
        dataset = LoadStreams(source, img_size=imgsz)
    else:
        save_img = True
        dataset = LoadImages(source, img_size=imgsz)
    # Get names and colors
    names = model.module.names if hasattr(model, 'module') else model.names
    colors = [[random.randint(0, 255) for _ in range(3)] for _ in names]
    # Run inference
    t0 = time.time()
    img = torch.zeros((1, 3, imgsz, imgsz), device=device)  # init img
    _ = model(img.half() if half else img) if device.type != 'cpu' else None  # run once
    for path, img, im0s, vid_cap in dataset:
        img = torch.from_numpy(img).to(device)
        img = img.half() if half else img.float()  # uint8 to fp16/32
        img /= 255.0  # 0 - 255 to 0.0 - 1.0
        if img.ndimension() == 3:
            img = img.unsqueeze(0)
        # Inference
        t1 = time_synchronized()
        pred = model(img, augment=opt.augment)[0]
        # Apply NMS
        pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms)
        t2 = time_synchronized()
        # Apply Classifier
        if classify:
            pred = apply_classifier(pred, modelc, img, im0s)
        # Process detections
        for i, det in enumerate(pred):  # detections per image
            if webcam:  # batch_size >= 1
                p, s, im0 = Path(path[i]), '%g: ' % i, im0s[i].copy()
            else:
                p, s, im0 = Path(path), '', im0s
            save_path = str(save_dir / p.name)
            txt_path = str(save_dir / 'labels' / p.stem) + ('_%g' % dataset.frame if dataset.mode == 'video' else '')
            s += '%gx%g ' % img.shape[2:]  # print string
            gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwh
            if len(det):
                # Rescale boxes from img_size to im0 size
                det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
                # Print results
                for c in det[:, -1].unique():
                    n = (det[:, -1] == c).sum()  # detections per class
                    s += '%g %ss, ' % (n, names[int(c)])  # add to string
                # Write results
                for *xyxy, conf, cls in reversed(det):
                    if save_txt:  # Write to file
                        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
                        line = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh)  # label format
                        with open(txt_path + '.txt', 'a') as f:
                            f.write(('%g ' * len(line)).rstrip() % line + '\n')
                    if save_img or view_img:  # Add bbox to image
                        label = '%s %.2f' % (names[int(cls)], conf)
                        plot_one_box(xyxy, im0, label=label, color=colors[int(cls)], line_thickness=3)
            # Print time (inference + NMS)
            print('%sDone. (%.3fs)' % (s, t2 - t1))
            # Stream results
            if view_img:
                cv2.imshow(p, im0)
                if cv2.waitKey(1) == ord('q'):  # q to quit
                    raise StopIteration
            # Save results (image with detections)
            if save_img:
                if dataset.mode == 'images':
                    cv2.imwrite(save_path, im0)
                else:
                    if vid_path != save_path:  # new video
                        vid_path = save_path
                        if isinstance(vid_writer, cv2.VideoWriter):
                            vid_writer.release()  # release previous video writer
                        fourcc = 'mp4v'  # output video codec
                        fps = vid_cap.get(cv2.CAP_PROP_FPS)
                        w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
                        h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
                        vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*fourcc), fps, (w, h))
                    vid_writer.write(im0)
    if save_txt or save_img:
        print('Results saved to %s' % save_dir)
    print('Done. (%.3fs)' % (time.time() - t0))
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', nargs='+', type=str, default='yolov5s.pt', help='model.pt path(s)')
    parser.add_argument('--source', type=str, default='data/images', help='source')  # file/folder, 0 for webcam
    parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)')
    parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold')
    parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--view-img', action='store_true', help='display results')
    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3')
    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
    parser.add_argument('--augment', action='store_true', help='augmented inference')
    parser.add_argument('--update', action='store_true', help='update all models')
    parser.add_argument('--project', default='runs/detect', help='save results to project/name')
    parser.add_argument('--name', default='exp', help='save results to project/name')
    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
    opt = parser.parse_args()
    print(opt)
    with torch.no_grad():
        if opt.update:  # update all models (to fix SourceChangeWarning)
            for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:
                detect()
                strip_optimizer(opt.weights)
        else:
            detect()
--- a/export.py
+++ b/export.py
@ -0,0 +1,94 @@
 """Exports a YOLOv5 *.pt model to ONNX and TorchScript formats
 Usage:
    $ export PYTHONPATH="$PWD" && python models/export.py --weights ./weights/yolov5s.pt --img 640 --batch 1
 """
 import argparse
 import sys
 import time
 sys.path.append('./')  # to run '$ python *.py' files in subdirectories
 import torch
 import torch.nn as nn
 import models
 from models.experimental import attempt_load
 from utils.activations import Hardswish
 from utils.general import set_logging, check_img_size
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path')  # from yolov5/models/
    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size')  # height, width
    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
    opt = parser.parse_args()
    opt.img_size *= 2 if len(opt.img_size) == 1 else 1  # expand
    print(opt)
    set_logging()
    t = time.time()
    # Load PyTorch model
    model = attempt_load(opt.weights, map_location=torch.device('cpu'))  # load FP32 model
    labels = model.names
    # Checks
    gs = int(max(model.stride))  # grid size (max stride)
    opt.img_size = [check_img_size(x, gs) for x in opt.img_size]  # verify img_size are gs-multiples
    # Input
    img = torch.zeros(opt.batch_size, 3, *opt.img_size)  # image size(1,3,320,192) iDetection
    # Update model
    for k, m in model.named_modules():
        m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
        if isinstance(m, models.common.Conv) and isinstance(m.act, nn.Hardswish):
            m.act = Hardswish()  # assign activation
        # if isinstance(m, models.yolo.Detect):
        #     m.forward = m.forward_export  # assign forward (optional)
    model.model[-1].export = True  # set Detect() layer export=True
    y = model(img)  # dry run
    # TorchScript export
    try:
        print('\nStarting TorchScript export with torch %s...' % torch.__version__)
        f = opt.weights.replace('.pt', '.torchscript.pt')  # filename
        ts = torch.jit.trace(model, img)
        ts.save(f)
        print('TorchScript export success, saved as %s' % f)
    except Exception as e:
        print('TorchScript export failure: %s' % e)
    # ONNX export
    try:
        import onnx
        print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
        f = opt.weights.replace('.pt', '.onnx')  # filename
        torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'],
                          output_names=['classes', 'boxes'] if y is None else ['output'])
        # Checks
        onnx_model = onnx.load(f)  # load onnx model
        onnx.checker.check_model(onnx_model)  # check onnx model
        # print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
        print('ONNX export success, saved as %s' % f)
    except Exception as e:
        print('ONNX export failure: %s' % e)
    # CoreML export
    try:
        import coremltools as ct
        print('\nStarting CoreML export with coremltools %s...' % ct.__version__)
        # convert model from torchscript and apply pixel scaling as per detect.py
        model = ct.convert(ts, inputs=[ct.ImageType(name='image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])])
        f = opt.weights.replace('.pt', '.mlmodel')  # filename
        model.save(f)
        print('CoreML export success, saved as %s' % f)
    except Exception as e:
        print('CoreML export failure: %s' % e)
    # Finish
    print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t))
--- a/hubconf.py
+++ b/hubconf.py
@ -0,0 +1,119 @@
 """File for accessing YOLOv5 via PyTorch Hub https://pytorch.org/hub/
 Usage:
    import torch
    model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True, channels=3, classes=80)
 """
 from pathlib import Path
 import torch
 from models.yolo import Model
 from utils.general import set_logging
 from utils.google_utils import attempt_download
 dependencies = ['torch', 'yaml']
 set_logging()
 def create(name, pretrained, channels, classes):
    """Creates a specified YOLOv5 model
    Arguments:
        name (str): name of model, i.e. 'yolov5s'
        pretrained (bool): load pretrained weights into the model
        channels (int): number of input channels
        classes (int): number of model classes
    Returns:
        pytorch model
    """
    config = Path(__file__).parent / 'models' / f'{name}.yaml'  # model.yaml path
    try:
        model = Model(config, channels, classes)
        if pretrained:
            fname = f'{name}.pt'  # checkpoint filename
            attempt_download(fname)  # download if not found locally
            ckpt = torch.load(fname, map_location=torch.device('cpu'))  # load
            state_dict = ckpt['model'].float().state_dict()  # to FP32
            state_dict = {k: v for k, v in state_dict.items() if model.state_dict()[k].shape == v.shape}  # filter
            model.load_state_dict(state_dict, strict=False)  # load
            if len(ckpt['model'].names) == classes:
                model.names = ckpt['model'].names  # set class names attribute
            # model = model.autoshape()  # for PIL/cv2/np inputs and NMS
        return model
    except Exception as e:
        help_url = 'https://github.com/ultralytics/yolov5/issues/36'
        s = 'Cache maybe be out of date, try force_reload=True. See %s for help.' % help_url
        raise Exception(s) from e
 def yolov5s(pretrained=False, channels=3, classes=80):
    """YOLOv5-small model from https://github.com/ultralytics/yolov5
    Arguments:
        pretrained (bool): load pretrained weights into the model, default=False
        channels (int): number of input channels, default=3
        classes (int): number of model classes, default=80
    Returns:
        pytorch model
    """
    return create('yolov5s', pretrained, channels, classes)
 def yolov5m(pretrained=False, channels=3, classes=80):
    """YOLOv5-medium model from https://github.com/ultralytics/yolov5
    Arguments:
        pretrained (bool): load pretrained weights into the model, default=False
        channels (int): number of input channels, default=3
        classes (int): number of model classes, default=80
    Returns:
        pytorch model
    """
    return create('yolov5m', pretrained, channels, classes)
 def yolov5l(pretrained=False, channels=3, classes=80):
    """YOLOv5-large model from https://github.com/ultralytics/yolov5
    Arguments:
        pretrained (bool): load pretrained weights into the model, default=False
        channels (int): number of input channels, default=3
        classes (int): number of model classes, default=80
    Returns:
        pytorch model
    """
    return create('yolov5l', pretrained, channels, classes)
 def yolov5x(pretrained=False, channels=3, classes=80):
    """YOLOv5-xlarge model from https://github.com/ultralytics/yolov5
    Arguments:
        pretrained (bool): load pretrained weights into the model, default=False
        channels (int): number of input channels, default=3
        classes (int): number of model classes, default=80
    Returns:
        pytorch model
    """
    return create('yolov5x', pretrained, channels, classes)
 if __name__ == '__main__':
    model = create(name='yolov5s', pretrained=True, channels=3, classes=80)  # example
    model = model.fuse().autoshape()  # for PIL/cv2/np inputs and NMS
    # Verify inference
    from PIL import Image
    imgs = [Image.open(x) for x in Path('data/images').glob('*.jpg')]
    results = model(imgs)
    results.show()
    results.print()
--- a/kneron/bie2nef.py
+++ b/kneron/bie2nef.py
@ -0,0 +1,44 @@
 import ktc
 import os
 import shutil
 import subprocess
 # 設定 `.bie` 和 `.nef` 的存放目錄
 onnx_dir = "runs/train/exp24/weights/"
 bie_file = os.path.join(onnx_dir, "input.kdp720.scaled.bie")  # 確保 `.bie` 路徑正確
 # 確保 `.bie` 檔案存在
 if not os.path.exists(bie_file):
    raise FileNotFoundError(f"❌ Error: BIE file not found at {bie_file}")
 print(f"✅ Found BIE file: {bie_file}")
 # 初始化 ModelConfig
 km = ktc.ModelConfig(20008, "0001", "720", bie_path=bie_file)
 # 執行 `.nef` 轉換
 nef_model_path = ktc.compile([km])
 # 打印出 `.nef` 生成的路徑
 print(f"🔍 Generated NEF file at: {nef_model_path}")
 # 確保 `.nef` 轉換成功
 if not nef_model_path or not os.path.exists(nef_model_path):
    raise RuntimeError(f"❌ Error: NEF model was not generated at {nef_model_path}")
 # 確保目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 移動 `.nef` 到指定資料夾
 nef_save_path = os.path.join(onnx_dir, os.path.basename(nef_model_path))
 shutil.copy(nef_model_path, nef_save_path)
 # **立即檢查 `.nef` 是否真的存在**
 if os.path.exists(nef_save_path):
    print(f"\n✅ NEF file successfully saved to: {nef_save_path}")
 else:
    raise RuntimeError(f"❌ Error: NEF file NOT found in {nef_save_path} after copying!")
 # **執行 `ls` 指令來確認 Python 內部真的看到 `.nef`**
 print("\n🔍 Listing files in target directory:")
 subprocess.run(["ls", "-lh", onnx_dir])
--- a/kneron/exporting/yolov5/init.py
+++ b/kneron/exporting/yolov5/init.py
--- a/kneron/exporting/yolov5/common.py
+++ b/kneron/exporting/yolov5/common.py
@ -0,0 +1,224 @@
 # This file contains modules common to various models
 import torch.nn as nn
 import torch
 import torch.nn.functional as F
 def autopad(k, p=None):  # kernel, padding
    # Pad to 'same'
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
    return p
 def DWConv(c1, c2, k=1, s=1, act=True):
    # Depthwise convolution
    return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
 class Conv(nn.Module):
    # Standard convolution
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Conv, self).__init__()
        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity()
    def forward(self, x):
        return self.act(self.bn(self.conv(x)))
    def fuseforward(self, x):
        return self.act(self.conv(x))
 class Bottleneck(nn.Module):
    # Standard bottleneck
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
        super(Bottleneck, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2
    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
 class BottleneckCSP(nn.Module):
    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super(BottleneckCSP, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
        self.cv4 = Conv(2 * c_, c2, 1, 1)
        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
        self.act = nn.LeakyReLU(0.1, inplace=True)
        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
    def forward(self, x):
        y1 = self.cv3(self.m(self.cv1(x)))
        y2 = self.cv2(x)
        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
 class SPP(nn.Module):
    # Spatial pyramid pooling layer used in YOLOv3-SPP
    def __init__(self, c1, c2, k=(5, 9, 13)):
        super(SPP, self).__init__()
        c_ = c1 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
    def forward(self, x):
        x = self.cv1(x)
        return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
 class Focus(nn.Module):#
    # Focus wh information into c-space
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Focus, self).__init__()
        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
        w1_1 = torch.tensor([[[1., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w1_2 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[1., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w1_3 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[1., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w3_1 = torch.tensor([[[0., 1., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w3_2 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 1., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w3_3 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 1., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w2_1 = torch.tensor([[[0., 0., 0.],[1., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w2_2 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[1., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w2_3 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[1., 0., 0.],[0., 0., 0.]]])
        w4_1 = torch.tensor([[[0., 0., 0.],[0., 1., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w4_2 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 1., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]]])
        w4_3 = torch.tensor([[[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 0., 0.],[0., 0., 0.]], [[0., 0., 0.],[0., 1., 0.],[0., 0., 0.]]])
        w1_1 = w1_1.view(1, 3, 3, 3)
        w1_2 = w1_2.view(1, 3, 3, 3)
        w1_3 = w1_3.view(1, 3, 3, 3)
        w2_1 = w2_1.view(1, 3, 3, 3)
        w2_2 = w2_2.view(1, 3, 3, 3)
        w2_3 = w2_3.view(1, 3, 3, 3)
        w3_1 = w3_1.view(1, 3, 3, 3)
        w3_2 = w3_2.view(1, 3, 3, 3)
        w3_3 = w3_3.view(1, 3, 3, 3)
        w4_1 = w4_1.view(1, 3, 3, 3)
        w4_2 = w4_2.view(1, 3, 3, 3)
        w4_3 = w4_3.view(1, 3, 3, 3)    
        self.w_cat = torch.cat([w1_1, w1_2,w1_3, w2_1,w2_2,w2_3, w3_1,w3_2,w3_3, w4_1,w4_2,w4_3], 0) 
        self.p2d = (0, 2, 0, 2)
    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)      
        # x = x.type(torch.cuda.FloatTensor)
        #x_gt = self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
        x_pad = F.pad(x, self.p2d, 'constant', 0)
        xx = F.conv2d(x_pad, self.w_cat.to(x.device),stride=2) 
        xx = self.conv(xx)
        #print(torch.sum(x_gt - xx))
        return xx
 class Focus_ori(nn.Module):#
    # Focus wh information into c-space
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Focus, self).__init__()
        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
        x = x.type(torch.cuda.FloatTensor)
        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
 class Concat(nn.Module):
    # Concatenate a list of tensors along dimension
    def __init__(self, dimension=1):
        super(Concat, self).__init__()
        self.d = dimension
    def forward(self, x):
        return torch.cat(x, self.d)
 class Flatten(nn.Module):
    # Use after nn.AdaptiveAvgPool2d(1) to remove last 2 dimensions
    @staticmethod
    def forward(x):
        return x.view(x.size(0), -1)
 class Classify(nn.Module):
    # Classification head, i.e. x(b,c1,20,20) to x(b,c2)
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Classify, self).__init__()
        self.aap = nn.AdaptiveAvgPool2d(1)  # to x(b,c1,1,1)
        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)  # to x(b,c2,1,1)
        self.flat = Flatten()
    def forward(self, x):
        z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1)  # cat if list
        return self.flat(self.conv(z))  # flatten to x(b,c2)
 class MixConv2d(nn.Module):
    # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595
    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
        super(MixConv2d, self).__init__()
        groups = len(k)
        if equal_ch:  # equal c_ per group
            i = torch.linspace(0, groups - 1E-6, c2).floor()  # c2 indices
            c_ = [(i == g).sum() for g in range(groups)]  # intermediate channels
        else:  # equal weight.numel() per group
            b = [c2] + [0] * groups
            a = np.eye(groups + 1, groups, k=-1)
            a -= np.roll(a, 1, axis=1)
            a *= np.array(k) ** 2
            a[0] = 1
            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b
        self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.LeakyReLU(0.1, inplace=True)
    def forward(self, x):
        return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
 class CrossConv(nn.Module):
    # Cross Convolution Downsample
    def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
        # ch_in, ch_out, kernel, stride, groups, expansion, shortcut
        super(CrossConv, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, (1, k), (1, s))
        self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
        self.add = shortcut and c1 == c2
    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
 # class C3(nn.Module):
 #     # Cross Convolution CSP
 #     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
 #         super(C3, self).__init__()
 #         c_ = int(c2 * e)  # hidden channels
 #         self.cv1 = Conv(c1, c_, 1, 1)
 #         self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
 #         self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
 #         self.cv4 = Conv(2 * c_, c2, 1, 1)
 #         self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
 #         self.act = nn.LeakyReLU(0.1, inplace=True)
 #         self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
 #     def forward(self, x):
 #         y1 = self.cv3(self.m(self.cv1(x)))
 #         y2 = self.cv2(x)
 #         return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
 class C3(nn.Module):
    # CSP Bottleneck with 3 convolutions
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super(C3, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c1, c_, 1, 1)
        self.cv3 = Conv(2 * c_, c2, 1)  # act=FReLU(c2)
        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
    def forward(self, x):
        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))
--- a/kneron/exporting/yolov5/kneron_preprocessing/API.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/API.py
@ -0,0 +1,684 @@
 # -*- coding: utf-8 -*-
 import numpy as np
 import os
 from .funcs.utils import str2int, str2bool
 from . import Flow
 flow = Flow()
 flow.set_numerical_type('floating')
 flow_520 = Flow()
 flow_520.set_numerical_type('520')
 flow_720 = Flow()
 flow_720.set_numerical_type('720')
 DEFAULT = None
 default = {
    'crop':{
        'align_w_to_4':False
        },
    'resize':{
        'type':'bilinear',
        'calculate_ratio_using_CSim':False
        }
 }
 def set_default_as_520():
    """
    Set some default parameter as 520 setting
    crop.align_w_to_4 = True
    crop.pad_square_to_4 = True
    resize.type = 'fixed_520'
    resize.calculate_ratio_using_CSim = True
    """
    global default
    default['crop']['align_w_to_4'] = True
    default['resize']['type'] = 'fixed_520'
    default['resize']['calculate_ratio_using_CSim'] = True
    return
 def set_default_as_floating():
    """
    Set some default parameter as floating setting
    crop.align_w_to_4 = False
    crop.pad_square_to_4 = False
    resize.type = 'bilinear'
    resize.calculate_ratio_using_CSim = False
    """
    global default
    default['crop']['align_w_to_4'] = False
    default['resize']['type'] = 'bilinear'
    default['resize']['calculate_ratio_using_CSim'] = False
    pass
 def print_info_on():
    """
    turn print infomation on.
    """
    flow.set_print_info(True)
    flow_520.set_print_info(True)
 def print_info_off():
    """
    turn print infomation off.
    """
    flow.set_print_info(False)
    flow_520.set_print_info(False)
 def load_image(image):
    """
    load_image function
    load load_image and output as rgb888 format np.array
    Args:
        image: [np.array/str], can be np.array or image file path
    Returns:
        out: [np.array], rgb888 format
    Examples:
    """
    image = flow.load_image(image, is_raw = False)
    return image
 def load_bin(image, fmt=None, size=None):
    """
    load_bin function
    load bin file and output as rgb888 format np.array
    Args:
        image: [str], bin file path
        fmt: [str], "rgb888" / "rgb565" / "nir"
        size: [tuble], (image_w, image_h)
    Returns:
        out: [np.array], rgb888 format
    Examples:
        >>> image_data = kneron_preprocessing.API.load_bin(image,'rgb565',(raw_w,raw_h))
    """    
    assert isinstance(size, tuple)
    assert isinstance(fmt, str)
    # assert (fmt.lower() in ['rgb888', "rgb565" , "nir",'RGB888', "RGB565" , "NIR", 'NIR888', 'nir888'])
    image = flow.load_image(image, is_raw = True, raw_img_type='bin', raw_img_fmt = fmt, img_in_width = size[0], img_in_height = size[1])
    flow.set_color_conversion(source_format=fmt, out_format = 'rgb888')
    image,_ = flow.funcs['color'](image)
    return image
 def load_hex(file, fmt=None, size=None):
    """
    load_hex function
    load hex file and output as rgb888 format np.array
    Args:
        image: [str], hex file path
        fmt: [str], "rgb888" / "yuv444" / "ycbcr444" / "yuv422" / "ycbcr422" / "rgb565"
        size: [tuble], (image_w, image_h)
    Returns:
        out: [np.array], rgb888 format
    Examples:
        >>> image_data = kneron_preprocessing.API.load_hex(image,'rgb565',(raw_w,raw_h))
    """  
    assert isinstance(size, tuple)
    assert isinstance(fmt, str)
    assert (fmt.lower() in ['rgb888',"yuv444" , "ycbcr444" , "yuv422" , "ycbcr422" , "rgb565"])
    image = flow.load_image(file, is_raw = True, raw_img_type='hex', raw_img_fmt = fmt, img_in_width = size[0], img_in_height = size[1])
    flow.set_color_conversion(source_format=fmt, out_format = 'rgb888')
    image,_ = flow.funcs['color'](image)
    return image
 def dump_image(image, output=None, file_fmt='txt',image_fmt='rgb888',order=0):
    """
    dump_image function
    dump txt, bin or hex, default is txt
    image format as following format: RGB888, RGBA8888, RGB565, NIR, YUV444, YCbCr444, YUV422, YCbCr422, default is RGB888
    Args:
        image: [np.array/str], can be np.array or image file path
        output: [str], dump file path
        file_fmt: [str], "bin" / "txt" / "hex", set dump file format, default is txt
        image_fmt: [str], RGB888 / RGBA8888 / RGB565 / NIR / YUV444 / YCbCr444 / YUV422 / YCbCr422, default is RGB888
    Examples:
        >>> kneron_preprocessing.API.dump_image(image_data,out_path,fmt='bin')
    """
    if isinstance(image, str):
        image = load_image(image)
    assert isinstance(image, np.ndarray)
    if output is None:
        return
    flow.set_output_setting(is_dump=False, dump_format=file_fmt, image_format=image_fmt ,output_file=output)
    flow.dump_image(image)
    return
 def convert(image, out_fmt = 'RGB888', source_fmt = 'RGB888'):
    """
    color convert
    Args:
        image: [np.array], input
        out_fmt: [str], "rgb888" / "rgba8888" / "rgb565" / "yuv" / "ycbcr" / "yuv422" / "ycbcr422"
        source_fmt: [str], "rgb888" / "rgba8888" / "rgb565" / "yuv" / "ycbcr" / "yuv422" / "ycbcr422"
    Returns:
        out: [np.array]
    Examples:
    """  
    flow.set_color_conversion(source_format = source_fmt, out_format=out_fmt, simulation=False)
    image,_ = flow.funcs['color'](image)
    return image
 def get_crop_range(box,align_w_to_4=DEFAULT, pad_square_to_4=False,rounding_type=0):
    """
    get exact crop box according different setting
    Args:
        box: [tuble], (x1, y1, x2, y2)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
    Returns:
        out: [tuble,4], (crop_x1, crop_y1, crop_x2, crop_y2) 
    Examples:
        >>> image_data = kneron_preprocessing.API.get_crop_range((272,145,461,341), align_w_to_4=True, pad_square_to_4=True)
        (272, 145, 460, 341)
    """  
    if box is None:
        return (0,0,0,0)
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='specific', start_x=box[0],start_y=box[1],end_x=box[2],end_y=box[3], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4,rounding_type=rounding_type)
    image = np.zeros((1,1,3)).astype('uint8')
    _,info = flow.funcs['crop'](image)
    return info['box']
 def crop(image, box=None, align_w_to_4=DEFAULT, pad_square_to_4=False,rounding_type=0 ,info_out = {}):
    """
    crop function
    specific crop range by box
    Args:
        image: [np.array], input
        box: [tuble], (x1, y1, x2, y2)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
        info_out: [dic], save the final crop box into info_out['box']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop(image_data,(272,145,461,341), align_w_to_4=True, info_out=info)
        >>> info['box']
        (272, 145, 460, 341)
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop(image_data,(272,145,461,341), pad_square_to_4=True, info_out=info)
        >>> info['box']
        (268, 145, 464, 341)
    """  
    assert isinstance(image, np.ndarray)
    if box is None:
        return image
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='specific', start_x=box[0],start_y=box[1],end_x=box[2],end_y=box[3], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4,rounding_type=rounding_type)
    image,info = flow.funcs['crop'](image)
    info_out['box'] = info['box']
    return image
 def crop_center(image, range=None, align_w_to_4=DEFAULT, pad_square_to_4=False,rounding_type=0 ,info_out = {}):
    """
    crop function
    center crop by range
    Args:
        image: [np.array], input
        range: [tuble], (crop_w, crop_h)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
        info_out: [dic], save the final crop box into info_out['box']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_center(image_data,(102,40), align_w_to_4=True,info_out=info)
        >>> info['box']
        (268, 220, 372, 260)
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_center(image_data,(102,40), pad_square_to_4=True, info_out=info)
        >>> info['box']
        (269, 192, 371, 294)
    """   
    assert isinstance(image, np.ndarray)
    if range is None:
        return image
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='center', crop_w=range[0],crop_h=range[1], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4,rounding_type=rounding_type)
    image,info = flow.funcs['crop'](image)
    info_out['box'] = info['box']
    return image
 def crop_corner(image, range=None, align_w_to_4=DEFAULT,pad_square_to_4=False,rounding_type=0 ,info_out = {}):
    """
    crop function
    corner crop by range
    Args:
        image: [np.array], input
        range: [tuble], (crop_w, crop_h)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
        info_out: [dic], save the final crop box into info_out['box']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_corner(image_data,(102,40), align_w_to_4=True,info_out=info)
        >>> info['box']
        (0, 0, 104, 40)
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_corner(image_data,(102,40), pad_square_to_4=True,info_out=info)
        >>> info['box']
        (0, -28, 102, 74)
    """
    assert isinstance(image, np.ndarray)
    if range is None:
        return image
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='corner', crop_w=range[0],crop_h=range[1], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4)
    image, info = flow.funcs['crop'](image)
    info_out['box'] = info['box']
    return image
 def resize(image, size=None, keep_ratio = True, zoom = True, type=DEFAULT, calculate_ratio_using_CSim = DEFAULT, info_out = {}):
    """
    resize function
    resize type can be bilinear or bilicubic as floating type, fixed or fixed_520/fixed_720 as fixed type.
    fixed_520/fixed_720 type has add some function to simulate 520/720 bug.
    Args:
        image: [np.array], input
        size: [tuble], (input_w, input_h)
        keep_ratio: [bool], keep_ratio or not, default True
        zoom: [bool], enable resize can zoom image or not, default True
        type: [str], "bilinear" / "bilicubic" / "cv2" / "fixed" / "fixed_520" / "fixed_720"
        calculate_ratio_using_CSim: [bool], calculate the ratio and scale using Csim function and C float, default False
        info_out: [dic], save the final scale size(w,h) into info_out['size']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.resize(image_data,size=(56,56),type='fixed',info_out=info)
        >>> info_out['size']
        (54,56)
    """
    assert isinstance(image, np.ndarray)
    if size is None:
        return image
    if type is None:
        type = default['resize']['type']
    if calculate_ratio_using_CSim is None:
        calculate_ratio_using_CSim = default['resize']['calculate_ratio_using_CSim']
    flow.set_resize(resize_w = size[0], resize_h = size[1], type=type, keep_ratio=keep_ratio,zoom=zoom, calculate_ratio_using_CSim=calculate_ratio_using_CSim)
    image, info = flow.funcs['resize'](image)
    info_out['size'] = info['size']
    return image
 def pad(image, pad_l=0, pad_r=0, pad_t=0, pad_b=0, pad_val=0):
    """
    pad function
    specific left, right, top and bottom pad size.
    Args:
        image[np.array]: input
        pad_l: [int], pad size from left, default 0
        pad_r: [int], pad size form right, default 0
        pad_t: [int], pad size from top, default 0
        pad_b: [int], pad size form bottom, default 0
        pad_val: [float], the value of pad, , default 0 
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.pad(image_data,20,40,20,40,-0.5)
    """
    assert isinstance(image, np.ndarray)
    flow.set_padding(type='specific',pad_l=pad_l,pad_r=pad_r,pad_t=pad_t,pad_b=pad_b,pad_val=pad_val)
    image, _ = flow.funcs['padding'](image)
    return image
 def pad_center(image,size=None, pad_val=0):
    """
    pad function
    center pad with pad size.
    Args:
        image[np.array]: input
        size: [tuble], (padded_size_w, padded_size_h)
        pad_val: [float], the value of pad, , default 0 
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.pad_center(image_data,size=(56,56),pad_val=-0.5)
    """
    assert isinstance(image, np.ndarray)
    if size is None:
        return image
    assert ( (image.shape[0] <= size[1]) & (image.shape[1] <= size[0]) )
    flow.set_padding(type='center',padded_w=size[0],padded_h=size[1],pad_val=pad_val)
    image, _ = flow.funcs['padding'](image)
    return image
 def pad_corner(image,size=None, pad_val=0):
    """
    pad function
    corner pad with pad size.
    Args:
        image[np.array]: input
        size: [tuble], (padded_size_w, padded_size_h)
        pad_val: [float], the value of pad, , default 0 
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.pad_corner(image_data,size=(56,56),pad_val=-0.5)
    """   
    assert isinstance(image, np.ndarray)
    if size is None:
        return image
    assert ( (image.shape[0] <= size[1]) & (image.shape[1] <= size[0]) )
    flow.set_padding(type='corner',padded_w=size[0],padded_h=size[1],pad_val=pad_val)
    image, _ = flow.funcs['padding'](image)
    return image
 def norm(image,scale=256.,bias=-0.5, mean=None, std=None):
    """
    norm function
    x = (x/scale - bias)
    x[0,1,2] = x - mean[0,1,2]
    x[0,1,2] = x / std[0,1,2]
    Args:
        image: [np.array], input
        scale: [float], default = 256
        bias: [float], default = -0.5
        mean: [tuble,3], default = None
        std: [tuble,3], default = None
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.norm(image_data)
        >>> image_data = kneron_preprocessing.API.norm(image_data,mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    """  
    assert isinstance(image, np.ndarray)
    flow.set_normalize(type='specific',scale=scale,  bias=bias, mean=mean, std =std)
    image, _ = flow.funcs['normalize'](image)
    return image
 def inproc_520(image,raw_fmt='rgb565',raw_size=None,npu_size=None, crop_box=None, pad_mode=0, norm='kneron', gray=False, rotate=0, radix=8, bit_width=8, round_w_to_16=True, NUM_BANK_LINE=32,BANK_ENTRY_CNT=512,MAX_IMG_PREPROC_ROW_NUM=511,MAX_IMG_PREPROC_COL_NUM=256):
    """
    inproc_520
    Args:
        image: [np.array], input
        crop_box: [tuble], (x1, y1, x2, y2), if None will skip crop
        pad_mode: [int], 0: pad 2 sides, 1: pad 1 side, 2: no pad. default = 0
        norm: [str], default = 'kneron'
        rotate: [int], 0 / 1 / 2 ,default = 0
        radix: [int], default = 8
        bit_width: [int], default = 8
        round_w_to_16: [bool], default = True
        gray: [bool], default = False
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.inproc_520(image_data,npu_size=(56,56),crop_box=(272,145,460,341),pad_mode=1)
    """  
    # assert isinstance(image, np.ndarray)
    if (not isinstance(image, np.ndarray)):
        flow_520.set_raw_img(is_raw_img='yes',raw_img_type = 'bin',raw_img_fmt=raw_fmt, img_in_width=raw_size[0], img_in_height=raw_size[1])
    else:
        flow_520.set_raw_img(is_raw_img='no')
        flow_520.set_color_conversion(source_format='rgb888')
    if npu_size is None:
        return image
    flow_520.set_model_size(w=npu_size[0],h=npu_size[1])
    ## Crop
    if crop_box != None:
        flow_520.set_crop(start_x=crop_box[0],start_y=crop_box[1],end_x=crop_box[2],end_y=crop_box[3])
        crop_fisrt = True
    else:
        crop_fisrt = False
    ## Color
    if gray:
        flow_520.set_color_conversion(out_format='l',simulation='no')
    else:
        flow_520.set_color_conversion(out_format='rgb888',simulation='no')
    ## Resize & Pad
    pad_mode = str2int(pad_mode)
    if (pad_mode == 0):
        pad_type = 'center'
        resize_keep_ratio = 'yes'
    elif (pad_mode == 1):
        pad_type = 'corner'
        resize_keep_ratio = 'yes'
    else:
        pad_type = 'center'
        resize_keep_ratio = 'no'
    flow_520.set_resize(keep_ratio=resize_keep_ratio)
    flow_520.set_padding(type=pad_type)
    ## Norm
    flow_520.set_normalize(type=norm)
    ## 520 inproc
    flow_520.set_520_setting(radix=radix,bit_width=bit_width,rotate=rotate,crop_fisrt=crop_fisrt,round_w_to_16=round_w_to_16,NUM_BANK_LINE=NUM_BANK_LINE,BANK_ENTRY_CNT=BANK_ENTRY_CNT,MAX_IMG_PREPROC_ROW_NUM=MAX_IMG_PREPROC_ROW_NUM,MAX_IMG_PREPROC_COL_NUM=MAX_IMG_PREPROC_COL_NUM)
    image_data, _ = flow_520.run_whole_process(image)
    return image_data
 def inproc_720(image,raw_fmt='rgb565',raw_size=None,npu_size=None, crop_box=None, pad_mode=0, norm='kneron', gray=False):
    """
    inproc_720
    Args:
        image: [np.array], input
        crop_box: [tuble], (x1, y1, x2, y2), if None will skip crop
        pad_mode: [int], 0: pad 2 sides, 1: pad 1 side, 2: no pad. default = 0
        norm: [str], default = 'kneron'
        rotate: [int], 0 / 1 / 2 ,default = 0
        radix: [int], default = 8
        bit_width: [int], default = 8
        round_w_to_16: [bool], default = True
        gray: [bool], default = False
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.inproc_520(image_data,npu_size=(56,56),crop_box=(272,145,460,341),pad_mode=1)
    """  
    # assert isinstance(image, np.ndarray)
    if (not isinstance(image, np.ndarray)):
        flow_720.set_raw_img(is_raw_img='yes',raw_img_type = 'bin',raw_img_fmt=raw_fmt, img_in_width=raw_size[0], img_in_height=raw_size[1])
    else:
        flow_720.set_raw_img(is_raw_img='no')
        flow_720.set_color_conversion(source_format='rgb888')
    if npu_size is None:
        return image
    flow_720.set_model_size(w=npu_size[0],h=npu_size[1])
    ## Crop
    if crop_box != None:
        flow_720.set_crop(start_x=crop_box[0],start_y=crop_box[1],end_x=crop_box[2],end_y=crop_box[3])
        crop_fisrt = True
    else:
        crop_fisrt = False
    ## Color
    if gray:
        flow_720.set_color_conversion(out_format='l',simulation='no')
    else:
        flow_720.set_color_conversion(out_format='rgb888',simulation='no')
    ## Resize & Pad
    pad_mode = str2int(pad_mode)
    if (pad_mode == 0):
        pad_type = 'center'
        resize_keep_ratio = 'yes'
    elif (pad_mode == 1):
        pad_type = 'corner'
        resize_keep_ratio = 'yes'
    else:
        pad_type = 'center'
        resize_keep_ratio = 'no'
    flow_720.set_resize(keep_ratio=resize_keep_ratio)
    flow_720.set_padding(type=pad_type)
    ## 720 inproc
    # flow_720.set_720_setting(radix=radix,bit_width=bit_width,rotate=rotate,crop_fisrt=crop_fisrt,round_w_to_16=round_w_to_16,NUM_BANK_LINE=NUM_BANK_LINE,BANK_ENTRY_CNT=BANK_ENTRY_CNT,MAX_IMG_PREPROC_ROW_NUM=MAX_IMG_PREPROC_ROW_NUM,MAX_IMG_PREPROC_COL_NUM=MAX_IMG_PREPROC_COL_NUM)
    image_data, _ = flow_720.run_whole_process(image)
    return image_data
 def bit_match(data1, data2):
    """
    bit_match function
    check data1 is equal to data2 or not.
    Args:
        data1: [np.array / str], can be array or txt/bin file
        data2: [np.array / str], can be array or txt/bin file
    Returns:
        out1: [bool], is match or not
        out2: [np.array], if not match, save the position for mismatched data
    Examples:
        >>> result, mismatched = kneron_preprocessing.API.bit_match(data1,data2)
    """
    if isinstance(data1, str):
        if os.path.splitext(data1)[1] == '.bin':
            data1 = np.fromfile(data1, dtype='uint8')
        elif os.path.splitext(data1)[1] == '.txt':
            data1 = np.loadtxt(data1)
    assert isinstance(data1, np.ndarray)
    if isinstance(data2, str):
        if os.path.splitext(data2)[1] == '.bin':
            data2 = np.fromfile(data2, dtype='uint8')
        elif os.path.splitext(data2)[1] == '.txt':
            data2 = np.loadtxt(data2)
    assert isinstance(data2, np.ndarray)
    data1 = data1.reshape((-1,1))
    data2 = data2.reshape((-1,1))
    if not(len(data1) == len(data2)):
        print('error len')
        return False, np.zeros((1))
    else: 
        ans = data2 - data1    
        if len(np.where(ans>0)[0]) > 0:
            print('error',np.where(ans>0)[0])
            return False, np.where(ans>0)[0]
        else:
            print('pass')
            return True, np.zeros((1))
 def cpr_to_crp(x_start, x_end, y_start, y_end, pad_l, pad_r, pad_t, pad_b, rx_start, rx_end, ry_start, ry_end):
    """
    calculate the parameters of crop->pad->resize flow  to HW crop->resize->padding flow
    Args:
    Returns:
    Examples:
    """
    pad_l = round(pad_l * (rx_end-rx_start) / (x_end - x_start + pad_l + pad_r))
    pad_r = round(pad_r * (rx_end-rx_start) / (x_end - x_start + pad_l + pad_r)) 
    pad_t = round(pad_t * (ry_end-ry_start) / (y_end - y_start + pad_t + pad_b))
    pad_b = round(pad_b * (ry_end-ry_start) / (y_end - y_start + pad_t + pad_b))
    rx_start +=pad_l
    rx_end -=pad_r
    ry_start +=pad_t
    ry_end -=pad_b
    return x_start, x_end, y_start, y_end, pad_l, pad_r, pad_t, pad_b, rx_start, rx_end, ry_start, ry_end
--- a/kneron/exporting/yolov5/kneron_preprocessing/Cflow.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/Cflow.py
@ -0,0 +1,172 @@
 import numpy as np
 import argparse
 import kneron_preprocessing
 def main_(args):
    image = args.input_file
    filefmt = args.file_fmt
    if filefmt == 'bin':
        raw_format = args.raw_format
        raw_w = args.input_width
        raw_h = args.input_height
        image_data = kneron_preprocessing.API.load_bin(image,raw_format,(raw_w,raw_h))
    else:
        image_data = kneron_preprocessing.API.load_image(image)
    npu_w = args.width
    npu_h = args.height
    crop_first = True if args.crop_first == "True" else False
    if crop_first:
        x1 = args.x_pos
        y1 = args.y_pos
        x2 = args.crop_w + x1
        y2 = args.crop_h + y1
        crop_box = [x1,y1,x2,y2]
    else:
        crop_box = None
    pad_mode = args.pad_mode
    norm_mode = args.norm_mode
    bitwidth = args.bitwidth
    radix = args.radix
    rotate = args.rotate_mode
    ##
    image_data = kneron_preprocessing.API.inproc_520(image_data,npu_size=(npu_w,npu_h),crop_box=crop_box,pad_mode=pad_mode,norm=norm_mode,rotate=rotate,radix=radix,bit_width=bitwidth)
    output_file = args.output_file
    kneron_preprocessing.API.dump_image(image_data,output_file,'bin','rgba')
    return
 if __name__ == "__main__":
    argparser = argparse.ArgumentParser(
        description="preprocessing"
        )
    argparser.add_argument(
        '-i',
        '--input_file',
        help="input file name"
        )
    argparser.add_argument(
        '-ff',
        '--file_fmt',
        help="input file format, jpg or bin"
        )
    argparser.add_argument(
        '-rf',
        '--raw_format',
        help="input file image format, rgb or rgb565 or nir"
        )
    argparser.add_argument(
        '-i_w',
        '--input_width',
        type=int,
        help="input image width"
        )
    argparser.add_argument(
        '-i_h',
        '--input_height',
        type=int,
        help="input image height"
        )
    argparser.add_argument(
        '-o',
        '--output_file',
        help="output file name"
        )
    argparser.add_argument(
        '-s_w',
        '--width',
        type=int,
        help="output width for npu input",
        )
    argparser.add_argument(
        '-s_h',
        '--height',
        type=int,
        help="output height for npu input",
        )
    argparser.add_argument(
        '-c_f',
        '--crop_first',
        help="crop first True or False",
        )
    argparser.add_argument(
        '-x',
        '--x_pos',
        type=int,
        help="left up coordinate x",
        )
    argparser.add_argument(
        '-y',
        '--y_pos',
        type=int,
        help="left up coordinate y",
        )
    argparser.add_argument(
        '-c_w',
        '--crop_w',
        type=int,
        help="crop width",
        )
    argparser.add_argument(
        '-c_h',
        '--crop_h',
        type=int,
        help="crop height",
        )
    argparser.add_argument(
        '-p_m',
        '--pad_mode',
        type=int,
        help=" 0: pad 2 sides, 1: pad 1 side, 2: no pad.",
        )
    argparser.add_argument(
        '-n_m',
        '--norm_mode',
        help="normalizaton mode: yolo, kneron, tf."
        )
    argparser.add_argument(
        '-r_m',
        '--rotate_mode',
        type=int,
        help="rotate mode:0,1,2"
        )
    argparser.add_argument(
        '-bw',
        '--bitwidth',
        type=int,
        help="Int for bitwidth"
        )
    argparser.add_argument(
        '-r',
        '--radix',
        type=int,
        help="Int for radix"
        )
    args = argparser.parse_args()
    main_(args)
--- a/kneron/exporting/yolov5/kneron_preprocessing/Flow.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/Flow.py
--- a/kneron/exporting/yolov5/kneron_preprocessing/init.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/init.py
@ -0,0 +1,2 @@
 from .Flow import *
 from .API import *
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/ColorConversion.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/ColorConversion.py
@ -0,0 +1,285 @@
 import numpy as np
 from PIL import Image
 from .utils import signed_rounding, clip, str2bool
 format_bit = 10
 c00_yuv = 1
 c02_yuv = 1436
 c10_yuv = 1
 c11_yuv = -354
 c12_yuv = -732
 c20_yuv = 1
 c21_yuv = 1814
 c00_ycbcr = 1192
 c02_ycbcr = 1634
 c10_ycbcr = 1192
 c11_ycbcr = -401
 c12_ycbcr = -833
 c20_ycbcr = 1192
 c21_ycbcr = 2065
 Matrix_ycbcr_to_rgb888 = np.array(
    [[1.16438356e+00,  1.16438356e+00,  1.16438356e+00],
     [2.99747219e-07, - 3.91762529e-01,  2.01723263e+00],
     [1.59602686e+00, - 8.12968294e-01,  3.04059479e-06]])
 Matrix_rgb888_to_ycbcr = np.array(
    [[0.25678824, - 0.14822353,  0.43921569],
     [0.50412941, - 0.29099216, - 0.36778824],
     [0.09790588,  0.43921569, - 0.07142745]])
 Matrix_rgb888_to_yuv = np.array(
    [[ 0.29899106, -0.16877996,  0.49988381],
    [ 0.5865453,  -0.33110385, -0.41826072],
    [ 0.11446364,  0.49988381, -0.08162309]])
 # Matrix_rgb888_to_yuv = np.array(
 #     [[0.299, - 0.147,   0.615],
 #      [0.587, - 0.289, - 0.515],
 #      [0.114,   0.436, - 0.100]])
 # Matrix_yuv_to_rgb888 = np.array(
 #     [[1.000,   1.000,  1.000],
 #      [0.000, - 0.394,  2.032],
 #      [1.140, - 0.581,  0.000]])
 class runner(object):
    def __init__(self):
        self.set = {
            'print_info':'no',
            'model_size':[0,0],
            'numerical_type':'floating',
            "source_format": "rgb888",
            "out_format": "rgb888",
            "options": {
                "simulation": "no",
                "simulation_format": "rgb888"
            }
        }
    def update(self, **kwargs):
        #
        self.set.update(kwargs)
        ## simulation
        self.funs = []
        if str2bool(self.set['options']['simulation']) and self.set['source_format'].lower() in ['RGB888', 'rgb888', 'RGB', 'rgb']:
            if self.set['options']['simulation_format'].lower() in ['YUV422', 'yuv422', 'YUV', 'yuv']:
                self.funs.append(self._ColorConversion_RGB888_to_YUV422)
                self.set['source_format'] = 'YUV422'
            elif self.set['options']['simulation_format'].lower() in ['YCBCR422', 'YCbCr422', 'ycbcr422', 'YCBCR', 'YCbCr', 'ycbcr']:
                self.funs.append(self._ColorConversion_RGB888_to_YCbCr422)
                self.set['source_format'] = 'YCbCr422'
            elif self.set['options']['simulation_format'].lower() in['RGB565', 'rgb565']:
                self.funs.append(self._ColorConversion_RGB888_to_RGB565)
                self.set['source_format'] = 'RGB565'
        ## to rgb888
        if self.set['source_format'].lower() in ['YUV444', 'yuv444','YUV422', 'yuv422', 'YUV', 'yuv']:
            self.funs.append(self._ColorConversion_YUV_to_RGB888)
        elif self.set['source_format'].lower() in ['YCBCR444', 'YCbCr444', 'ycbcr444','YCBCR422', 'YCbCr422', 'ycbcr422', 'YCBCR', 'YCbCr', 'ycbcr']:
            self.funs.append(self._ColorConversion_YCbCr_to_RGB888)
        elif self.set['source_format'].lower() in ['RGB565', 'rgb565']:
            self.funs.append(self._ColorConversion_RGB565_to_RGB888)
        elif self.set['source_format'].lower() in ['l', 'L' , 'nir', 'NIR']:
            self.funs.append(self._ColorConversion_L_to_RGB888)
        elif self.set['source_format'].lower() in ['RGBA8888', 'rgba8888' , 'RGBA', 'rgba']:
            self.funs.append(self._ColorConversion_RGBA8888_to_RGB888)
        ## output format
        if self.set['out_format'].lower() in ['L', 'l']:
            self.funs.append(self._ColorConversion_RGB888_to_L)
        elif self.set['out_format'].lower() in['RGB565', 'rgb565']:
            self.funs.append(self._ColorConversion_RGB888_to_RGB565)
        elif self.set['out_format'].lower() in['RGBA', 'RGBA8888','rgba','rgba8888']:
            self.funs.append(self._ColorConversion_RGB888_to_RGBA8888)
        elif self.set['out_format'].lower() in['YUV', 'YUV444','yuv','yuv444']:
            self.funs.append(self._ColorConversion_RGB888_to_YUV444)
        elif self.set['out_format'].lower() in['YUV422','yuv422']:
            self.funs.append(self._ColorConversion_RGB888_to_YUV422)
        elif self.set['out_format'].lower() in['YCBCR', 'YCBCR444','YCbCr','YCbCr444','ycbcr','ycbcr444']:
            self.funs.append(self._ColorConversion_RGB888_to_YCbCr444)
        elif self.set['out_format'].lower() in['YCBCR422','YCbCr422','ycbcr422']:
            self.funs.append(self._ColorConversion_RGB888_to_YCbCr422)
    def print_info(self):
        print("<colorConversion>",
              "source_format:", self.set['source_format'],
              ', out_format:', self.set['out_format'],
              ', simulation:', self.set['options']['simulation'],
              ', simulation_format:', self.set['options']['simulation_format'])
    def run(self, image_data):
        assert isinstance(image_data, np.ndarray)
        # print info
        if str2bool(self.set['print_info']):
            self.print_info()
        # color
        for _, f in enumerate(self.funs):
            image_data = f(image_data)
        # output
        info = {}
        return image_data, info
    def _ColorConversion_RGB888_to_YUV444(self, image):
        ## floating
        image = image.astype('float')
        image = (image @ Matrix_rgb888_to_yuv + 0.5).astype('uint8')
        return image
    def _ColorConversion_RGB888_to_YUV422(self, image):
        # rgb888 to yuv444
        image = self._ColorConversion_RGB888_to_YUV444(image)
        # yuv444 to yuv422
        u2 = image[:, 0::2, 1]
        u4 = np.repeat(u2, 2, axis=1)
        v2 = image[:, 1::2, 2]
        v4 = np.repeat(v2, 2, axis=1)
        image[..., 1] = u4
        image[..., 2] = v4
        return image
    def _ColorConversion_YUV_to_RGB888(self, image):
        ## fixed
        h, w, c = image.shape
        image_f = image.reshape((h * w, c))
        image_rgb_f = np.zeros(image_f.shape, dtype=np.uint8)
        for i in range(h * w):
            image_y = image_f[i, 0] *1024
            if image_f[i, 1] > 127:
                image_u = -((~(image_f[i, 1] - 1)) & 0xFF)
            else:
                image_u = image_f[i, 1]
            if image_f[i, 2] > 127:
                image_v = -((~(image_f[i, 2] - 1)) & 0xFF)
            else:
                image_v = image_f[i, 2]
            image_r = c00_yuv * image_y + c02_yuv * image_v
            image_g = c10_yuv * image_y + c11_yuv * image_u + c12_yuv * image_v
            image_b = c20_yuv * image_y + c21_yuv * image_u
            image_r = signed_rounding(image_r, format_bit)
            image_g = signed_rounding(image_g, format_bit)
            image_b = signed_rounding(image_b, format_bit)
            image_r = image_r >> format_bit
            image_g = image_g >> format_bit
            image_b = image_b >> format_bit
            image_rgb_f[i, 0] = clip(image_r, 0, 255)
            image_rgb_f[i, 1] = clip(image_g, 0, 255)
            image_rgb_f[i, 2] = clip(image_b, 0, 255)
        image_rgb = image_rgb_f.reshape((h, w, c))
        return image_rgb
    def _ColorConversion_RGB888_to_YCbCr444(self, image):
        ## floating
        image = image.astype('float')
        image = (image @ Matrix_rgb888_to_ycbcr + 0.5).astype('uint8')
        image[:, :, 0] += 16
        image[:, :, 1] += 128
        image[:, :, 2] += 128
        return image
    def _ColorConversion_RGB888_to_YCbCr422(self, image):
        # rgb888 to ycbcr444
        image = self._ColorConversion_RGB888_to_YCbCr444(image)
        # ycbcr444 to ycbcr422
        cb2 = image[:, 0::2, 1]
        cb4 = np.repeat(cb2, 2, axis=1)
        cr2 = image[:, 1::2, 2]
        cr4 = np.repeat(cr2, 2, axis=1)
        image[..., 1] = cb4
        image[..., 2] = cr4
        return image
    def _ColorConversion_YCbCr_to_RGB888(self, image):
        ## floating
        if (self.set['numerical_type'] == 'floating'):
            image = image.astype('float')
            image[:, :, 0] -= 16
            image[:, :, 1] -= 128
            image[:, :, 2] -= 128
            image = ((image @ Matrix_ycbcr_to_rgb888) + 0.5).astype('uint8')
            return image
        ## fixed
        h, w, c = image.shape
        image_f = image.reshape((h * w, c))
        image_rgb_f = np.zeros(image_f.shape, dtype=np.uint8)
        for i in range(h * w):
            image_y = (image_f[i, 0] - 16) * c00_ycbcr
            image_cb = image_f[i, 1] - 128
            image_cr = image_f[i, 2] - 128
            image_r = image_y + c02_ycbcr * image_cr
            image_g = image_y + c11_ycbcr * image_cb + c12_ycbcr * image_cr
            image_b = image_y + c21_ycbcr * image_cb
            image_r = signed_rounding(image_r, format_bit)
            image_g = signed_rounding(image_g, format_bit)
            image_b = signed_rounding(image_b, format_bit)
            image_r = image_r >> format_bit
            image_g = image_g >> format_bit
            image_b = image_b >> format_bit
            image_rgb_f[i, 0] = clip(image_r, 0, 255)
            image_rgb_f[i, 1] = clip(image_g, 0, 255)
            image_rgb_f[i, 2] = clip(image_b, 0, 255)
        image_rgb = image_rgb_f.reshape((h, w, c))
        return image_rgb
    def _ColorConversion_RGB888_to_RGB565(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]>=3)
        image_rgb565 = np.zeros(image.shape, dtype=np.uint8)
        image_rgb = image.astype('uint8')
        image_rgb565[:, :, 0] = image_rgb[:, :, 0] >> 3
        image_rgb565[:, :, 1] = image_rgb[:, :, 1] >> 2
        image_rgb565[:, :, 2] = image_rgb[:, :, 2] >> 3
        return image_rgb565
    def _ColorConversion_RGB565_to_RGB888(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]==3)
        image_rgb = np.zeros(image.shape, dtype=np.uint8)
        image_rgb[:, :, 0] = image[:, :, 0] << 3
        image_rgb[:, :, 1] = image[:, :, 1] << 2
        image_rgb[:, :, 2] = image[:, :, 2] << 3
        return image_rgb
    def _ColorConversion_L_to_RGB888(self, image):
        image_L = image.astype('uint8')
        img = Image.fromarray(image_L).convert('RGB')
        image_data = np.array(img).astype('uint8')
        return image_data
    def _ColorConversion_RGB888_to_L(self, image):
        image_rgb = image.astype('uint8')
        img = Image.fromarray(image_rgb).convert('L')
        image_data = np.array(img).astype('uint8')
        return image_data
    def _ColorConversion_RGBA8888_to_RGB888(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]==4)
        return image[:,:,:3]
    def _ColorConversion_RGB888_to_RGBA8888(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]==3)
        imageA = np.concatenate((image, np.zeros((image.shape[0], image.shape[1], 1), dtype=np.uint8) ), axis=2)
        return imageA
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/Crop.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/Crop.py
@ -0,0 +1,145 @@
 import numpy as np
 from PIL import Image
 from .utils import str2int, str2float, str2bool, pad_square_to_4
 from .utils_520 import round_up_n
 from .Runner_base import Runner_base, Param_base
 class General(Param_base):
    type = 'center'
    align_w_to_4 = False
    pad_square_to_4 = False
    rounding_type = 0
    crop_w = 0
    crop_h = 0
    start_x = 0.
    start_y = 0.
    end_x = 0.
    end_y = 0.
    def update(self, **dic):
        self.type = dic['type']
        self.align_w_to_4 = str2bool(dic['align_w_to_4'])
        self.rounding_type = str2int(dic['rounding_type'])
        self.crop_w = str2int(dic['crop_w'])
        self.crop_h = str2int(dic['crop_h'])
        self.start_x = str2float(dic['start_x'])
        self.start_y = str2float(dic['start_y'])
        self.end_x = str2float(dic['end_x'])
        self.end_y = str2float(dic['end_y'])
    def __str__(self):
        str_out = [
            ', type:',str(self.type),
            ', align_w_to_4:',str(self.align_w_to_4),
            ', pad_square_to_4:',str(self.pad_square_to_4),
            ', crop_w:',str(self.crop_w),
            ', crop_h:',str(self.crop_h),
            ', start_x:',str(self.start_x),
            ', start_y:',str(self.start_y),
            ', end_x:',str(self.end_x),
            ', end_y:',str(self.end_y)]
        return(' '.join(str_out))
 class runner(Runner_base):
    ## overwrite the class in Runner_base
    general = General()
    def __str__(self):
        return('<Crop>')
    def update(self, **kwargs):
        ##
        super().update(**kwargs)
        ##
        if (self.general.start_x != self.general.end_x) and (self.general.start_y != self.general.end_y):
            self.general.type = 'specific'
        elif(self.general.type != 'specific'):
            if self.general.crop_w == 0 or self.general.crop_h == 0:
                self.general.crop_w = self.common.model_size[0]
                self.general.crop_h = self.common.model_size[1]
            assert(self.general.crop_w > 0)
            assert(self.general.crop_h > 0)
            assert(self.general.type.lower() in ['CENTER', 'Center', 'center', 'CORNER', 'Corner', 'corner'])
        else:
            assert(self.general.type == 'specific')
    def run(self, image_data):
        ## init
        img = Image.fromarray(image_data)
        w, h = img.size
        ## get range
        if self.general.type.lower() in ['CENTER', 'Center', 'center']:
            x1, y1, x2, y2 = self._calcuate_xy_center(w, h)
        elif self.general.type.lower() in ['CORNER', 'Corner', 'corner']:
            x1, y1, x2, y2 = self._calcuate_xy_corner(w, h)
        else:
            x1 = self.general.start_x
            y1 = self.general.start_y
            x2 = self.general.end_x
            y2 = self.general.end_y
            assert( ((x1 != x2) and (y1 != y2)) )
        ## rounding
        if self.general.rounding_type == 0:
            x1 = int(np.floor(x1))
            y1 = int(np.floor(y1))
            x2 = int(np.ceil(x2))
            y2 = int(np.ceil(y2))
        else:
            x1 = int(round(x1))
            y1 = int(round(y1))
            x2 = int(round(x2))
            y2 = int(round(y2))
        if self.general.align_w_to_4:
            # x1 = (x1+1) &(~3)  #//+2
            # x2 = (x2+2) &(~3)  #//+1
            x1 = (x1+3) &(~3)  #//+2
            left = w - x2
            left = (left+3) &(~3)
            x2 = w - left
        ## pad_square_to_4
        if str2bool(self.general.pad_square_to_4):
            x1,x2,y1,y2 = pad_square_to_4(x1,x2,y1,y2)
        # do crop
        box = (x1,y1,x2,y2)
        img = img.crop(box)
        # print info
        if str2bool(self.common.print_info):
            self.general.start_x = x1
            self.general.start_y = y1
            self.general.end_x = x2
            self.general.end_y = y2
            self.general.crop_w = x2 - x1
            self.general.crop_h = y2 - y1
            self.print_info()
        # output
        image_data = np.array(img)
        info = {}
        info['box'] = box
        return image_data, info
    ## protect fun
    def _calcuate_xy_center(self, w, h):
        x1 = w/2 - self.general.crop_w / 2
        y1 = h/2 - self.general.crop_h / 2
        x2 = w/2 + self.general.crop_w / 2
        y2 = h/2 + self.general.crop_h / 2
        return x1, y1, x2, y2
    def _calcuate_xy_corner(self, _1, _2):
        x1 = 0
        y1 = 0
        x2 = self.general.crop_w
        y2 = self.general.crop_h
        return x1, y1, x2, y2
    def do_crop(self, image_data, startW, startH, endW, endH):
        return image_data[startH:endH, startW:endW, :]
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/Normalize.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/Normalize.py
@ -0,0 +1,186 @@
 import numpy as np
 from .utils import str2bool, str2int, str2float, clip_ary
 class runner(object):
    def __init__(self):
        self.set = {
            'general': {
                'print_info':'no',
                'model_size':[0,0],
                'numerical_type':'floating',
                'type': 'kneron'
            },
            'floating':{
                "scale": 1,
                "bias": 0,
                "mean": "",
                "std": "",
            },
            'hw':{
                "radix":8,
                "shift":"",
                "sub":""
            }
        }
        return
    def update(self, **kwargs):
        #
        self.set.update(kwargs)
        #
        if self.set['general']['numerical_type'] == '520':
            if self.set['general']['type'].lower() in ['TF', 'Tf', 'tf']:
                self.fun_normalize = self._chen_520
                self.shift = 7 - self.set['hw']['radix']
                self.sub = 128
            elif self.set['general']['type'].lower() in ['YOLO', 'Yolo', 'yolo']:
                self.fun_normalize = self._chen_520
                self.shift = 8 - self.set['hw']['radix']
                self.sub = 0
            elif self.set['general']['type'].lower() in ['KNERON', 'Kneron', 'kneron']:
                self.fun_normalize = self._chen_520
                self.shift = 8 - self.set['hw']['radix']
                self.sub = 128
            else:
                self.fun_normalize = self._chen_520
                self.shift = 0
                self.sub = 0      
        elif self.set['general']['numerical_type'] == '720':
                self.fun_normalize = self._chen_720
                self.shift = 0
                self.sub = 0                   
        else:
            if self.set['general']['type'].lower() in ['TORCH', 'Torch', 'torch']:
                self.fun_normalize = self._normalize_torch
                self.set['floating']['scale'] = 255.
                self.set['floating']['mean'] = [0.485, 0.456, 0.406]
                self.set['floating']['std'] = [0.229, 0.224, 0.225]
            elif self.set['general']['type'].lower() in ['TF', 'Tf', 'tf']:
                self.fun_normalize = self._normalize_tf
                self.set['floating']['scale'] = 127.5
                self.set['floating']['bias'] = -1.
            elif self.set['general']['type'].lower() in ['CAFFE', 'Caffe', 'caffe']:
                self.fun_normalize = self._normalize_caffe
                self.set['floating']['mean'] = [103.939, 116.779, 123.68]
            elif self.set['general']['type'].lower() in ['YOLO', 'Yolo', 'yolo']:
                self.fun_normalize = self._normalize_yolo
                self.set['floating']['scale'] = 255.
            elif self.set['general']['type'].lower() in ['KNERON', 'Kneron', 'kneron']:
                self.fun_normalize = self._normalize_kneron
                self.set['floating']['scale'] = 256.
                self.set['floating']['bias'] = -0.5
            else:
                self.fun_normalize = self._normalize_customized
                self.set['floating']['scale'] = str2float(self.set['floating']['scale'])
                self.set['floating']['bias'] = str2float(self.set['floating']['bias'])
                if self.set['floating']['mean'] != None:
                    if len(self.set['floating']['mean']) != 3:
                        self.set['floating']['mean'] = None
                if self.set['floating']['std'] != None:
                    if len(self.set['floating']['std']) != 3:
                        self.set['floating']['std'] = None
    def print_info(self):
        if self.set['general']['numerical_type'] == '520':
            print("<normalize>",
            'numerical_type', self.set['general']['numerical_type'],
            ", type:", self.set['general']['type'],
            ', shift:',self.shift, 
            ', sub:', self.sub)
        else:
            print("<normalize>",
            'numerical_type', self.set['general']['numerical_type'],
            ", type:", self.set['general']['type'],
            ', scale:',self.set['floating']['scale'], 
            ', bias:', self.set['floating']['bias'],
            ', mean:', self.set['floating']['mean'],
            ', std:',self.set['floating']['std'])
    def run(self, image_data):
        # print info
        if str2bool(self.set['general']['print_info']):
            self.print_info()
        # norm
        image_data = self.fun_normalize(image_data)
        # output
        info = {}
        return image_data, info
    def _normalize_torch(self, x):
        if len(x.shape) != 3:
            return x
        x = x.astype('float')
        x = x / self.set['floating']['scale']
        x[..., 0] -= self.set['floating']['mean'][0]
        x[..., 1] -= self.set['floating']['mean'][1]
        x[..., 2] -= self.set['floating']['mean'][2]
        x[..., 0] /= self.set['floating']['std'][0]
        x[..., 1] /= self.set['floating']['std'][1]
        x[..., 2] /= self.set['floating']['std'][2]
        return x
    def _normalize_tf(self, x):
        # print('_normalize_tf')
        x = x.astype('float')
        x = x / self.set['floating']['scale']
        x = x + self.set['floating']['bias']
        return x
    def _normalize_caffe(self, x):
        if len(x.shape) != 3:
            return x
        x = x.astype('float')
        x = x[..., ::-1]
        x[..., 0] -= self.set['floating']['mean'][0]
        x[..., 1] -= self.set['floating']['mean'][1]
        x[..., 2] -= self.set['floating']['mean'][2]
        return x
    def _normalize_yolo(self, x):
        # print('_normalize_yolo')
        x = x.astype('float')
        x = x / self.set['floating']['scale']
        return x
    def _normalize_kneron(self, x):
        # print('_normalize_kneron')
        x = x.astype('float')
        x = x/self.set['floating']['scale']
        x = x + self.set['floating']['bias']
        return x
    def _normalize_customized(self, x):
        # print('_normalize_customized')
        x = x.astype('float')
        if  self.set['floating']['scale'] != 0:
            x = x/ self.set['floating']['scale'] 
        x = x + self.set['floating']['bias'] 
        if self.set['floating']['mean'] is not None:
            x[..., 0] -= self.set['floating']['mean'][0]
            x[..., 1] -= self.set['floating']['mean'][1]
            x[..., 2] -= self.set['floating']['mean'][2]
        if self.set['floating']['std'] is not None:
            x[..., 0] /= self.set['floating']['std'][0]
            x[..., 1] /= self.set['floating']['std'][1]
            x[..., 2] /= self.set['floating']['std'][2]
        return x
    def _chen_520(self, x):
        # print('_chen_520')
        x = (x - self.sub).astype('uint8')
        x = (np.right_shift(x,self.shift))
        x=x.astype('uint8')
        return x
    def _chen_720(self, x):
        # print('_chen_720')
        if self.shift == 1:
            x = x + np.array([[self.sub], [self.sub], [self.sub]])
        else:
            x = x + np.array([[self.sub], [self.sub], [self.sub]])
        return x
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/Padding.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/Padding.py
@ -0,0 +1,187 @@
 import numpy as np
 from PIL import Image
 from .utils import str2bool, str2int, str2float
 from .Runner_base import Runner_base, Param_base
 class General(Param_base):
    type = ''
    pad_val = ''
    padded_w = ''
    padded_h = ''
    pad_l = ''
    pad_r = ''
    pad_t = ''
    pad_b = ''
    padding_ch = 3
    padding_ch_type = 'RGB'
    def update(self, **dic):
        self.type = dic['type']
        self.pad_val = dic['pad_val']
        self.padded_w = str2int(dic['padded_w'])
        self.padded_h = str2int(dic['padded_h'])
        self.pad_l = str2int(dic['pad_l'])
        self.pad_r = str2int(dic['pad_r'])
        self.pad_t = str2int(dic['pad_t'])
        self.pad_b = str2int(dic['pad_b'])
    def __str__(self):
        str_out = [
            ', type:',str(self.type),
            ', pad_val:',str(self.pad_val),
            ', pad_l:',str(self.pad_l),
            ', pad_r:',str(self.pad_r),
            ', pad_r:',str(self.pad_t),
            ', pad_b:',str(self.pad_b),
            ', padding_ch:',str(self.padding_ch)]
        return(' '.join(str_out))
 class Hw(Param_base):
    radix = 8
    normalize_type = 'floating'
    def update(self, **dic):
        self.radix = dic['radix']
        self.normalize_type = dic['normalize_type']
    def __str__(self):
        str_out = [
            ', radix:', str(self.radix),
            ', normalize_type:',str(self.normalize_type)]
        return(' '.join(str_out))
 class runner(Runner_base):
    ## overwrite the class in Runner_base
    general = General()
    hw = Hw()
    def __str__(self):
        return('<Padding>')
    def update(self, **kwargs):
        super().update(**kwargs)
        ## update pad type & pad length
        if (self.general.pad_l != 0) or (self.general.pad_r != 0) or (self.general.pad_t != 0) or (self.general.pad_b != 0):
            self.general.type = 'specific'
            assert(self.general.pad_l >= 0)
            assert(self.general.pad_r >= 0)
            assert(self.general.pad_t >= 0)
            assert(self.general.pad_b >= 0)
        elif(self.general.type != 'specific'):
            if self.general.padded_w == 0 or self.general.padded_h == 0:
                self.general.padded_w = self.common.model_size[0]
                self.general.padded_h = self.common.model_size[1]
            assert(self.general.padded_w > 0)
            assert(self.general.padded_h > 0)
            assert(self.general.type.lower() in ['CENTER', 'Center', 'center', 'CORNER', 'Corner', 'corner'])
        else:
            assert(self.general.type == 'specific')
        ## decide pad_val & padding ch
        # if numerical_type is floating
        if (self.common.numerical_type == 'floating'):
            if self.general.pad_val != 'edge':
                self.general.pad_val = str2float(self.general.pad_val)
            self.general.padding_ch = 3
            self.general.padding_ch_type = 'RGB'
        # if numerical_type is 520 or 720
        else: 
            if self.general.pad_val == '':
                if self.hw.normalize_type.lower() in ['TF', 'Tf', 'tf']:
                    self.general.pad_val = np.uint8(-128 >> (7 - self.hw.radix))
                elif self.hw.normalize_type.lower() in ['YOLO', 'Yolo', 'yolo']:
                    self.general.pad_val = np.uint8(0 >> (8 - self.hw.radix))
                elif self.hw.normalize_type.lower() in ['KNERON', 'Kneron', 'kneron']:
                    self.general.pad_val = np.uint8(-128 >> (8 - self.hw.radix))
                else:
                    self.general.pad_val = np.uint8(0 >> (8 - self.hw.radix))
            else:
                self.general.pad_val = str2int(self.general.pad_val)
            self.general.padding_ch = 4
            self.general.padding_ch_type = 'RGBA'
    def run(self, image_data):
        # init
        shape = image_data.shape
        w = shape[1]
        h = shape[0]
        if len(shape) < 3:
            self.general.padding_ch = 1
            self.general.padding_ch_type = 'L'
        else:
            if shape[2] == 3 and self.general.padding_ch == 4:
                image_data = np.concatenate((image_data, np.zeros((h, w, 1), dtype=np.uint8) ), axis=2)
        ## padding
        if self.general.type.lower() in ['CENTER',  'Center',  'center']:
            img_pad = self._padding_center(image_data, w, h)
        elif self.general.type.lower() in ['CORNER',  'Corner',  'corner']:
            img_pad = self._padding_corner(image_data, w, h)
        else:
            img_pad = self._padding_sp(image_data, w, h)
        # print info
        if str2bool(self.common.print_info):
            self.print_info()
        # output
        info = {}
        return img_pad, info
    ## protect fun
    def _padding_center(self, img, ori_w, ori_h):
        # img_pad = Image.new(self.general.padding_ch_type, (self.general.padded_w, self.general.padded_h), int(self.general.pad_val[0]))
        # img = Image.fromarray(img)
        # img_pad.paste(img, ((self.general.padded_w-ori_w)//2, (self.general.padded_h-ori_h)//2))
        # return img_pad
        padH = self.general.padded_h - ori_h
        padW = self.general.padded_w - ori_w
        self.general.pad_t = padH // 2
        self.general.pad_b = (padH // 2) + (padH % 2)
        self.general.pad_l = padW // 2
        self.general.pad_r = (padW // 2) + (padW % 2)
        if self.general.pad_l < 0 or self.general.pad_r <0 or self.general.pad_t <0 or self.general.pad_b<0:
            return img
        img_pad = self._padding_sp(img,ori_w,ori_h)
        return img_pad
    def _padding_corner(self, img, ori_w, ori_h):
        # img_pad = Image.new(self.general.padding_ch_type, (self.general.padded_w, self.general.padded_h), self.general.pad_val)
        # img_pad.paste(img, (0, 0))
        self.general.pad_l = 0
        self.general.pad_r = self.general.padded_w - ori_w
        self.general.pad_t = 0
        self.general.pad_b = self.general.padded_h - ori_h
        if self.general.pad_l < 0 or self.general.pad_r <0 or self.general.pad_t <0 or self.general.pad_b<0:
            return img
        img_pad = self._padding_sp(img,ori_w,ori_h)
        return img_pad
    def _padding_sp(self, img, ori_w, ori_h):
        # block_t = np.zeros((self.general.pad_t, self.general.pad_l + self.general.pad_r + ori_w, self.general.padding_ch), dtype=np.float)
        # block_l = np.zeros((ori_h, self.general.pad_l, self.general.padding_ch), dtype=np.float)
        # block_r = np.zeros((ori_h, self.general.pad_r, self.general.padding_ch), dtype=np.float)
        # block_b = np.zeros((self.general.pad_b, self.general.pad_l + self.general.pad_r + ori_w, self.general.padding_ch), dtype=np.float)
        # for i in range(self.general.padding_ch):
        #     block_t[:, :, i] = np.ones(block_t[:, :, i].shape, dtype=np.float) * self.general.pad_val
        #     block_l[:, :, i] = np.ones(block_l[:, :, i].shape, dtype=np.float) * self.general.pad_val
        #     block_r[:, :, i] = np.ones(block_r[:, :, i].shape, dtype=np.float) * self.general.pad_val
        #     block_b[:, :, i] = np.ones(block_b[:, :, i].shape, dtype=np.float) * self.general.pad_val
        # padded_image_hor = np.concatenate((block_l, img, block_r), axis=1)
        # padded_image = np.concatenate((block_t, padded_image_hor, block_b), axis=0)
        # return padded_image
        if self.general.padding_ch == 1:
            pad_range = ( (self.general.pad_t, self.general.pad_b),(self.general.pad_l, self.general.pad_r) )
        else:
            pad_range = ((self.general.pad_t, self.general.pad_b),(self.general.pad_l, self.general.pad_r),(0,0))
        if isinstance(self.general.pad_val, str):
            if self.general.pad_val == 'edge':
                padded_image = np.pad(img, pad_range, mode="edge")
            else:
                padded_image = np.pad(img, pad_range, mode="constant",constant_values=0)
        else:
            padded_image = np.pad(img, pad_range, mode="constant",constant_values=self.general.pad_val)
        return padded_image
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/Resize.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/Resize.py
@ -0,0 +1,237 @@
 import numpy as np
 import cv2
 from PIL import Image
 from .utils import str2bool, str2int
 from ctypes import c_float
 from .Runner_base import Runner_base, Param_base
 class General(Param_base):
    type = 'bilinear'
    keep_ratio = True
    zoom = True
    calculate_ratio_using_CSim = True
    resize_w = 0
    resize_h = 0
    resized_w = 0
    resized_h = 0
    def update(self, **dic):
        self.type = dic['type']
        self.keep_ratio = str2bool(dic['keep_ratio'])
        self.zoom = str2bool(dic['zoom'])
        self.calculate_ratio_using_CSim = str2bool(dic['calculate_ratio_using_CSim'])
        self.resize_w = str2int(dic['resize_w'])
        self.resize_h = str2int(dic['resize_h'])
    def __str__(self):
        str_out = [
            ', type:',str(self.type),
            ', keep_ratio:',str(self.keep_ratio),
            ', zoom:',str(self.zoom),
            ', calculate_ratio_using_CSim:',str(self.calculate_ratio_using_CSim),
            ', resize_w:',str(self.resize_w),
            ', resize_h:',str(self.resize_h),
            ', resized_w:',str(self.resized_w),
            ', resized_h:',str(self.resized_h)]
        return(' '.join(str_out))
 class Hw(Param_base):
    resize_bit = 12
    def update(self, **dic):
        pass
    def __str__(self):
        str_out = [
            ', resize_bit:',str(self.resize_bit)]
        return(' '.join(str_out))
 class runner(Runner_base):
    ## overwrite the class in Runner_base
    general = General()
    hw = Hw()
    def __str__(self):
        return('<Resize>')
    def update(self, **kwargs):
        super().update(**kwargs)
        ## if resize size has not been assigned, then it will take model size as resize size
        if self.general.resize_w == 0 or self.general.resize_h == 0:
            self.general.resize_w = self.common.model_size[0]
            self.general.resize_h = self.common.model_size[1]
        assert(self.general.resize_w > 0)
        assert(self.general.resize_h > 0)
        ##
        if self.common.numerical_type == '520':
            self.general.type = 'fixed_520'
        elif self.common.numerical_type == '720':
            self.general.type = 'fixed_720'
        assert(self.general.type.lower() in ['BILINEAR',  'Bilinear',  'bilinear', 'BICUBIC',  'Bicubic',  'bicubic', 'FIXED',  'Fixed', 'fixed', 'FIXED_520',  'Fixed_520',  'fixed_520', 'FIXED_720', 'Fixed_720', 'fixed_720','CV', 'cv', 'opencv', 'OpenCV', 'CV2', 'cv2'])
    def run(self, image_data):
        ## init
        ori_w = image_data.shape[1]
        ori_h = image_data.shape[0]
        info = {}
        ##
        if self.general.keep_ratio:
            self.general.resized_w, self.general.resized_h = self.calcuate_scale_keep_ratio(self.general.resize_w,self.general.resize_h, ori_w, ori_h, self.general.calculate_ratio_using_CSim)
        else:
            self.general.resized_w = int(self.general.resize_w)
            self.general.resized_h = int(self.general.resize_h)
        assert(self.general.resized_w > 0)
        assert(self.general.resized_h > 0)
        ##
        if (self.general.resized_w > ori_w) or (self.general.resized_h > ori_h):
            if not self.general.zoom: 
                info['size'] = (ori_w,ori_h)
                if str2bool(self.common.print_info):
                    print('no resize')
                    self.print_info()
                return image_data, info
        ## resize
        if self.general.type.lower() in ['BILINEAR',  'Bilinear',  'bilinear']:
            image_data = self.do_resize_bilinear(image_data, self.general.resized_w, self.general.resized_h)
        elif self.general.type.lower() in ['BICUBIC',  'Bicubic',  'bicubic']:
            image_data = self.do_resize_bicubic(image_data, self.general.resized_w, self.general.resized_h)
        elif self.general.type.lower() in ['CV',  'cv',  'opencv', 'OpenCV',  'CV2',  'cv2']:
            image_data = self.do_resize_cv2(image_data, self.general.resized_w, self.general.resized_h)
        elif self.general.type.lower() in ['FIXED',  'Fixed',  'fixed', 'FIXED_520',  'Fixed_520',  'fixed_520', 'FIXED_720', 'Fixed_720', 'fixed_720']:
            image_data = self.do_resize_fixed(image_data, self.general.resized_w, self.general.resized_h, self.hw.resize_bit, self.general.type)
        # output
        info['size'] = (self.general.resized_w, self.general.resized_h)
        # print info
        if str2bool(self.common.print_info):
            self.print_info()
        return image_data, info
    def calcuate_scale_keep_ratio(self, tar_w, tar_h, ori_w, ori_h, calculate_ratio_using_CSim):
        if not calculate_ratio_using_CSim:
            scale_w = tar_w * 1.0 / ori_w*1.0
            scale_h = tar_h * 1.0 / ori_h*1.0
            scale = scale_w if scale_w < scale_h else scale_h
            new_w = int(round(ori_w * scale))
            new_h = int(round(ori_h * scale))
            return new_w, new_h
        ## calculate_ratio_using_CSim
        scale_w = c_float(tar_w * 1.0 / (ori_w * 1.0)).value
        scale_h = c_float(tar_h * 1.0 / (ori_h * 1.0)).value
        scale_ratio = 0.0
        scale_target_w = 0
        scale_target_h = 0
        padH = 0
        padW = 0
        bScaleW = True if scale_w < scale_h else False
        if bScaleW:
            scale_ratio = scale_w
            scale_target_w = int(c_float(scale_ratio * ori_w + 0.5).value)
            scale_target_h = int(c_float(scale_ratio * ori_h + 0.5).value)
            assert (abs(scale_target_w - tar_w) <= 1), "Error: scale down width cannot meet expectation\n"
            padH = tar_h - scale_target_h
            padW = 0
            assert (padH >= 0), "Error: padH shouldn't be less than zero\n"
        else:
            scale_ratio = scale_h 
            scale_target_w = int(c_float(scale_ratio * ori_w + 0.5).value)
            scale_target_h = int(c_float(scale_ratio * ori_h + 0.5).value)
            assert (abs(scale_target_h - tar_h) <= 1), "Error: scale down height cannot meet expectation\n"
            padW = tar_w - scale_target_w
            padH = 0
            assert (padW >= 0), "Error: padW shouldn't be less than zero\n"
        new_w = tar_w - padW
        new_h = tar_h - padH
        return new_w, new_h
    def do_resize_bilinear(self, image_data, resized_w, resized_h):
        img = Image.fromarray(image_data)
        img = img.resize((resized_w, resized_h), Image.BILINEAR)
        image_data = np.array(img).astype('uint8')
        return image_data        
    def do_resize_bicubic(self, image_data, resized_w, resized_h):
        img = Image.fromarray(image_data)
        img = img.resize((resized_w, resized_h), Image.BICUBIC)
        image_data = np.array(img).astype('uint8')
        return image_data
    def do_resize_cv2(self, image_data, resized_w, resized_h):
        image_data = cv2.resize(image_data, (resized_w, resized_h))
        image_data = np.array(image_data)
        # image_data = np.array(image_data).astype('uint8')
        return image_data
    def do_resize_fixed(self, image_data, resized_w, resized_h, resize_bit, type):
        if len(image_data.shape) < 3:
            m, n = image_data.shape
            tmp = np.zeros((m,n,3), dtype=np.uint8)
            tmp[:,:,0] = image_data
            image_data = tmp
            c = 3
            gray = True
        else:
            m, n, c = image_data.shape
            gray = False
        resolution = 1 << resize_bit
        # Width
        ratio = int(((n - 1) << resize_bit) / (resized_w - 1))
        ratio_cnt = 0
        src_x = 0
        resized_image_w = np.zeros((m, resized_w, c), dtype=np.uint8)
        for dst_x in range(resized_w):
            while ratio_cnt > resolution:
                ratio_cnt = ratio_cnt - resolution
                src_x = src_x + 1
            mul1 = np.ones((m, c)) * (resolution - ratio_cnt)
            mul2 = np.ones((m, c)) * ratio_cnt
            resized_image_w[:, dst_x, :] = np.multiply(np.multiply(
                image_data[:, src_x, :], mul1) + np.multiply(image_data[:, src_x + 1, :], mul2), 1/resolution)
            ratio_cnt = ratio_cnt + ratio
        # Height
        ratio = int(((m - 1) << resize_bit) / (resized_h - 1))
        ## NPU HW special case 2 , only on 520
        if type.lower() in ['FIXED_520',  'Fixed_520',  'fixed_520']:
            if (((ratio * (resized_h - 1)) % 4096 == 0) and ratio != 4096):
                ratio -= 1
        ratio_cnt = 0
        src_x = 0
        resized_image = np.zeros(
            (resized_h, resized_w, c), dtype=np.uint8)
        for dst_x in range(resized_h):
            while ratio_cnt > resolution:
                ratio_cnt = ratio_cnt - resolution
                src_x = src_x + 1
            mul1 = np.ones((resized_w, c)) * (resolution - ratio_cnt)
            mul2 = np.ones((resized_w, c)) * ratio_cnt
            ## NPU HW special case 1 , both on 520 / 720
            if (((dst_x > 0) and ratio_cnt == resolution) and (ratio != resolution)):
                if type.lower() in ['FIXED_520',  'Fixed_520',  'fixed_520','FIXED_720',  'Fixed_720',  'fixed_720' ]:
                    resized_image[dst_x, :, :] = np.multiply(np.multiply(
                        resized_image_w[src_x+1, :, :], mul1) + np.multiply(resized_image_w[src_x + 2, :, :], mul2), 1/resolution)
            else:
                resized_image[dst_x, :, :] = np.multiply(np.multiply(
                    resized_image_w[src_x, :, :], mul1) + np.multiply(resized_image_w[src_x + 1, :, :], mul2), 1/resolution)
            ratio_cnt = ratio_cnt + ratio
        if gray:
            resized_image = resized_image[:,:,0]
        return resized_image
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/Rotate.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/Rotate.py
@ -0,0 +1,45 @@
 import numpy as np
 from .utils import str2bool, str2int
 class runner(object):
    def __init__(self, *args, **kwargs):
        self.set = {
            'operator': '',
            "rotate_direction": 0,
        }
        self.update(*args, **kwargs)
    def update(self, *args, **kwargs):
        self.set.update(kwargs)
        self.rotate_direction = str2int(self.set['rotate_direction'])
        # print info
        if str2bool(self.set['b_print']):
            self.print_info()
    def print_info(self):
        print("<rotate>",
            'rotate_direction', self.rotate_direction,)
    def run(self, image_data):
        image_data = self._rotate(image_data)
        return image_data
    def _rotate(self,img):
        if self.rotate_direction == 1 or self.rotate_direction == 2:
            col, row, unit = img.shape
            pInBuf = img.reshape((-1,1))
            pOutBufTemp = np.zeros((col* row* unit))
            for r in range(row):
                for c in range(col):
                    for u in range(unit):
                        if self.rotate_direction == 1:
                            pOutBufTemp[unit * (c * row + (row - r - 1))+u] = pInBuf[unit * (r * col + c)+u]
                        elif self.rotate_direction == 2:
                            pOutBufTemp[unit * (row * (col - c - 1) + r)+u] = pInBuf[unit * (r * col + c)+u]
            img = pOutBufTemp.reshape((col,row,unit))
        return img
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/Runner_base.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/Runner_base.py
@ -0,0 +1,59 @@
 from abc import ABCMeta, abstractmethod
 class Param_base(object):
    @abstractmethod
    def update(self,**dic):
        raise NotImplementedError("Must override")
    def load_dic(self, key, **dic):
        if key in dic:
            param = eval('self.'+key)
            param = dic[key]
    def __str__(self):
        str_out = []
        return(' '.join(str_out))
 class Common(Param_base):
    print_info = False
    model_size = [0,0]
    numerical_type = 'floating'
    def update(self, **dic):
        self.print_info = dic['print_info']
        self.model_size = dic['model_size']
        self.numerical_type = dic['numerical_type']
    def __str__(self):
        str_out = ['numerical_type:',str(self.numerical_type)]
        return(' '.join(str_out))
 class Runner_base(metaclass=ABCMeta):
    common = Common()
    general = Param_base()
    floating = Param_base()
    hw = Param_base()
    def update(self, **kwargs):
        ## update param
        self.common.update(**kwargs['common'])
        self.general.update(**kwargs['general'])
        assert(self.common.numerical_type.lower() in ['floating', '520', '720'])
        if (self.common.numerical_type == 'floating'):
            if (self.floating.__class__.__name__ != 'Param_base'):
                self.floating.update(**kwargs['floating'])
        else:
            if (self.hw.__class__.__name__ != 'Param_base'):
                self.hw.update(**kwargs['hw'])
    def print_info(self):
        if (self.common.numerical_type == 'floating'):
            print(self, self.common, self.general, self.floating)
        else:
            print(self, self.common, self.general, self.hw)
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/init.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/init.py
@ -0,0 +1,2 @@
 from . import ColorConversion, Padding, Resize, Crop, Normalize, Rotate
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/utils.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/utils.py
@ -0,0 +1,372 @@
 import numpy as np
 from PIL import Image
 import struct
 def pad_square_to_4(x_start, x_end, y_start, y_end):
    w_int = x_end - x_start 
    h_int = y_end - y_start
    pad = w_int - h_int
    if pad > 0:
        pad_s = (pad >> 1) &(~3)
        pad_e = pad - pad_s
        y_start -= pad_s
        y_end += pad_e
    else:#//pad <=0
        pad_s = -(((pad) >> 1) &(~3))
        pad_e = (-pad) - pad_s
        x_start -= pad_s
        x_end += pad_e
    return x_start, x_end, y_start, y_end
 def str_fill(value):
    if len(value) == 1:
        value = "0" + value
    elif len(value) == 0:
        value = "00"
    return value
 def clip_ary(value):
    list_v = []
    for i in range(len(value)):
        v = value[i] % 256
        list_v.append(v)
    return list_v
 def str2bool(v):
    if isinstance(v,bool):
        return v
    return v.lower() in ('TRUE', 'True', 'true', '1', 'T', 't', 'Y', 'YES', 'y', 'yes')
 def str2int(s):
    if s == "":
        s = 0
    s = int(s)
    return s
 def str2float(s):
    if s == "":
        s = 0
    s = float(s)
    return s
 def clip(value, mini, maxi):
    if value < mini:
        result = mini
    elif value > maxi:
        result = maxi
    else:
        result = value
    return result
 def clip_ary(value):
    list_v = []
    for i in range(len(value)):
        v = value[i] % 256
        list_v.append(v)
    return list_v
 def signed_rounding(value, bit):
    if value < 0:
        value = value - (1 << (bit - 1))
    else:
        value = value + (1 << (bit - 1))
    return value
 def hex_loader(data_folder,**kwargs):
    format_mode = kwargs['raw_img_fmt']
    src_h = kwargs['img_in_height']
    src_w = kwargs['img_in_width']
    if format_mode in ['YUV444', 'yuv444', 'YCBCR444', 'YCbCr444', 'ycbcr444']:
        output = hex_yuv444(data_folder,src_h,src_w)
    elif format_mode in ['RGB565', 'rgb565']:
        output = hex_rgb565(data_folder,src_h,src_w)
    elif format_mode in ['YUV422', 'yuv422', 'YCBCR422', 'YCbCr422', 'ycbcr422']:
        output = hex_yuv422(data_folder,src_h,src_w)
    return output
 def hex_rgb565(hex_folder,src_h,src_w):
    pix_per_line = 8
    byte_per_line = 16
    f = open(hex_folder)
    pixel_r = []
    pixel_g = []
    pixel_b = []
    # Ignore the first line
    f.readline()
    input_line = int((src_h * src_w)/pix_per_line)
    for i in range(input_line):
        readline = f.readline()
        for j in range(int(byte_per_line/2)-1, -1, -1):
            data1 = int(readline[(j * 4 + 0):(j * 4 + 2)], 16)
            data0 = int(readline[(j * 4 + 2):(j * 4 + 4)], 16)
            r = ((data1 & 0xf8) >> 3)
            g = (((data0 & 0xe0) >> 5) + ((data1 & 0x7) << 3))
            b = (data0 & 0x1f)
            pixel_r.append(r)
            pixel_g.append(g)
            pixel_b.append(b)
    ary_r = np.array(pixel_r, dtype=np.uint8)
    ary_g = np.array(pixel_g, dtype=np.uint8)
    ary_b = np.array(pixel_b, dtype=np.uint8)
    output = np.concatenate((ary_r[:, None], ary_g[:, None], ary_b[:, None]), axis=1)
    output = output.reshape((src_h, src_w, 3))
    return output
 def hex_yuv444(hex_folder,src_h,src_w):
    pix_per_line = 4
    byte_per_line = 16
    f = open(hex_folder)
    byte0 = []
    byte1 = []
    byte2 = []
    byte3 = []
    # Ignore the first line
    f.readline()
    input_line = int((src_h * src_w)/pix_per_line)
    for i in range(input_line):
        readline = f.readline()
        for j in range(byte_per_line-1, -1, -1):
            data = int(readline[(j*2):(j*2+2)], 16)
            if (j+1) % 4 == 0:
                byte0.append(data)
            elif (j+2) % 4 == 0:
                byte1.append(data)
            elif (j+3) % 4 == 0:
                byte2.append(data)
            elif (j+4) % 4 == 0:
                byte3.append(data)
    # ary_a = np.array(byte0, dtype=np.uint8)
    ary_v = np.array(byte1, dtype=np.uint8)
    ary_u = np.array(byte2, dtype=np.uint8)
    ary_y = np.array(byte3, dtype=np.uint8)
    output = np.concatenate((ary_y[:, None], ary_u[:, None], ary_v[:, None]), axis=1)
    output = output.reshape((src_h, src_w, 3))
    return output
 def hex_yuv422(hex_folder,src_h,src_w):
    pix_per_line = 8
    byte_per_line = 16
    f = open(hex_folder)
    pixel_y = []
    pixel_u = []
    pixel_v = []
    # Ignore the first line
    f.readline()
    input_line = int((src_h * src_w)/pix_per_line)
    for i in range(input_line):
        readline = f.readline()
        for j in range(int(byte_per_line/4)-1, -1, -1):
            data3 = int(readline[(j * 8 + 0):(j * 8 + 2)], 16)
            data2 = int(readline[(j * 8 + 2):(j * 8 + 4)], 16)
            data1 = int(readline[(j * 8 + 4):(j * 8 + 6)], 16)
            data0 = int(readline[(j * 8 + 6):(j * 8 + 8)], 16)
            pixel_y.append(data3)
            pixel_y.append(data1)
            pixel_u.append(data2)
            pixel_u.append(data2)
            pixel_v.append(data0)
            pixel_v.append(data0)
    ary_y = np.array(pixel_y, dtype=np.uint8)
    ary_u = np.array(pixel_u, dtype=np.uint8)
    ary_v = np.array(pixel_v, dtype=np.uint8)
    output = np.concatenate((ary_y[:, None], ary_u[:, None], ary_v[:, None]), axis=1)
    output = output.reshape((src_h, src_w, 3))
    return output
 def bin_loader(data_folder,**kwargs):
    format_mode = kwargs['raw_img_fmt']
    src_h = kwargs['img_in_height']
    src_w = kwargs['img_in_width']
    if format_mode in ['YUV','yuv','YUV444', 'yuv444', 'YCBCR','YCbCr','ycbcr','YCBCR444', 'YCbCr444', 'ycbcr444']:
        output = bin_yuv444(data_folder,src_h,src_w)
    elif format_mode in ['RGB565', 'rgb565']:
        output = bin_rgb565(data_folder,src_h,src_w)
    elif format_mode in ['NIR', 'nir','NIR888', 'nir888']:
        output = bin_nir(data_folder,src_h,src_w)
    elif format_mode in ['YUV422', 'yuv422', 'YCBCR422', 'YCbCr422', 'ycbcr422']:
        output = bin_yuv422(data_folder,src_h,src_w)
    elif format_mode in ['RGB888','rgb888']:
        output = np.fromfile(data_folder, dtype='uint8')
        output = output.reshape(src_h,src_w,3)
    elif format_mode in ['RGBA8888','rgba8888', 'RGBA' , 'rgba']:
        output_temp = np.fromfile(data_folder, dtype='uint8')
        output_temp = output_temp.reshape(src_h,src_w,4)
        output = output_temp[:,:,0:3]
    return output
 def bin_yuv444(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    row = src_h
    col = src_w
    pixels = row*col
    raw = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            raw.append(s[0])
    raw = raw[:pixels*4]
    #
    output = np.zeros((pixels * 3), dtype=np.uint8)
    cnt = 0
    for i in range(0, pixels*4, 4):
        #Y
        output[cnt] = raw[i+3]
        #U
        cnt += 1
        output[cnt] = raw[i+2]
        #V
        cnt += 1
        output[cnt] = raw[i+1]
        cnt += 1          
    output = output.reshape((src_h,src_w,3))
    return output
 def bin_yuv422(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    row = src_h
    col = src_w
    pixels = row*col
    raw = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            raw.append(s[0])
    raw = raw[:pixels*2]
    #
    output = np.zeros((pixels * 3), dtype=np.uint8)
    cnt = 0
    for i in range(0, pixels*2, 4):
        #Y0
        output[cnt] = raw[i+3]
        #U0
        cnt += 1
        output[cnt] = raw[i+2]
        #V0
        cnt += 1
        output[cnt] = raw[i]
        #Y1
        cnt += 1
        output[cnt] = raw[i+1]
        #U1
        cnt += 1
        output[cnt] = raw[i+2]
        #V1
        cnt += 1
        output[cnt] = raw[i]
        cnt += 1          
    output = output.reshape((src_h,src_w,3))
    return output
 def bin_rgb565(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    row = src_h
    col = src_w
    pixels = row*col
    rgba565 = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            rgba565.append(s[0])
    rgba565 = rgba565[:pixels*2]
    # rgb565_bin to numpy_array
    output = np.zeros((pixels * 3), dtype=np.uint8)
    cnt = 0
    for i in range(0, pixels*2, 2):
        temp = rgba565[i]
        temp2 = rgba565[i+1]
        #R-5
        output[cnt] = (temp2 >>3)
        #G-6
        cnt += 1
        output[cnt] = ((temp & 0xe0) >> 5) + ((temp2 & 0x07) << 3)
        #B-5
        cnt += 1
        output[cnt] = (temp & 0x1f)
        cnt += 1          
    output = output.reshape((src_h,src_w,3))
    return output
 def bin_nir(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    nir = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            nir.append(s[0])
    nir = nir[:src_h*src_w]
    pixels = len(nir)
    # nir_bin to numpy_array
    output = np.zeros((len(nir) * 3), dtype=np.uint8)
    for i in range(0, pixels):
        output[i*3]=nir[i]
        output[i*3+1]=nir[i]
        output[i*3+2]=nir[i]
    output = output.reshape((src_h,src_w,3))
    return output
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/utils_520.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/utils_520.py
@ -0,0 +1,50 @@
 import math
 def round_up_16(num):
    return ((num + (16 - 1)) & ~(16 - 1))
 def round_up_n(num, n):
    if (num > 0):
        temp = float(num) / n
        return math.ceil(temp) * n
    else:
        return -math.ceil(float(-num) / n) * n
 def cal_img_row_offset(crop_num, pad_num, start_row, out_row, orig_row):
    scaled_img_row = int(out_row - (pad_num[1] + pad_num[3]))
    if ((start_row - pad_num[1]) > 0):
        img_str_row = int((start_row - pad_num[1]))
    else:
        img_str_row = 0
    valid_row = int(orig_row - (crop_num[1] + crop_num[3]))
    img_str_row = int(valid_row * img_str_row / scaled_img_row)
    return int(img_str_row + crop_num[1])
 def get_pad_num(pad_num_orig, left, up, right, bottom):
    pad_num = [0]*4
    for i in range(0,4):
        pad_num[i] = pad_num_orig[i]
    if not (left):
        pad_num[0] = 0
    if not (up):
        pad_num[1] = 0
    if not (right):
        pad_num[2] = 0
    if not (bottom):
        pad_num[3] = 0
    return pad_num
 def get_byte_per_pixel(raw_fmt):
    if raw_fmt.lower() in ['RGB888', 'rgb888', 'RGB', 'rgb888']:
        return 4
    elif raw_fmt.lower() in ['YUV', 'yuv', 'YUV422', 'yuv422']:
        return 2
    elif raw_fmt.lower() in ['RGB565', 'rgb565']:
        return 2
    elif raw_fmt.lower() in ['NIR888', 'nir888', 'NIR', 'nir']:
        return 1
    else:
        return -1
--- a/kneron/exporting/yolov5/kneron_preprocessing/funcs/utils_720.py
+++ b/kneron/exporting/yolov5/kneron_preprocessing/funcs/utils_720.py
@ -0,0 +1,42 @@
 import numpy as np
 from PIL import Image
 def twos_complement(value):
    value = int(value)
    # msb = (value & 0x8000) * (1/np.power(2, 15))
    msb = (value & 0x8000) >> 15
    if msb == 1:
        if (((~value) & 0xFFFF) + 1) >= 0xFFFF:
            result = ((~value) & 0xFFFF)
        else:
            result = (((~value) & 0xFFFF) + 1)
        result = result * (-1)
    else:
        result = value
    return result
 def twos_complement_pix(value):
    h, _ = value.shape
    for i in range(h):
        value[i, 0] = twos_complement(value[i, 0])
    return value
 def clip(value, mini, maxi):
    if value < mini:
        result = mini
    elif value > maxi:
        result = maxi
    else:
        result = value
    return result
 def clip_pix(value, mini, maxi):
    h, _ = value.shape
    for i in range(h):
        value[i, 0] = clip(value[i, 0], mini, maxi)
    return value
--- a/kneron/exporting/yolov5/quantize_yolov5.py
+++ b/kneron/exporting/yolov5/quantize_yolov5.py
@ -0,0 +1,45 @@
 import os
 import numpy as np
 import torch
 import ktc  # Kneron Toolchain
 from yolov5_preprocess import Yolov5_preprocess  # 使用你的預處理
 import kneron_preprocessing
 # 設定裝置
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 # 設定圖片大小（與訓練時一致）
 imgsz_h, imgsz_w = 640, 640
 # 量化數據集目錄（請確保這個資料夾存在）
 data_path = "/data50"
 img_list = []
 # 設定 ONNX 模型路徑（確保這個路徑在 Docker 內部是否正確）
 onnx_model_path = "/workspace/yolov5/latest.opt.onnx"
 # **初始化 Kneron ModelConfig 物件**
 km = ktc.ModelConfig(20008, "0001", "720", onnx_model=onnx_model_path)
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # 執行與訓練相同的預處理
        img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
        # 確保格式為 NumPy 陣列
        img_data = img_data.cpu().numpy()
        print(f"Processed: {fullpath}")
        img_list.append(img_data)
 # 轉為 NumPy 格式
 img_list = np.array(img_list)
 # **執行 BIE 量化分析**
 bie_model_path = km.analysis({"input": img_list})
 # 輸出成功訊息
 print("\n✅ Fixed-point analysis done! BIE model saved to:", bie_model_path)
--- a/kneron/exporting/yolov5/readme.txt
+++ b/kneron/exporting/yolov5/readme.txt
@ -0,0 +1,68 @@
 (1)yolov5_app.py for ploting moldel inference results
 cd applications
 python yolov5_app.py
 (2)yolov5_evaluation.py for evaluating moldel mAP at hw_repo
 cd applications
 python yolov5_evaluation.py
 #mAP @ yolov5s_v2_op9_sig_batch1_input05_640x640_nearest_convert.onnx with(upsampling rearest)
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.346
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.533
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.372
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.196
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.391
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.442
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.279
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.456
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.503
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.320
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.557
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.618
 (3)yolov5_compare_pth_onnx.py for comparing the results of pytorch moldel and onnx model
 cd applications
 python yolov5_compare_pth_onnx.py
 (4)v2 model is in the below link.
 10.200.210.221:/mnt/models/Object_models/YOLOv5/yolov5s_v2_state_dict_input05.pt
 10.200.210.221:/mnt/models/Object_models/YOLOv5/yolov5s_v2_op9_sig_batch1_input05_640x640_nearest_convert.onnx
 (5)the parameters setting
 (5.1)In order to get high mAP in coco val2017, please use 
 101620_yolov5_init_params.json
 {
    "model_path": "/mnt/models/Object_models/YOLOv5/yolov5s_v2_state_dict_input05.pt",
    "grid20_path": "/mnt/models/Object_models/YOLOv5/20_640x640.npy",
    "grid40_path": "/mnt/models/Object_models/YOLOv5/40_640x640.npy",
    "grid80_path": "/mnt/models/Object_models/YOLOv5/80_640x640.npy",
    "num_classes": 80,
    "imgsz_h": 640,
 	"imgsz_w": 640,
 	"conf_thres": 0.001,
 	"iou_thres": 0.65,
    "top_k_num": 3000
 }
 (5.2)For video usage scenarios, please use
 102320_yolov5_init_params.json
 {
    "model_path": "/mnt/models/Object_models/YOLOv5/yolov5s_v2_state_dict_input05.pt",
    "grid20_path": "/mnt/models/Object_models/YOLOv5/20_640x352.npy",
    "grid40_path": "/mnt/models/Object_models/YOLOv5/40_640x352.npy",
    "grid80_path": "/mnt/models/Object_models/YOLOv5/80_640x352.npy",
    "num_classes": 80,
    "imgsz_h": 352,
 	"imgsz_w": 640,
 	"conf_thres": 0.3,
 	"iou_thres": 0.5,
    "top_k_num": 3000
 }
 (5.3)The differences of above setting are
 (5.3.1) Video uses input (640w*352h) to run faster.
 Coco has high or flat wide images, so it is better to use input (640w*640h) 
 (5.3.2) Using the yolov5 official website setting test coco val2017, the confidence setting is low "conf_thres": 0.001, and the iou setting of NMS is high "iou_thres": 0.65, which gets a better mAP.
 But running video needs to be set to "conf_thres": 0.3, so that there are not too many false positives, and the iou setting of NMS "iou_thres": 0.5 is more friendly to close objects
--- a/kneron/exporting/yolov5/yolo_v2.py
+++ b/kneron/exporting/yolov5/yolo_v2.py
@ -0,0 +1,164 @@
 import argparse
 from copy import deepcopy
 import torch
 #from experimental import *
 from .common import *
 #from .common_v3 import *
 from pathlib import Path
 import math
 import yaml
 class Detect(nn.Module):
    def __init__(self, nc=80, anchors=(), ch=()):  # detection layer
        super(Detect, self).__init__()
        self.stride = None  # strides computed during build
        self.nc = nc  # number of classes
        self.no = nc + 5  # number of outputs per anchor
        self.nl = len(anchors)  # number of detection layers
        self.na = len(anchors[0]) // 2  # number of anchors
        self.grid = [torch.zeros(1)] * self.nl  # init grid
        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
        self.register_buffer('anchors', a)  # shape(nl,na,2)
        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
        self.export = False  # onnx export
    def forward(self, x):
        # x = x.copy()  # for profiling
        z = []  # inference output
        self.training |= self.export
        for i in range(self.nl):
            x[i] = self.m[i](x[i])  # conv
            # bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
            # x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
            x[i] = x[i].sigmoid()
        # return x if self.training else (torch.cat(z, 1), x)
        return x
    @staticmethod
    def _make_grid(nx=20, ny=20):
        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
 class Model(nn.Module):
    def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None):  # model, input channels, number of classes
        super(Model, self).__init__()   
        with open(cfg) as f:
            self.yaml = yaml.load(f, Loader=yaml.FullLoader)  # model dict
        # Define model
        if nc and nc != self.yaml['nc']:
            print('Overriding %s nc=%g with nc=%g' % (cfg, self.yaml['nc'], nc))
            self.yaml['nc'] = nc  # override yaml value
        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist, ch_out
        # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))])
        # Build strides, anchors
        m = self.model[-1]  # Detect()
        if isinstance(m, Detect):
            s = 128  # 2x min stride
            #m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
            # FocusNoSliceCat
            m.stride = torch.tensor([8.0,16.0,32.0])  # forward
            m.anchors /= m.stride.view(-1, 1, 1)
            check_anchor_order(m)
            self.stride = m.stride
            self._initialize_biases()  # only run once
            # print('Strides: %s' % m.stride.tolist())
        # Init weights, biases
        initialize_weights(self)
    def forward(self, x, augment=False, profile=False):
        y, dt = [], []  # outputs
        for m in self.model:
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            x = m(x)  # run
            y.append(x if m.i in self.save else None)  # save output
        return x
    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
        m = self.model[-1]  # Detect() module
        for mi, s in zip(m.m, m.stride):  #  from
            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
            b[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
            b[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
 def make_divisible(x, divisor):
    # Returns x evenly divisble by divisor
    return math.ceil(x / divisor) * divisor
 def check_anchor_order(m):
    # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
    a = m.anchor_grid.prod(-1).view(-1)  # anchor area
    da = a[-1] - a[0]  # delta a
    ds = m.stride[-1] - m.stride[0]  # delta s
    if da.sign() != ds.sign():  # same order
        print('Reversing anchor order')
        m.anchors[:] = m.anchors.flip(0)
        m.anchor_grid[:] = m.anchor_grid.flip(0)
 def initialize_weights(model):
    for m in model.modules():
        t = type(m)
        if t is nn.Conv2d:
            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
        elif t is nn.BatchNorm2d:
            m.eps = 1e-3
            m.momentum = 0.03
        elif t in [nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
            m.inplace = True
 def parse_model(d, ch):  # model_dict, input_channels(3) #original
    #print('\n%3s%18s%3s%10s  %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))
    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
        m = eval(m) if isinstance(m, str) else m  # eval strings
        for j, a in enumerate(args):
            try:
                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
            except:
                pass
        n = max(round(n * gd), 1) if n > 1 else n  # depth gain
        if m in [nn.Conv2d, Conv, Bottleneck, SPP, DWConv,  Focus,  BottleneckCSP]:
            #print('*m',m)
            c1, c2 = ch[f], args[0]
            c2 = make_divisible(c2 * gw, 8) if c2 != no else c2
            args = [c1, c2, *args[1:]]
            if m in [BottleneckCSP]:
                args.insert(2, n)
                n = 1
        elif m is nn.BatchNorm2d:
            args = [ch[f]]
        elif m is Concat:
            c2 = sum([ch[-1 if x == -1 else x + 1] for x in f])
        elif m is Detect:
            args.append([ch[x + 1] for x in f])
            if isinstance(args[1], int):  # number of anchors
                args[1] = [list(range(args[1] * 2))] * len(f)
        else:
            c2 = ch[f]
        m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args)  # module
        t = str(m)[8:-2].replace('__main__.', '')  # module type
        np = sum([x.numel() for x in m_.parameters()])  # number params
        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
        #print('%3s%18s%3s%10.0f  %-40s%-30s' % (i, f, n, np, t, args))  # print
        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
        layers.append(m_)
        ch.append(c2)
    return nn.Sequential(*layers), sorted(save)
--- a/kneron/exporting/yolov5/yolov5_postprocess.py
+++ b/kneron/exporting/yolov5/yolov5_postprocess.py
@ -0,0 +1,334 @@
 # coding: utf-8
 import torch
 import torchvision
 import time
 import numpy as np
 import sys
 np.set_printoptions(threshold=sys.maxsize)
 def box_iou(box1, box2):
    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
    """
    Return intersection-over-union (Jaccard index) of boxes.
    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
    Arguments:
        box1 (Tensor[N, 4])
        box2 (Tensor[M, 4])
    Returns:
        iou (Tensor[N, M]): the NxM matrix containing the pairwise
            IoU values for every element in boxes1 and boxes2
    """
    def box_area(box):
        # box = 4xn
        return (box[2] - box[0]) * (box[3] - box[1])
    area1 = box_area(box1.t())
    area2 = box_area(box2.t())
    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
    return inter / (area1[:, None] + area2 - inter)  # iou = inter / (area1 + area2 - inter)
 def xywh2xyxy(x):
    # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
    y = torch.zeros_like(x) if isinstance(x, torch.Tensor) else np.zeros_like(x)
    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
    return y
 def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, top_k_num=3000, merge=False, classes=None, agnostic=False):
    """Performs Non-Maximum Suppression (NMS) on inference results
    Returns:
         detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
    """
    # print('conf_thres',conf_thres)
    if prediction.dtype is torch.float16:
        prediction = prediction.float()  # to FP32
    nc = prediction[0].shape[1] - 5  # number of classes
    xc = prediction[..., 4] > conf_thres  # candidates
    # Settings
    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
    max_det = 300  # maximum number of detections per image
    time_limit = 10.0  # seconds to quit after
    multi_label = nc > 1  # multiple labels per box (adds 0.5ms/img)
    t = time.time()
    output = [None] * prediction.shape[0]
    for xi, x in enumerate(prediction):  # image index, image inference
        # Apply constraints
        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
        x = x[xc[xi]]  # confidence
        # If none remain process next image
        if not x.shape[0]:
            continue
        # Compute conf
        x[:, 5:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
        box = xywh2xyxy(x[:, :4])
        # Detections matrix nx6 (xyxy, conf, cls)
        if multi_label:
            i, j = (x[:, 5:] > conf_thres).nonzero().t()
            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
        else:  # best class only
            conf, j = x[:, 5:].max(1, keepdim=True)
            x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
        # If none remain process next image
        n = x.shape[0]  # number of boxes
        if not n:
            continue
        # Sort by confidence
        # x = x[x[:, 4].argsort(descending=True)]
        # Batched NMS
        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
        # # Sort by confidence
        ind_Sort_by_confidence = x[:, 4].argsort(descending=True)
        boxes = boxes[ind_Sort_by_confidence][:top_k_num] #
        scores = scores[ind_Sort_by_confidence][:top_k_num] #
        x = x[ind_Sort_by_confidence][:top_k_num] #
        # cross classes nms
        i = torchvision.ops.boxes.nms(boxes, scores, iou_thres)
        if i.shape[0] > max_det:  # limit detections
            i = i[:max_det]
        # if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
        #     update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
        #     iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
        #     weights = iou * scores[None]  # box weights
        #     x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
        #     if redundant:
        #         i = i[iou.sum(1) > 1]  # require redundancy
        output[xi] = x[i]
        if (time.time() - t) > time_limit:
            break  # time limit exceeded
    return output
 def non_max_suppression_kneron(prediction, conf_thres=0.1, iou_thres=0.6, top_k_num=3000, merge=False, classes=None, agnostic=False):
    """Performs Non-Maximum Suppression (NMS) on inference results
    Returns:
         detections with shape: nx6 (x1, y1, x2, y2, conf, cls)
    """
    if prediction.dtype is torch.float16:
        prediction = prediction.float()  # to FP32
    nc = prediction[0].shape[1] - 5  # number of classes
    xc = prediction[..., 4] > conf_thres  # candidates
    # Settings
    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
    max_det = 300  # maximum number of detections per image
    time_limit = 10.0  # seconds to quit after
    multi_label = nc > 1  # multiple labels per box (adds 0.5ms/img)
    t = time.time()
    output = [None] * prediction.shape[0]
    for xi, x in enumerate(prediction):  # image index, image inference
        # Apply constraints
        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
        x = x[xc[xi]]  # confidence
        # If none remain process next image
        if not x.shape[0]:
            continue
        # Compute conf
        x[:, 5:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
        box = xywh2xyxy(x[:, :4])
        # Detections matrix nx6 (xyxy, conf, cls)
        if multi_label:
            i, j = (x[:, 5:] > conf_thres).nonzero().t()
            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
        else:  # best class only
            conf, j = x[:, 5:].max(1, keepdim=True)
            x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
        # If none remain process next image
        n = x.shape[0]  # number of boxes
        if not n:
            continue
        # Sort by confidence
        # x = x[x[:, 4].argsort(descending=True)]
        # Batched NMS
        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
        # Sort by confidence
        ind_Sort_by_confidence = x[:, 4].argsort(descending=True)
        boxes = boxes[ind_Sort_by_confidence][:top_k_num] #
        scores = scores[ind_Sort_by_confidence][:top_k_num] #
        x = x[ind_Sort_by_confidence][:top_k_num] #
        # cross classes nms
        i = torchvision.ops.boxes.nms(boxes, scores, iou_thres)
        if i.shape[0] > max_det:  # limit detections
            i = i[:max_det]
        output[xi] = x[i]
        if (time.time() - t) > time_limit:
            break  # time limit exceeded
    return output
 def clip_coords(boxes, img_shape):
    # Clip bounding xyxy bounding boxes to image shape (height, width)
    boxes[:, 0].clamp_(0, img_shape[1])  # x1
    boxes[:, 1].clamp_(0, img_shape[0])  # y1
    boxes[:, 2].clamp_(0, img_shape[1])  # x2
    boxes[:, 3].clamp_(0, img_shape[0])  # y2
 def scale_coords_ori(img1_shape, coords, img0_shape, ratio_pad=None):
    # Rescale coords (xyxy) from img1_shape to img0_shape
    if ratio_pad is None:  # calculate from img0_shape
        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
        pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
    else:
        gain = ratio_pad[0][0]
        pad = ratio_pad[1]
    coords[:, [0, 2]] -= pad[0]  # x padding
    coords[:, [1, 3]] -= pad[1]  # y padding
    coords[:, :4] /= gain
    clip_coords(coords, img0_shape)
    return coords
 def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None):
    # Rescale coords (xyxy) from img1_shape to img0_shape
    if ratio_pad is None:  # calculate from img0_shape
        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])  # gain  = old / new
        #pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2  # wh padding
    else:
        gain = ratio_pad[0][0]
        #pad = ratio_pad[1]
    # coords[:, [0, 2]] -= pad[0]  # x padding
    # coords[:, [1, 3]] -= pad[1]  # y padding
    coords[:, :4] /= gain
    clip_coords(coords, img0_shape)
    return coords
 def scale_coords_test(img1_shape, coords, img0_shape, ratio_pad=None):
    coords[:, 0] /= (img1_shape[1] / img0_shape[1])
    coords[:, 2] /= (img1_shape[1] / img0_shape[1])
    coords[:, 1] /= (img1_shape[0] / img0_shape[0])
    coords[:, 3] /= (img1_shape[0] / img0_shape[0])
    clip_coords(coords, img0_shape)
    return coords
 def classes_mapping(det, num_classes):
    det[:, 5] = det[:, 5] + 1.0
 def Yolov5_postprocess(pred, img_shape, im0_shape, conf_thres, iou_thres, top_k_num, num_classes, vanish_point, e2e_coco) :
    classes, agnostic_nms = None, False#
    img_h = im0_shape[0]
    vanish_y2 = vanish_point * float(img_h)
    # Apply NMS
    pred = non_max_suppression(pred, conf_thres, iou_thres, top_k_num, classes=classes, agnostic=agnostic_nms)
    #return pred
    dets = []
    for i, det in enumerate(pred):  # detections per image
        gn = torch.tensor(im0_shape)[[1, 0, 1, 0]]  # normalization gain whwh
        if det is not None and len(det):
            # Rescale boxes from img_size to im0 size
            det[:, :4] = scale_coords(img_shape[2:], det[:, :4], im0_shape).round()
            det = det[det[:,3]>=vanish_y2]
            # (x1,y1,x2,y2) -> (x1,y1,w,h) for public_field.py
            det[:, 2] = det[:, 2] - det[:, 0] 
            det[:, 3] = det[:, 3] - det[:, 1]  
            # classes(0~79) -> classes(1~80) for public_field.py
            if e2e_coco:
                classes_mapping(det, num_classes)
            det = det.cpu().numpy()  
            dets.append(det)                                
    if dets and len(dets) > 0:
        dets = np.asarray(dets)
        dets = np.squeeze(dets, axis=0) # remove outer []
        dets = dets.tolist()
    return dets
 def make_grid(nx=20, ny=20):
    yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
    grids = torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
    return grids
 def Yolov5_postprocess_onnx_sig(out,img_shape, im0_shape, conf_thres, iou_thres, top_k_num, grids, num_classes, anchors,vanish_point, e2e_coco) :
    nc = num_classes  # number of classes
    no = nc + 5  # number of outputs per anchor    
    nl = len(anchors)  # number of detection layers
    na = len(anchors[0]) // 2  # number of anchors
    a = torch.tensor(anchors).float().view(3, -1, 2)
    anchor_grid = a.clone().view(3, 1, -1, 1, 1, 2)
    stride = torch.tensor([ 8., 16., 32.])     
    z = []
    for i in range(nl):
        x = torch.from_numpy(out[i])
        # print('x.shape',x.shape)
        bs, _, ny, nx = x.shape  # x(bs,3,20,20,85)
        x = x.view(bs, na, no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
        # grid_r = make_grid(nx, ny) ##grid
        # grid_r = grid_r.numpy() ##grid
        # file_name = str(i)+'.npy' ##grid
        # np.save(file_name,grid_r) ##grid
        grid = grids[i]#
        #y = x.sigmoid()
        y = x
        y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + grid) * stride[i]  # xy
        y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * anchor_grid[i]  # wh
        z.append(y.view(bs, -1, no))
    pred = torch.cat(z, 1)
    return Yolov5_postprocess(pred, img_shape, im0_shape, conf_thres, iou_thres, top_k_num, num_classes,vanish_point, e2e_coco)
 def Yolov5_postprocess_sig(out,img_shape, im0_shape, conf_thres, iou_thres, top_k_num, grids, num_classes, anchors,vanish_point, e2e_coco) :
    nc = num_classes  # number of classes
    no = nc + 5  # number of outputs per anchor    
    nl = len(anchors)  # number of detection layers
    na = len(anchors[0]) // 2  # number of anchors
    a = torch.tensor(anchors).float().view(3, -1, 2)
    anchor_grid = a.clone().view(3, 1, -1, 1, 1, 2).to(out[0].device)
    stride = torch.tensor([ 8., 16., 32.]).to(out[0].device)     
    z = []
    for i in range(nl):
        x = out[i]
        bs, _, ny, nx = x.shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
        # print('x.shape',x.shape)
        x = x.view(bs, na, no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
        # grid_r = make_grid(nx, ny) ##grid
        # grid_r = grid_r.numpy() ##grid
        # file_name = str(i)+'.npy' ##grid
        # np.save(file_name,grid_r) ##grid
        grid = grids[i].to(out[0].device) #
        #y = x.sigmoid()
        y = x
        y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + grid) * stride[i]  # xy
        y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * anchor_grid[i]  # wh
        z.append(y.view(bs, -1, no))   
    # exit(0)
    pred = torch.cat(z, 1)
    return Yolov5_postprocess(pred, img_shape, im0_shape, conf_thres, iou_thres, top_k_num, num_classes,vanish_point, e2e_coco)
--- a/kneron/exporting/yolov5/yolov5_preprocess.py
+++ b/kneron/exporting/yolov5/yolov5_preprocess.py
@ -0,0 +1,160 @@
 # coding: utf-8
 import torch
 import cv2
 import numpy as np
 import math
 import time
 from . import kneron_preprocessing
 kneron_preprocessing.API.set_default_as_520()
 torch.backends.cudnn.deterministic = True
 img_formats = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.tiff', '.dng']
 def make_divisible(x, divisor):
    # Returns x evenly divisble by divisor
    return math.ceil(x / divisor) * divisor
 def check_img_size(img_size, s=32):
    # Verify img_size is a multiple of stride s
    new_size = make_divisible(img_size, int(s))  # ceil gs-multiple
    if new_size != img_size:
        print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size))
    return new_size
 def letterbox_ori(img, new_shape=(640, 640), color=(0, 0, 0), auto=True, scaleFill=False, scaleup=True):
    # Resize image to a 32-pixel-multiple rectangle https://github.com/ultralytics/yolov3/issues/232
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)
    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)
    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) # width, height 
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    dw /= 2  # divide padding into 2 sides
    dh /= 2
    if shape[::-1] != new_unpad:  # resize
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
        #img = kneron_preprocessing.API.resize(img,size=new_unpad, keep_ratio = False)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    # top, bottom = int(0), int(round(dh + 0.1))
    # left, right = int(0), int(round(dw + 0.1))    
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    #img = kneron_preprocessing.API.pad(img, left, right, top, bottom, 0)
    return img, ratio, (dw, dh)
 def letterbox(img, new_shape=(640, 640), color=(0, 0, 0), auto=True, scaleFill=False, scaleup=True):
    # Resize image to a 32-pixel-multiple rectangle https://github.com/ultralytics/yolov3/issues/232
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)
    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)
    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) # width, height 
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    # dw /= 2  # divide padding into 2 sides
    # dh /= 2
    if shape[::-1] != new_unpad:  # resize
        #img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
        img = kneron_preprocessing.API.resize(img,size=new_unpad, keep_ratio = False)
    # top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    # left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    top, bottom = int(0), int(round(dh + 0.1))
    left, right = int(0), int(round(dw + 0.1))    
    #img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    img = kneron_preprocessing.API.pad(img, left, right, top, bottom, 0)
    return img, ratio, (dw, dh)
 def letterbox_test(img, new_shape=(640, 640), color=(0, 0, 0), auto=True, scaleFill=False, scaleup=True):
    ratio = 1.0, 1.0
    dw, dh = 0, 0
    img = kneron_preprocessing.API.resize(img, size=(480, 256), keep_ratio=False, type='bilinear')
    return img, ratio, (dw, dh)
 def LoadImages(path,img_size):  #_rgb # for inference
    if isinstance(path, str):
        img0 = cv2.imread(path)  # BGR       
    else:
        img0 = path  # BGR
    # Padded resize
    img = letterbox(img0, new_shape=img_size)[0]
    # Convert
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    img = np.ascontiguousarray(img)
    return img, img0
 def LoadImages_yyy(path,img_size): #_yyy # for inference
    if isinstance(path, str):
        img0 = cv2.imread(path)  # BGR       
    else:
        img0 = path  # BGR
    yvu = cv2.cvtColor(img0, cv2.COLOR_BGR2YCrCb)
    y, v, u = cv2.split(yvu)
    img0 = np.stack((y,)*3, axis=-1)
    # Padded resize
    img = letterbox(img0, new_shape=img_size)[0]
    # Convert
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    img = np.ascontiguousarray(img)
    return img, img0
 def LoadImages_yuv420(path,img_size):  #_yuv420 # for inference 
    if isinstance(path, str):
        img0 = cv2.imread(path)  # BGR       
    else:
        img0 = path  # BGR
    img_h, img_w = img0.shape[:2]
    img_h = (img_h // 2) * 2
    img_w = (img_w // 2) * 2
    img = img0[:img_h,:img_w,:]
    yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV_I420)
    img0= cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR_I420) #yuv420
    # Padded resize
    img = letterbox(img0, new_shape=img_size)[0]
    # Convert
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    img = np.ascontiguousarray(img)
    return img, img0
 def Yolov5_preprocess(image_path, device, imgsz_h, imgsz_w) : 
    model_stride_max = 32
    imgsz_h = check_img_size(imgsz_h, s=model_stride_max)  # check img_size
    imgsz_w = check_img_size(imgsz_w, s=model_stride_max)  # check img_size
    img, im0 = LoadImages(image_path, img_size=(imgsz_h,imgsz_w))
    img = kneron_preprocessing.API.norm(img) #path1
    #print('img',img.shape)
    img = torch.from_numpy(img).to(device) #path1,path2
    # img = img.float()  # uint8 to fp16/32 #path2
    # img /= 255.0#256.0 - 0.5 # 0 - 255 to -0.5 - 0.5 #path2
    if img.ndimension() == 3:
        img = img.unsqueeze(0)
    return img, im0
--- a/kneron/exporting/yolov5/yolov5_runner.py
+++ b/kneron/exporting/yolov5/yolov5_runner.py
@ -0,0 +1,91 @@
 import torch
 torch.set_printoptions(precision=10)
 torch.set_printoptions(threshold=99999999999)
 torch.backends.cudnn.deterministic = True
 from .yolov5_preprocess import *
 from .yolov5_postprocess import *
 from .yolo_v2 import Model as Model_v2
 import onnxruntime
 import time
 import os
 from collections import Counter
 import torch.nn.functional as F
 import random
 from pathlib import Path
 class Yolov5Runner:
    def __init__(self, model_path, yaml_path, grid20_path, grid40_path, grid80_path, num_classes, imgsz_h, imgsz_w, conf_thres, iou_thres, top_k_num, vanish_point, **kwargs):#is_onnx,
        """
        inputs :
            model_path : str ,path to model 
        """
        self.model_path = model_path
        self.imgsz_h = imgsz_h
        self.imgsz_w = imgsz_w
        self.conf_thres = conf_thres
        self.iou_thres = iou_thres
        self.top_k_num = top_k_num
        self.vanish_point = vanish_point
        self.num_classes = num_classes
        self.DEVICE = torch.device("cpu")#torch.device('cuda:0')#
        self.grid20 = torch.from_numpy(np.load(grid20_path))
        self.grid40 = torch.from_numpy(np.load(grid40_path))
        self.grid80 = torch.from_numpy(np.load(grid80_path)) 
        self.grids = [self.grid80, self.grid40, self.grid20]        
        if 'onnx' not in model_path:
            self.yolov5_model = Model_v2(yaml_path, nc=num_classes)                    
            self.yolov5_model.load_state_dict(torch.load(model_path, map_location=self.DEVICE))#,strict=False)
            self.yolov5_model.float().eval()
            self.yolov5_model.to(self.DEVICE)
            self.yolov5_model.eval()
        else:
            #onnxruntime.set_default_logger_severity(0)
            self.sess = onnxruntime.InferenceSession(model_path)
            # self.sess.set_providers(['CUDAExecutionProvider'])
            self.input_name = self.sess.get_inputs()[0].name
            self.onnx_batch_size = self.sess.get_inputs()[0].shape[0]
            self.onnx_img_size_h = self.sess.get_inputs()[0].shape[2]
            self.onnx_img_size_w = self.sess.get_inputs()[0].shape[3]  
        self.anchors = [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]] #yolov5            
        print('self.vanish_point',self.vanish_point)                                                               
        self.e2e_coco = kwargs.get('e2e_coco', False)
    def run(self, img_path):
        """
        inputs : 
            img_path : path of the image
        outputs :
            dets : list 
        """
        self.yolov5_model.eval()
        with torch.no_grad():
            img, im0 = Yolov5_preprocess(img_path, self.DEVICE, self.imgsz_h, self.imgsz_w)
            if next(self.yolov5_model.parameters()).is_cuda:
                img = img.type(torch.cuda.FloatTensor)
            else:
                img = img.type(torch.FloatTensor)
            pred = self.yolov5_model(img, augment=False)
            img_shape, im0_shape = img.shape, im0.shape               
            dets = Yolov5_postprocess_sig(pred,img_shape, im0_shape, self.conf_thres, self.iou_thres, self.top_k_num, self.grids, self.num_classes, self.anchors,self.vanish_point)
            return dets 
    def run_onnx(self, img_path):
        """
        inputs : 
            img_path : path of the image
        outputs :
            dets : list 
        """
        with torch.no_grad():
            img, im0 = Yolov5_preprocess(img_path, self.DEVICE, self.imgsz_h, self.imgsz_w)
            np_images = np.array(img.cpu())
            np_images = np_images.astype(np.float32)
            pred_onnx = self.sess.run(None, {self.input_name: np_images })
            img_shape, im0_shape = img.shape, im0.shape 
            # print('img_shape',img_shape)
            # print('im0_shape', im0_shape)
            dets_onnx = Yolov5_postprocess_onnx_sig(pred_onnx,img_shape, im0_shape, self.conf_thres, self.iou_thres, self.top_k_num, self.grids, self.num_classes, self.anchors,self.vanish_point, self.e2e_coco)
            return dets_onnx
--- a/kneron/exporting/yolov5_export.py
+++ b/kneron/exporting/yolov5_export.py
@ -0,0 +1,80 @@
 import os
 import torch
 import sys
 import yaml
 import argparse
 from yolov5.yolov5_runner import Yolov5Runner
 def save_weight(num_classes): 
    current_path=os.getcwd()
    par_path = os.path.dirname(current_path)
    sys.path.append(os.path.join(par_path, 'yolov5'))
    from models.yolo import Model  
    num_classes = num_classes 
    device=torch.device('cpu')
    ckpt = torch.load(path, map_location=device)
    model = Model(yaml_path, nc=num_classes)
    ckpt['model'] = {k: v for k, v in ckpt['model'].float().state_dict().items() if k in model.state_dict() and model.state_dict()[k].shape == v.shape}
    model.load_state_dict(ckpt['model'])
    torch.save(model.state_dict(),pt_path,_use_new_zipfile_serialization=False)
 def export_onnx(input_h, input_w, num_classes):
    onnx_batch_size, onnx_img_h, onnx_img_w = 1, input_h, input_w
    yolov5_model = Yolov5Runner(model_path=pt_path, yaml_path=yaml_path, grid20_path=grid20_path, grid40_path=grid40_path, grid80_path=grid80_path, num_classes=num_classes, imgsz_h=onnx_img_h, imgsz_w=onnx_img_w, conf_thres=0.001, iou_thres=0.65, top_k_num=3000, vanish_point=0.0) 
    # Input
    img = torch.zeros((onnx_batch_size, 3, onnx_img_h, onnx_img_w))  
    # img = img.type(torch.cuda.FloatTensor)
    # Load PyTorch model
    model = yolov5_model.yolov5_model
    model.eval()
    model.model[-1].export = True  # set Detect() layer export=True
    y = model(img)  # dry run
    # ONNX export
    try:
        import onnx
        print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
        print('****onnx file****',onnx_export_file)
        torch.onnx.export(model, img, onnx_export_file, verbose=False, opset_version=11, keep_initializers_as_inputs=True, input_names=['images'], output_names=['classes', 'boxes'] if y is None else ['output'])
        # Checks
        onnx_model = onnx.load(onnx_export_file)  # load onnx model
        onnx.checker.check_model(onnx_model)  # check onnx model
        print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
        print('ONNX export success, saved as %s' % onnx_export_file)
    except Exception as e:
        print('ONNX export failure: %s' % e)
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--data', type=str, default='../yolov5/data/pretrained_paths_520.yaml', help='the path to pretrained model paths yaml file')
    args = parser.parse_args()
    with open(args.data) as f:
        data_dict = yaml.load(f, Loader=yaml.FullLoader)  # data dict
    os.environ["CUDA_VISIBLE_DEVICES"] = '0'
    num_classes = data_dict['nc']
    input_w = data_dict['input_w']
    input_h = data_dict['input_h']
    grid_dir = data_dict['grid_dir']
    grid20_path = data_dict['grid20_path']
    grid40_path = data_dict['grid40_path']
    grid80_path = data_dict['grid80_path']
    path = data_dict['path']
    pt_path=data_dict['pt_path']
    yaml_path=data_dict['yaml_path']
    onnx_export_file = data_dict['onnx_export_file']
    save_weight(num_classes)
    export_onnx(input_h, input_w, num_classes)
--- a/kneron/inference.py
+++ b/kneron/inference.py
@ -0,0 +1,64 @@
 import os
 import sys
 import argparse
 import yaml
 import cv2
 import numpy as np
 def draw(img_path, bboxes, save_path = None, names = None):
    img = cv2.imread(img_path)
    for bbox in bboxes:
        l,t,w,h,score,class_id=bbox
        if names is not None:
            class_id = names[int(class_id)]
        img = cv2.rectangle(img,(int(l),int(t)),(int(l+w),int(t+h)),(0, 255, 0),6)
        text = "{}".format(class_id) + "  {}".format(np.round(score, 3))
        img = cv2.putText(img, text, (int(l), int(t)-10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
    if save_path is None:
        save_path = img_path
    output_file = os.path.join(save_path, "output.jpg")  # 確保有圖片副檔名
    cv2.imwrite(output_file, img)
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--img-path', type=str, default=None, help='path to image')
    parser.add_argument('--save-path', type=str, default=None, help='path to save image')
    parser.add_argument('--data', type=str, default='data/pretrained_paths_520.yaml', help='the path to pretrained model paths yaml file')
    parser.add_argument('--conf_thres', type=float, default=0.3, help='confidence threshold')
    parser.add_argument('--iou_thres', type=float, default=0.5, help='iou threshold for NMS')
    parser.add_argument('--onnx', help='inference onnx model',action='store_true')
    args = parser.parse_args()
    par_path = os.path.abspath(os.path.join(os.getcwd(), os.pardir))
    sys.path.append(par_path)
    sys.path.append(os.path.join(par_path, 'exporting') )
    from yolov5.yolov5_runner import Yolov5Runner
    with open(args.data) as f:
        data_dict = yaml.load(f, Loader=yaml.FullLoader)  # data dict
    num_classes = data_dict['nc']
    input_w = data_dict['input_w']
    input_h = data_dict['input_h']
    grid_dir = data_dict['grid_dir']
    grid20_path = data_dict['grid20_path']
    grid40_path = data_dict['grid40_path']
    grid80_path = data_dict['grid80_path']
    path = data_dict['path']
    if args.onnx:
        yolov5_model = Yolov5Runner(model_path=data_dict['onnx_export_file'], yaml_path=data_dict['yaml_path'], grid20_path=grid20_path, grid40_path=grid40_path, grid80_path=grid80_path, num_classes=num_classes, imgsz_h=input_h, imgsz_w=input_w, conf_thres=args.conf_thres, iou_thres=args.iou_thres, top_k_num=3000, vanish_point=0.0) 
        bboxes = yolov5_model.run_onnx(args.img_path)
    else:
        yolov5_model = Yolov5Runner(model_path=data_dict['pt_path'], yaml_path=data_dict['yaml_path'], grid20_path=grid20_path, grid40_path=grid40_path, grid80_path=grid80_path, num_classes=num_classes, imgsz_h=input_h, imgsz_w=input_w, conf_thres=args.conf_thres, iou_thres=args.iou_thres, top_k_num=3000, vanish_point=0.0) 
        bboxes = yolov5_model.run(args.img_path)
    print(bboxes)
    if args.save_path is not None:
        draw(args.img_path, bboxes, save_path = args.save_path, names = data_dict['names'])
--- a/kneron/inference_e2e.py
+++ b/kneron/inference_e2e.py
@ -0,0 +1,53 @@
 import os
 import sys
 import argparse
 import yaml
 from tqdm import tqdm
 import json
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--img-path', type=str, help='Path to the dataset directory.')
    parser.add_argument('--params', type=str, help='Path to the init params file.')
    parser.add_argument('--save-path', type=str, help='Path to save output in json.')
    args = parser.parse_args()
    par_path = os.path.abspath(os.path.join(os.getcwd(), os.pardir))
    sys.path.append(par_path)
    sys.path.append(os.path.join(par_path, 'exporting') )
    from yolov5.yolov5_runner import Yolov5Runner
    with open(args.params, "r", encoding="utf-8") as f:
        params_dict = yaml.load(f, Loader=yaml.FullLoader)  # data dict
    num_classes = params_dict['nc']
    input_w = params_dict['input_w']
    input_h = params_dict['input_h']
    grid20_path = params_dict['grid20_path']
    grid40_path = params_dict['grid40_path']
    grid80_path = params_dict['grid80_path']
    conf_thres = params_dict['conf_thres']
    iou_thres = params_dict['iou_thres']
    model_type = params_dict['model_type']
    e2e_coco = params_dict['e2e_coco']
    if model_type == 'onnx':
        yolov5_model = Yolov5Runner(model_path=params_dict['onnx_path'], yaml_path=params_dict['model_yaml_path'], grid20_path=grid20_path, grid40_path=grid40_path, grid80_path=grid80_path, num_classes=num_classes, imgsz_h=input_h, imgsz_w=input_w, conf_thres=conf_thres, iou_thres=iou_thres, top_k_num=3000, vanish_point=0.0, e2e_coco=e2e_coco)
    else:
        yolov5_model = Yolov5Runner(model_path=params_dict['pt_path'], yaml_path=params_dict['model_yaml_path'], grid20_path=grid20_path, grid40_path=grid40_path, grid80_path=grid80_path, num_classes=num_classes, imgsz_h=input_h, imgsz_w=input_w, conf_thres=conf_thres, iou_thres=iou_thres, top_k_num=3000, vanish_point=0.0, e2e_coco=e2e_coco)
    img_list = os.listdir(args.img_path)
    results = []
    for img_name in tqdm(img_list):
        if img_name.split('.')[-1] not in ['png', 'jpg']:
            continue
        img_path = os.path.join(args.img_path, img_name)
        if model_type == 'onnx':
            bboxes = yolov5_model.run_onnx(img_path)
        else:
            bboxes = yolov5_model.run(img_path)
        results.append({'img_path': img_path, 'bbox': bboxes } )
    with open(args.save_path, 'w') as fp:
        json.dump(results, fp)
--- a/kneron/ktc720.py
+++ b/kneron/ktc720.py
@ -0,0 +1,72 @@
 import ktc
 import numpy as np
 import os
 import onnx
 from PIL import Image
 import torch
 from yolov5_preprocess import Yolov5_preprocess
 import kneron_preprocessing
 onnx_path = 'runs/train/exp24/weights/best_simplified.onnx'
 m = onnx.load(onnx_path)
 m = ktc.onnx_optimizer.onnx2onnx_flow(m)
 onnx.save(m,'latest.opt.onnx')
 km = ktc.ModelConfig(20008, "0001", "720", onnx_model=m)
 eval_result = km.evaluate()
 print("\nNpu performance evaluation result:\n" + str(eval_result))
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 imgsz_h, imgsz_w = 640, 640
 data_path = "data50"
 files_found = [f for _, _, files in os.walk(data_path) for f in files if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp"))]
 if not files_found:
    raise FileNotFoundError(f"❌ Error: No images found in {data_path}! Please check your dataset.")
 print(f"✅ Found {len(files_found)} images in {data_path}")
 # **獲取 ONNX 模型的輸入名稱**
 input_name = m.graph.input[0].name  # 確保 key 與 ONNX input name 一致
 # 存儲預處理後的圖片數據
 img_list = []
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # **只處理圖片文件**
        if not f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
            print(f"⚠️ Skipping non-image file: {fullpath}")
            continue
        # **嘗試處理圖片**
        try:
            img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
            img_data = img_data.cpu().numpy()
            print(f"✅ Processed: {fullpath}")
            img_list.append(img_data)
        except Exception as e:
            print(f"❌ Failed to process {fullpath}: {e}")
 # **確保 img_list 不是空的**
 if not img_list:
    raise ValueError("❌ Error: No valid images were processed! Please check the image paths and formats.")
 # **執行 BIE 量化**
 bie_model_path = km.analysis({input_name: img_list})
 # **確認 BIE 模型是否生成**
 if not os.path.exists(bie_model_path):
    raise RuntimeError(f"❌ Error: BIE model was not generated! Please check your quantization process.")
 # 顯示成功訊息
 print("\n✅ Fixed-point analysis done! BIE model saved to:", bie_model_path)
 # 確保 `km` 已經初始化，並且 `.bie` 模型已生成
 nef_model_path = ktc.compile([km])
 # 顯示成功訊息
 print("\n✅ Compile done! NEF file saved to:", nef_model_path)
--- a/kneron/oldquantize_yolov5.py
+++ b/kneron/oldquantize_yolov5.py
@ -0,0 +1,33 @@
 import os
 import numpy as np
 import torch
 from yolov5_preprocess import Yolov5_preprocess  # 使用你的預處理
 import kneron_preprocessing
 # 設定裝置
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 # 設定圖片大小（與訓練時一致）
 imgsz_h, imgsz_w = 640, 640
 # 量化數據集目錄（請確保這個資料夾存在）
 data_path = "/data50"
 img_list = []
 # 遍歷 voc_data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # 執行與訓練相同的預處理
        img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
        print(f"Processed: {fullpath}")
        img_list.append(img_data)
 # 轉為 NumPy 格式
 img_list = np.array(img_list)
 # 執行 BIE 量化分析
 bie_model_path = km.analysis({"input": img_list})
 print("\nFixed-point analysis done. Saved bie model to '" + str(bie_model_path) + "'")
--- a/kneron/onnx2nef520.py
+++ b/kneron/onnx2nef520.py
@ -0,0 +1,110 @@
 import ktc
 import numpy as np
 import os
 import onnx
 import shutil
 from PIL import Image
 import torch
 from yolov5_preprocess import Yolov5_preprocess
 import kneron_preprocessing
 # 設定 ONNX 模型儲存路徑
 onnx_dir = 'runs/train/exp24/weights/'
 onnx_path = os.path.join(onnx_dir, 'best_no_sigmoid.onnx')
 # 確保目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 加載並優化 ONNX 模型
 m = onnx.load(onnx_path)
 m = ktc.onnx_optimizer.onnx2onnx_flow(m)
 opt_onnx_path = os.path.join(onnx_dir, 'latest.opt.onnx')
 onnx.save(m, opt_onnx_path)
 km = ktc.ModelConfig(20008, "0001", "520", onnx_model=m)
 eval_result = km.evaluate()
 print("\nNpu performance evaluation result:\n" + str(eval_result))
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 imgsz_h, imgsz_w = 640, 640
 data_path = "datacoin"
 files_found = [f for _, _, files in os.walk(data_path) for f in files if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp"))]
 if not files_found:
    raise FileNotFoundError(f"\u274c Error: No images found in {data_path}! Please check your dataset.")
 print(f"\u2705 Found {len(files_found)} images in {data_path}")
 # 獲取 ONNX 模型的輸入名稱
 input_name = m.graph.input[0].name  # 確保 key 與 ONNX input name 一致
 # 存儲預處理後的圖片數據
 img_list = []
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # 只處理圖片文件
        if not f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
            print(f"\u26a0\ufe0f Skipping non-image file: {fullpath}")
            continue
        # 嘗試處理圖片
        try:
            img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
            img_data = img_data.cpu().numpy()
            print(f"\u2705 Processed: {fullpath}")
            img_list.append(img_data)
        except Exception as e:
            print(f"\u274c Failed to process {fullpath}: {e}")
 # 確保 img_list 不是空的
 if not img_list:
    raise ValueError("\u274c Error: No valid images were processed! Please check the image paths and formats.")
 # 執行 BIE 量化
 bie_model_path = km.analysis({input_name: img_list})
 # 確保 BIE 檔案儲存到指定目錄
 bie_save_path = os.path.join(onnx_dir, os.path.basename(bie_model_path))
 shutil.copy(bie_model_path, bie_save_path)  # 使用 shutil.move 來處理跨磁碟移動
 # 確認 BIE 模型是否生成
 if not os.path.exists(bie_save_path):
    raise RuntimeError(f"\u274c Error: BIE model was not generated! Please check your quantization process.")
 print("\n\u2705 Fixed-point analysis done! BIE model saved to:", bie_save_path)
 # 確保 `km` 已經初始化，並且 `.bie` 模型已生成
 nef_model_path = ktc.compile([km])
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 NEF 目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 .nef 檔案存在
 if not os.path.exists(nef_model_path):
    raise RuntimeError(f"❌ Error: NEF model was not generated at {nef_model_path}! Please check your compilation process.")
 # 確保 NEF 檔案儲存到指定目錄
 nef_save_path = os.path.join(onnx_dir, os.path.basename(nef_model_path))
 if os.path.exists(nef_model_path):
    shutil.copy(nef_model_path, nef_save_path)
 else:
    raise RuntimeError(f"❌ Error: NEF model was expected at {nef_model_path}, but it does not exist!")  # 同樣使用 shutil.move
 if not os.path.exists(nef_save_path):
    raise RuntimeError(f"\u274c Error: NEF model was not generated! Please check your compilation process.")
 print("\n\u2705 Compile done! NEF file saved to:", nef_save_path)
--- a/kneron/onnx2nef630.py
+++ b/kneron/onnx2nef630.py
@ -0,0 +1,110 @@
 import ktc
 import numpy as np
 import os
 import onnx
 import shutil
 from PIL import Image
 import torch
 from yolov5_preprocess import Yolov5_preprocess
 import kneron_preprocessing
 # 設定 ONNX 模型儲存路徑
 onnx_dir = 'runs/train/exp29/weights/'
 onnx_path = os.path.join(onnx_dir, 'best_simplified.onnx')
 # 確保目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 加載並優化 ONNX 模型
 m = onnx.load(onnx_path)
 m = ktc.onnx_optimizer.onnx2onnx_flow(m)
 opt_onnx_path = os.path.join(onnx_dir, 'latest.opt.onnx')
 onnx.save(m, opt_onnx_path)
 km = ktc.ModelConfig(20008, "0001", "630", onnx_model=m)
 eval_result = km.evaluate()
 print("\nNpu performance evaluation result:\n" + str(eval_result))
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 imgsz_h, imgsz_w = 640, 640
 data_path = "data4"
 files_found = [f for _, _, files in os.walk(data_path) for f in files if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp"))]
 if not files_found:
    raise FileNotFoundError(f"\u274c Error: No images found in {data_path}! Please check your dataset.")
 print(f"\u2705 Found {len(files_found)} images in {data_path}")
 # 獲取 ONNX 模型的輸入名稱
 input_name = m.graph.input[0].name  # 確保 key 與 ONNX input name 一致
 # 存儲預處理後的圖片數據
 img_list = []
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # 只處理圖片文件
        if not f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
            print(f"\u26a0\ufe0f Skipping non-image file: {fullpath}")
            continue
        # 嘗試處理圖片
        try:
            img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
            img_data = img_data.cpu().numpy()
            print(f"\u2705 Processed: {fullpath}")
            img_list.append(img_data)
        except Exception as e:
            print(f"\u274c Failed to process {fullpath}: {e}")
 # 確保 img_list 不是空的
 if not img_list:
    raise ValueError("\u274c Error: No valid images were processed! Please check the image paths and formats.")
 # 執行 BIE 量化
 bie_model_path = km.analysis({input_name: img_list})
 # 確保 BIE 檔案儲存到指定目錄
 bie_save_path = os.path.join(onnx_dir, os.path.basename(bie_model_path))
 shutil.copy(bie_model_path, bie_save_path)  # 使用 shutil.move 來處理跨磁碟移動
 # 確認 BIE 模型是否生成
 if not os.path.exists(bie_save_path):
    raise RuntimeError(f"\u274c Error: BIE model was not generated! Please check your quantization process.")
 print("\n\u2705 Fixed-point analysis done! BIE model saved to:", bie_save_path)
 # 確保 `km` 已經初始化，並且 `.bie` 模型已生成
 nef_model_path = ktc.compile([km])
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 NEF 目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 .nef 檔案存在
 if not os.path.exists(nef_model_path):
    raise RuntimeError(f"❌ Error: NEF model was not generated at {nef_model_path}! Please check your compilation process.")
 # 確保 NEF 檔案儲存到指定目錄
 nef_save_path = os.path.join(onnx_dir, os.path.basename(nef_model_path))
 if os.path.exists(nef_model_path):
    shutil.copy(nef_model_path, nef_save_path)
 else:
    raise RuntimeError(f"❌ Error: NEF model was expected at {nef_model_path}, but it does not exist!")  # 同樣使用 shutil.move
 if not os.path.exists(nef_save_path):
    raise RuntimeError(f"\u274c Error: NEF model was not generated! Please check your compilation process.")
 print("\n\u2705 Compile done! NEF file saved to:", nef_save_path)
--- a/kneron/onnx2nef720.py
+++ b/kneron/onnx2nef720.py
@ -0,0 +1,110 @@
 import ktc
 import numpy as np
 import os
 import onnx
 import shutil
 from PIL import Image
 import torch
 from yolov5_preprocess import Yolov5_preprocess
 import kneron_preprocessing
 # 設定 ONNX 模型儲存路徑
 onnx_dir = 'runs/train/exp73/weights/'
 onnx_path = os.path.join(onnx_dir, 'best_simplified.onnx')
 # 確保目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 加載並優化 ONNX 模型
 m = onnx.load(onnx_path)
 m = ktc.onnx_optimizer.onnx2onnx_flow(m)
 opt_onnx_path = os.path.join(onnx_dir, 'latest.opt.onnx')
 onnx.save(m, opt_onnx_path)
 km = ktc.ModelConfig(20008, "0001", "720", onnx_model=m)
 eval_result = km.evaluate()
 print("\nNpu performance evaluation result:\n" + str(eval_result))
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 imgsz_h, imgsz_w = 640, 640
 data_path = "data50"
 files_found = [f for _, _, files in os.walk(data_path) for f in files if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp"))]
 if not files_found:
    raise FileNotFoundError(f"\u274c Error: No images found in {data_path}! Please check your dataset.")
 print(f"\u2705 Found {len(files_found)} images in {data_path}")
 # 獲取 ONNX 模型的輸入名稱
 input_name = m.graph.input[0].name  # 確保 key 與 ONNX input name 一致
 # 存儲預處理後的圖片數據
 img_list = []
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # 只處理圖片文件
        if not f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
            print(f"\u26a0\ufe0f Skipping non-image file: {fullpath}")
            continue
        # 嘗試處理圖片
        try:
            img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
            img_data = img_data.cpu().numpy()
            print(f"\u2705 Processed: {fullpath}")
            img_list.append(img_data)
        except Exception as e:
            print(f"\u274c Failed to process {fullpath}: {e}")
 # 確保 img_list 不是空的
 if not img_list:
    raise ValueError("\u274c Error: No valid images were processed! Please check the image paths and formats.")
 # 執行 BIE 量化
 bie_model_path = km.analysis({input_name: img_list})
 # 確保 BIE 檔案儲存到指定目錄
 bie_save_path = os.path.join(onnx_dir, os.path.basename(bie_model_path))
 shutil.copy(bie_model_path, bie_save_path)  # 使用 shutil.move 來處理跨磁碟移動
 # 確認 BIE 模型是否生成
 if not os.path.exists(bie_save_path):
    raise RuntimeError(f"\u274c Error: BIE model was not generated! Please check your quantization process.")
 print("\n\u2705 Fixed-point analysis done! BIE model saved to:", bie_save_path)
 # 確保 `km` 已經初始化，並且 `.bie` 模型已生成
 nef_model_path = ktc.compile([km])
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 NEF 目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 .nef 檔案存在
 if not os.path.exists(nef_model_path):
    raise RuntimeError(f"❌ Error: NEF model was not generated at {nef_model_path}! Please check your compilation process.")
 # 確保 NEF 檔案儲存到指定目錄
 nef_save_path = os.path.join(onnx_dir, os.path.basename(nef_model_path))
 if os.path.exists(nef_model_path):
    shutil.copy(nef_model_path, nef_save_path)
 else:
    raise RuntimeError(f"❌ Error: NEF model was expected at {nef_model_path}, but it does not exist!")  # 同樣使用 shutil.move
 if not os.path.exists(nef_save_path):
    raise RuntimeError(f"\u274c Error: NEF model was not generated! Please check your compilation process.")
 print("\n\u2705 Compile done! NEF file saved to:", nef_save_path)
--- a/kneron/onnx2nefSTDC630.py
+++ b/kneron/onnx2nefSTDC630.py
@ -0,0 +1,110 @@
 import ktc
 import numpy as np
 import os
 import onnx
 import shutil
 from PIL import Image
 import torch
 from yolov5_preprocess import Yolov5_preprocess
 import kneron_preprocessing
 # 設定 ONNX 模型儲存路徑
 onnx_dir = 'work_dirs/kn_stdc1_in1k-pre_512x1024_80k_cityscapes/'
 onnx_path = os.path.join(onnx_dir, 'latest.onnx')
 # 確保目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 加載並優化 ONNX 模型
 m = onnx.load(onnx_path)
 m = ktc.onnx_optimizer.onnx2onnx_flow(m)
 opt_onnx_path = os.path.join(onnx_dir, 'latest.opt.onnx')
 onnx.save(m, opt_onnx_path)
 km = ktc.ModelConfig(20008, "0001", "630", onnx_model=m)
 eval_result = km.evaluate()
 print("\nNpu performance evaluation result:\n" + str(eval_result))
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 imgsz_h, imgsz_w = 640, 640
 data_path = "data50"
 files_found = [f for _, _, files in os.walk(data_path) for f in files if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp"))]
 if not files_found:
    raise FileNotFoundError(f"\u274c Error: No images found in {data_path}! Please check your dataset.")
 print(f"\u2705 Found {len(files_found)} images in {data_path}")
 # 獲取 ONNX 模型的輸入名稱
 input_name = m.graph.input[0].name  # 確保 key 與 ONNX input name 一致
 # 存儲預處理後的圖片數據
 img_list = []
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # 只處理圖片文件
        if not f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
            print(f"\u26a0\ufe0f Skipping non-image file: {fullpath}")
            continue
        # 嘗試處理圖片
        try:
            img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
            img_data = img_data.cpu().numpy()
            print(f"\u2705 Processed: {fullpath}")
            img_list.append(img_data)
        except Exception as e:
            print(f"\u274c Failed to process {fullpath}: {e}")
 # 確保 img_list 不是空的
 if not img_list:
    raise ValueError("\u274c Error: No valid images were processed! Please check the image paths and formats.")
 # 執行 BIE 量化
 bie_model_path = km.analysis({input_name: img_list})
 # 確保 BIE 檔案儲存到指定目錄
 bie_save_path = os.path.join(onnx_dir, os.path.basename(bie_model_path))
 shutil.copy(bie_model_path, bie_save_path)  # 使用 shutil.move 來處理跨磁碟移動
 # 確認 BIE 模型是否生成
 if not os.path.exists(bie_save_path):
    raise RuntimeError(f"\u274c Error: BIE model was not generated! Please check your quantization process.")
 print("\n\u2705 Fixed-point analysis done! BIE model saved to:", bie_save_path)
 # 確保 `km` 已經初始化，並且 `.bie` 模型已生成
 nef_model_path = ktc.compile([km])
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 NEF 目標資料夾存在
 os.makedirs(onnx_dir, exist_ok=True)
 # 確保 nef_model_path 不是 None 或空值
 if not nef_model_path:
    raise RuntimeError("❌ Error: ktc.compile() did not return a valid .nef file path!")
 # 確保 .nef 檔案存在
 if not os.path.exists(nef_model_path):
    raise RuntimeError(f"❌ Error: NEF model was not generated at {nef_model_path}! Please check your compilation process.")
 # 確保 NEF 檔案儲存到指定目錄
 nef_save_path = os.path.join(onnx_dir, os.path.basename(nef_model_path))
 if os.path.exists(nef_model_path):
    shutil.copy(nef_model_path, nef_save_path)
 else:
    raise RuntimeError(f"❌ Error: NEF model was expected at {nef_model_path}, but it does not exist!")  # 同樣使用 shutil.move
 if not os.path.exists(nef_save_path):
    raise RuntimeError(f"\u274c Error: NEF model was not generated! Please check your compilation process.")
 print("\n\u2705 Compile done! NEF file saved to:", nef_save_path)
--- a/kneron/preprocessing/API.py
+++ b/kneron/preprocessing/API.py
@ -0,0 +1,684 @@
 # -*- coding: utf-8 -*-
 import numpy as np
 import os
 from .funcs.utils import str2int, str2bool
 from . import Flow
 flow = Flow()
 flow.set_numerical_type('floating')
 flow_520 = Flow()
 flow_520.set_numerical_type('520')
 flow_720 = Flow()
 flow_720.set_numerical_type('720')
 DEFAULT = None
 default = {
    'crop':{
        'align_w_to_4':False
        },
    'resize':{
        'type':'bilinear',
        'calculate_ratio_using_CSim':False
        }
 }
 def set_default_as_520():
    """
    Set some default parameter as 520 setting
    crop.align_w_to_4 = True
    crop.pad_square_to_4 = True
    resize.type = 'fixed_520'
    resize.calculate_ratio_using_CSim = True
    """
    global default
    default['crop']['align_w_to_4'] = True
    default['resize']['type'] = 'fixed_520'
    default['resize']['calculate_ratio_using_CSim'] = True
    return
 def set_default_as_floating():
    """
    Set some default parameter as floating setting
    crop.align_w_to_4 = False
    crop.pad_square_to_4 = False
    resize.type = 'bilinear'
    resize.calculate_ratio_using_CSim = False
    """
    global default
    default['crop']['align_w_to_4'] = False
    default['resize']['type'] = 'bilinear'
    default['resize']['calculate_ratio_using_CSim'] = False
    pass
 def print_info_on():
    """
    turn print infomation on.
    """
    flow.set_print_info(True)
    flow_520.set_print_info(True)
 def print_info_off():
    """
    turn print infomation off.
    """
    flow.set_print_info(False)
    flow_520.set_print_info(False)
 def load_image(image):
    """
    load_image function
    load load_image and output as rgb888 format np.array
    Args:
        image: [np.array/str], can be np.array or image file path
    Returns:
        out: [np.array], rgb888 format
    Examples:
    """
    image = flow.load_image(image, is_raw = False)
    return image
 def load_bin(image, fmt=None, size=None):
    """
    load_bin function
    load bin file and output as rgb888 format np.array
    Args:
        image: [str], bin file path
        fmt: [str], "rgb888" / "rgb565" / "nir"
        size: [tuble], (image_w, image_h)
    Returns:
        out: [np.array], rgb888 format
    Examples:
        >>> image_data = kneron_preprocessing.API.load_bin(image,'rgb565',(raw_w,raw_h))
    """    
    assert isinstance(size, tuple)
    assert isinstance(fmt, str)
    # assert (fmt.lower() in ['rgb888', "rgb565" , "nir",'RGB888', "RGB565" , "NIR", 'NIR888', 'nir888'])
    image = flow.load_image(image, is_raw = True, raw_img_type='bin', raw_img_fmt = fmt, img_in_width = size[0], img_in_height = size[1])
    flow.set_color_conversion(source_format=fmt, out_format = 'rgb888')
    image,_ = flow.funcs['color'](image)
    return image
 def load_hex(file, fmt=None, size=None):
    """
    load_hex function
    load hex file and output as rgb888 format np.array
    Args:
        image: [str], hex file path
        fmt: [str], "rgb888" / "yuv444" / "ycbcr444" / "yuv422" / "ycbcr422" / "rgb565"
        size: [tuble], (image_w, image_h)
    Returns:
        out: [np.array], rgb888 format
    Examples:
        >>> image_data = kneron_preprocessing.API.load_hex(image,'rgb565',(raw_w,raw_h))
    """  
    assert isinstance(size, tuple)
    assert isinstance(fmt, str)
    assert (fmt.lower() in ['rgb888',"yuv444" , "ycbcr444" , "yuv422" , "ycbcr422" , "rgb565"])
    image = flow.load_image(file, is_raw = True, raw_img_type='hex', raw_img_fmt = fmt, img_in_width = size[0], img_in_height = size[1])
    flow.set_color_conversion(source_format=fmt, out_format = 'rgb888')
    image,_ = flow.funcs['color'](image)
    return image
 def dump_image(image, output=None, file_fmt='txt',image_fmt='rgb888',order=0):
    """
    dump_image function
    dump txt, bin or hex, default is txt
    image format as following format: RGB888, RGBA8888, RGB565, NIR, YUV444, YCbCr444, YUV422, YCbCr422, default is RGB888
    Args:
        image: [np.array/str], can be np.array or image file path
        output: [str], dump file path
        file_fmt: [str], "bin" / "txt" / "hex", set dump file format, default is txt
        image_fmt: [str], RGB888 / RGBA8888 / RGB565 / NIR / YUV444 / YCbCr444 / YUV422 / YCbCr422, default is RGB888
    Examples:
        >>> kneron_preprocessing.API.dump_image(image_data,out_path,fmt='bin')
    """
    if isinstance(image, str):
        image = load_image(image)
    assert isinstance(image, np.ndarray)
    if output is None:
        return
    flow.set_output_setting(is_dump=False, dump_format=file_fmt, image_format=image_fmt ,output_file=output)
    flow.dump_image(image)
    return
 def convert(image, out_fmt = 'RGB888', source_fmt = 'RGB888'):
    """
    color convert
    Args:
        image: [np.array], input
        out_fmt: [str], "rgb888" / "rgba8888" / "rgb565" / "yuv" / "ycbcr" / "yuv422" / "ycbcr422"
        source_fmt: [str], "rgb888" / "rgba8888" / "rgb565" / "yuv" / "ycbcr" / "yuv422" / "ycbcr422"
    Returns:
        out: [np.array]
    Examples:
    """  
    flow.set_color_conversion(source_format = source_fmt, out_format=out_fmt, simulation=False)
    image,_ = flow.funcs['color'](image)
    return image
 def get_crop_range(box,align_w_to_4=DEFAULT, pad_square_to_4=False,rounding_type=0):
    """
    get exact crop box according different setting
    Args:
        box: [tuble], (x1, y1, x2, y2)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
    Returns:
        out: [tuble,4], (crop_x1, crop_y1, crop_x2, crop_y2) 
    Examples:
        >>> image_data = kneron_preprocessing.API.get_crop_range((272,145,461,341), align_w_to_4=True, pad_square_to_4=True)
        (272, 145, 460, 341)
    """  
    if box is None:
        return (0,0,0,0)
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='specific', start_x=box[0],start_y=box[1],end_x=box[2],end_y=box[3], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4,rounding_type=rounding_type)
    image = np.zeros((1,1,3)).astype('uint8')
    _,info = flow.funcs['crop'](image)
    return info['box']
 def crop(image, box=None, align_w_to_4=DEFAULT, pad_square_to_4=False,rounding_type=0 ,info_out = {}):
    """
    crop function
    specific crop range by box
    Args:
        image: [np.array], input
        box: [tuble], (x1, y1, x2, y2)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
        info_out: [dic], save the final crop box into info_out['box']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop(image_data,(272,145,461,341), align_w_to_4=True, info_out=info)
        >>> info['box']
        (272, 145, 460, 341)
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop(image_data,(272,145,461,341), pad_square_to_4=True, info_out=info)
        >>> info['box']
        (268, 145, 464, 341)
    """  
    assert isinstance(image, np.ndarray)
    if box is None:
        return image
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='specific', start_x=box[0],start_y=box[1],end_x=box[2],end_y=box[3], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4,rounding_type=rounding_type)
    image,info = flow.funcs['crop'](image)
    info_out['box'] = info['box']
    return image
 def crop_center(image, range=None, align_w_to_4=DEFAULT, pad_square_to_4=False,rounding_type=0 ,info_out = {}):
    """
    crop function
    center crop by range
    Args:
        image: [np.array], input
        range: [tuble], (crop_w, crop_h)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
        info_out: [dic], save the final crop box into info_out['box']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_center(image_data,(102,40), align_w_to_4=True,info_out=info)
        >>> info['box']
        (268, 220, 372, 260)
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_center(image_data,(102,40), pad_square_to_4=True, info_out=info)
        >>> info['box']
        (269, 192, 371, 294)
    """   
    assert isinstance(image, np.ndarray)
    if range is None:
        return image
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='center', crop_w=range[0],crop_h=range[1], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4,rounding_type=rounding_type)
    image,info = flow.funcs['crop'](image)
    info_out['box'] = info['box']
    return image
 def crop_corner(image, range=None, align_w_to_4=DEFAULT,pad_square_to_4=False,rounding_type=0 ,info_out = {}):
    """
    crop function
    corner crop by range
    Args:
        image: [np.array], input
        range: [tuble], (crop_w, crop_h)
        align_w_to_4: [bool], crop length in w direction align to 4 or not, default False
        pad_square_to_4: [bool], pad to square(align 4) or not, default False
        rounding_type: [int], 0-> x1,y1 take floor, x2,y2 take ceil; 1->all take rounding 
        info_out: [dic], save the final crop box into info_out['box']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_corner(image_data,(102,40), align_w_to_4=True,info_out=info)
        >>> info['box']
        (0, 0, 104, 40)
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.crop_corner(image_data,(102,40), pad_square_to_4=True,info_out=info)
        >>> info['box']
        (0, -28, 102, 74)
    """
    assert isinstance(image, np.ndarray)
    if range is None:
        return image
    if align_w_to_4 is None:
        align_w_to_4 = default['crop']['align_w_to_4']
    flow.set_crop(type='corner', crop_w=range[0],crop_h=range[1], align_w_to_4=align_w_to_4, pad_square_to_4=pad_square_to_4)
    image, info = flow.funcs['crop'](image)
    info_out['box'] = info['box']
    return image
 def resize(image, size=None, keep_ratio = True, zoom = True, type=DEFAULT, calculate_ratio_using_CSim = DEFAULT, info_out = {}):
    """
    resize function
    resize type can be bilinear or bilicubic as floating type, fixed or fixed_520/fixed_720 as fixed type.
    fixed_520/fixed_720 type has add some function to simulate 520/720 bug.
    Args:
        image: [np.array], input
        size: [tuble], (input_w, input_h)
        keep_ratio: [bool], keep_ratio or not, default True
        zoom: [bool], enable resize can zoom image or not, default True
        type: [str], "bilinear" / "bilicubic" / "cv2" / "fixed" / "fixed_520" / "fixed_720"
        calculate_ratio_using_CSim: [bool], calculate the ratio and scale using Csim function and C float, default False
        info_out: [dic], save the final scale size(w,h) into info_out['size']
    Returns:
        out: [np.array] 
    Examples:
        >>> info = {}
        >>> image_data = kneron_preprocessing.API.resize(image_data,size=(56,56),type='fixed',info_out=info)
        >>> info_out['size']
        (54,56)
    """
    assert isinstance(image, np.ndarray)
    if size is None:
        return image
    if type is None:
        type = default['resize']['type']
    if calculate_ratio_using_CSim is None:
        calculate_ratio_using_CSim = default['resize']['calculate_ratio_using_CSim']
    flow.set_resize(resize_w = size[0], resize_h = size[1], type=type, keep_ratio=keep_ratio,zoom=zoom, calculate_ratio_using_CSim=calculate_ratio_using_CSim)
    image, info = flow.funcs['resize'](image)
    info_out['size'] = info['size']
    return image
 def pad(image, pad_l=0, pad_r=0, pad_t=0, pad_b=0, pad_val=0):
    """
    pad function
    specific left, right, top and bottom pad size.
    Args:
        image[np.array]: input
        pad_l: [int], pad size from left, default 0
        pad_r: [int], pad size form right, default 0
        pad_t: [int], pad size from top, default 0
        pad_b: [int], pad size form bottom, default 0
        pad_val: [float], the value of pad, , default 0 
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.pad(image_data,20,40,20,40,-0.5)
    """
    assert isinstance(image, np.ndarray)
    flow.set_padding(type='specific',pad_l=pad_l,pad_r=pad_r,pad_t=pad_t,pad_b=pad_b,pad_val=pad_val)
    image, _ = flow.funcs['padding'](image)
    return image
 def pad_center(image,size=None, pad_val=0):
    """
    pad function
    center pad with pad size.
    Args:
        image[np.array]: input
        size: [tuble], (padded_size_w, padded_size_h)
        pad_val: [float], the value of pad, , default 0 
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.pad_center(image_data,size=(56,56),pad_val=-0.5)
    """
    assert isinstance(image, np.ndarray)
    if size is None:
        return image
    assert ( (image.shape[0] <= size[1]) & (image.shape[1] <= size[0]) )
    flow.set_padding(type='center',padded_w=size[0],padded_h=size[1],pad_val=pad_val)
    image, _ = flow.funcs['padding'](image)
    return image
 def pad_corner(image,size=None, pad_val=0):
    """
    pad function
    corner pad with pad size.
    Args:
        image[np.array]: input
        size: [tuble], (padded_size_w, padded_size_h)
        pad_val: [float], the value of pad, , default 0 
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.pad_corner(image_data,size=(56,56),pad_val=-0.5)
    """   
    assert isinstance(image, np.ndarray)
    if size is None:
        return image
    assert ( (image.shape[0] <= size[1]) & (image.shape[1] <= size[0]) )
    flow.set_padding(type='corner',padded_w=size[0],padded_h=size[1],pad_val=pad_val)
    image, _ = flow.funcs['padding'](image)
    return image
 def norm(image,scale=256.,bias=-0.5, mean=None, std=None):
    """
    norm function
    x = (x/scale - bias)
    x[0,1,2] = x - mean[0,1,2]
    x[0,1,2] = x / std[0,1,2]
    Args:
        image: [np.array], input
        scale: [float], default = 256
        bias: [float], default = -0.5
        mean: [tuble,3], default = None
        std: [tuble,3], default = None
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.norm(image_data)
        >>> image_data = kneron_preprocessing.API.norm(image_data,mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    """  
    assert isinstance(image, np.ndarray)
    flow.set_normalize(type='specific',scale=scale,  bias=bias, mean=mean, std =std)
    image, _ = flow.funcs['normalize'](image)
    return image
 def inproc_520(image,raw_fmt='rgb565',raw_size=None,npu_size=None, crop_box=None, pad_mode=0, norm='kneron', gray=False, rotate=0, radix=8, bit_width=8, round_w_to_16=True, NUM_BANK_LINE=32,BANK_ENTRY_CNT=512,MAX_IMG_PREPROC_ROW_NUM=511,MAX_IMG_PREPROC_COL_NUM=256):
    """
    inproc_520
    Args:
        image: [np.array], input
        crop_box: [tuble], (x1, y1, x2, y2), if None will skip crop
        pad_mode: [int], 0: pad 2 sides, 1: pad 1 side, 2: no pad. default = 0
        norm: [str], default = 'kneron'
        rotate: [int], 0 / 1 / 2 ,default = 0
        radix: [int], default = 8
        bit_width: [int], default = 8
        round_w_to_16: [bool], default = True
        gray: [bool], default = False
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.inproc_520(image_data,npu_size=(56,56),crop_box=(272,145,460,341),pad_mode=1)
    """  
    # assert isinstance(image, np.ndarray)
    if (not isinstance(image, np.ndarray)):
        flow_520.set_raw_img(is_raw_img='yes',raw_img_type = 'bin',raw_img_fmt=raw_fmt, img_in_width=raw_size[0], img_in_height=raw_size[1])
    else:
        flow_520.set_raw_img(is_raw_img='no')
        flow_520.set_color_conversion(source_format='rgb888')
    if npu_size is None:
        return image
    flow_520.set_model_size(w=npu_size[0],h=npu_size[1])
    ## Crop
    if crop_box != None:
        flow_520.set_crop(start_x=crop_box[0],start_y=crop_box[1],end_x=crop_box[2],end_y=crop_box[3])
        crop_fisrt = True
    else:
        crop_fisrt = False
    ## Color
    if gray:
        flow_520.set_color_conversion(out_format='l',simulation='no')
    else:
        flow_520.set_color_conversion(out_format='rgb888',simulation='no')
    ## Resize & Pad
    pad_mode = str2int(pad_mode)
    if (pad_mode == 0):
        pad_type = 'center'
        resize_keep_ratio = 'yes'
    elif (pad_mode == 1):
        pad_type = 'corner'
        resize_keep_ratio = 'yes'
    else:
        pad_type = 'center'
        resize_keep_ratio = 'no'
    flow_520.set_resize(keep_ratio=resize_keep_ratio)
    flow_520.set_padding(type=pad_type)
    ## Norm
    flow_520.set_normalize(type=norm)
    ## 520 inproc
    flow_520.set_520_setting(radix=radix,bit_width=bit_width,rotate=rotate,crop_fisrt=crop_fisrt,round_w_to_16=round_w_to_16,NUM_BANK_LINE=NUM_BANK_LINE,BANK_ENTRY_CNT=BANK_ENTRY_CNT,MAX_IMG_PREPROC_ROW_NUM=MAX_IMG_PREPROC_ROW_NUM,MAX_IMG_PREPROC_COL_NUM=MAX_IMG_PREPROC_COL_NUM)
    image_data, _ = flow_520.run_whole_process(image)
    return image_data
 def inproc_720(image,raw_fmt='rgb565',raw_size=None,npu_size=None, crop_box=None, pad_mode=0, norm='kneron', gray=False):
    """
    inproc_720
    Args:
        image: [np.array], input
        crop_box: [tuble], (x1, y1, x2, y2), if None will skip crop
        pad_mode: [int], 0: pad 2 sides, 1: pad 1 side, 2: no pad. default = 0
        norm: [str], default = 'kneron'
        rotate: [int], 0 / 1 / 2 ,default = 0
        radix: [int], default = 8
        bit_width: [int], default = 8
        round_w_to_16: [bool], default = True
        gray: [bool], default = False
    Returns:
        out: [np.array] 
    Examples:
        >>> image_data = kneron_preprocessing.API.inproc_520(image_data,npu_size=(56,56),crop_box=(272,145,460,341),pad_mode=1)
    """  
    # assert isinstance(image, np.ndarray)
    if (not isinstance(image, np.ndarray)):
        flow_720.set_raw_img(is_raw_img='yes',raw_img_type = 'bin',raw_img_fmt=raw_fmt, img_in_width=raw_size[0], img_in_height=raw_size[1])
    else:
        flow_720.set_raw_img(is_raw_img='no')
        flow_720.set_color_conversion(source_format='rgb888')
    if npu_size is None:
        return image
    flow_720.set_model_size(w=npu_size[0],h=npu_size[1])
    ## Crop
    if crop_box != None:
        flow_720.set_crop(start_x=crop_box[0],start_y=crop_box[1],end_x=crop_box[2],end_y=crop_box[3])
        crop_fisrt = True
    else:
        crop_fisrt = False
    ## Color
    if gray:
        flow_720.set_color_conversion(out_format='l',simulation='no')
    else:
        flow_720.set_color_conversion(out_format='rgb888',simulation='no')
    ## Resize & Pad
    pad_mode = str2int(pad_mode)
    if (pad_mode == 0):
        pad_type = 'center'
        resize_keep_ratio = 'yes'
    elif (pad_mode == 1):
        pad_type = 'corner'
        resize_keep_ratio = 'yes'
    else:
        pad_type = 'center'
        resize_keep_ratio = 'no'
    flow_720.set_resize(keep_ratio=resize_keep_ratio)
    flow_720.set_padding(type=pad_type)
    ## 720 inproc
    # flow_720.set_720_setting(radix=radix,bit_width=bit_width,rotate=rotate,crop_fisrt=crop_fisrt,round_w_to_16=round_w_to_16,NUM_BANK_LINE=NUM_BANK_LINE,BANK_ENTRY_CNT=BANK_ENTRY_CNT,MAX_IMG_PREPROC_ROW_NUM=MAX_IMG_PREPROC_ROW_NUM,MAX_IMG_PREPROC_COL_NUM=MAX_IMG_PREPROC_COL_NUM)
    image_data, _ = flow_720.run_whole_process(image)
    return image_data
 def bit_match(data1, data2):
    """
    bit_match function
    check data1 is equal to data2 or not.
    Args:
        data1: [np.array / str], can be array or txt/bin file
        data2: [np.array / str], can be array or txt/bin file
    Returns:
        out1: [bool], is match or not
        out2: [np.array], if not match, save the position for mismatched data
    Examples:
        >>> result, mismatched = kneron_preprocessing.API.bit_match(data1,data2)
    """
    if isinstance(data1, str):
        if os.path.splitext(data1)[1] == '.bin':
            data1 = np.fromfile(data1, dtype='uint8')
        elif os.path.splitext(data1)[1] == '.txt':
            data1 = np.loadtxt(data1)
    assert isinstance(data1, np.ndarray)
    if isinstance(data2, str):
        if os.path.splitext(data2)[1] == '.bin':
            data2 = np.fromfile(data2, dtype='uint8')
        elif os.path.splitext(data2)[1] == '.txt':
            data2 = np.loadtxt(data2)
    assert isinstance(data2, np.ndarray)
    data1 = data1.reshape((-1,1))
    data2 = data2.reshape((-1,1))
    if not(len(data1) == len(data2)):
        print('error len')
        return False, np.zeros((1))
    else: 
        ans = data2 - data1    
        if len(np.where(ans>0)[0]) > 0:
            print('error',np.where(ans>0)[0])
            return False, np.where(ans>0)[0]
        else:
            print('pass')
            return True, np.zeros((1))
 def cpr_to_crp(x_start, x_end, y_start, y_end, pad_l, pad_r, pad_t, pad_b, rx_start, rx_end, ry_start, ry_end):
    """
    calculate the parameters of crop->pad->resize flow  to HW crop->resize->padding flow
    Args:
    Returns:
    Examples:
    """
    pad_l = round(pad_l * (rx_end-rx_start) / (x_end - x_start + pad_l + pad_r))
    pad_r = round(pad_r * (rx_end-rx_start) / (x_end - x_start + pad_l + pad_r)) 
    pad_t = round(pad_t * (ry_end-ry_start) / (y_end - y_start + pad_t + pad_b))
    pad_b = round(pad_b * (ry_end-ry_start) / (y_end - y_start + pad_t + pad_b))
    rx_start +=pad_l
    rx_end -=pad_r
    ry_start +=pad_t
    ry_end -=pad_b
    return x_start, x_end, y_start, y_end, pad_l, pad_r, pad_t, pad_b, rx_start, rx_end, ry_start, ry_end
--- a/kneron/preprocessing/Cflow.py
+++ b/kneron/preprocessing/Cflow.py
@ -0,0 +1,172 @@
 import numpy as np
 import argparse
 import kneron_preprocessing
 def main_(args):
    image = args.input_file
    filefmt = args.file_fmt
    if filefmt == 'bin':
        raw_format = args.raw_format
        raw_w = args.input_width
        raw_h = args.input_height
        image_data = kneron_preprocessing.API.load_bin(image,raw_format,(raw_w,raw_h))
    else:
        image_data = kneron_preprocessing.API.load_image(image)
    npu_w = args.width
    npu_h = args.height
    crop_first = True if args.crop_first == "True" else False
    if crop_first:
        x1 = args.x_pos
        y1 = args.y_pos
        x2 = args.crop_w + x1
        y2 = args.crop_h + y1
        crop_box = [x1,y1,x2,y2]
    else:
        crop_box = None
    pad_mode = args.pad_mode
    norm_mode = args.norm_mode
    bitwidth = args.bitwidth
    radix = args.radix
    rotate = args.rotate_mode
    ##
    image_data = kneron_preprocessing.API.inproc_520(image_data,npu_size=(npu_w,npu_h),crop_box=crop_box,pad_mode=pad_mode,norm=norm_mode,rotate=rotate,radix=radix,bit_width=bitwidth)
    output_file = args.output_file
    kneron_preprocessing.API.dump_image(image_data,output_file,'bin','rgba')
    return
 if __name__ == "__main__":
    argparser = argparse.ArgumentParser(
        description="preprocessing"
        )
    argparser.add_argument(
        '-i',
        '--input_file',
        help="input file name"
        )
    argparser.add_argument(
        '-ff',
        '--file_fmt',
        help="input file format, jpg or bin"
        )
    argparser.add_argument(
        '-rf',
        '--raw_format',
        help="input file image format, rgb or rgb565 or nir"
        )
    argparser.add_argument(
        '-i_w',
        '--input_width',
        type=int,
        help="input image width"
        )
    argparser.add_argument(
        '-i_h',
        '--input_height',
        type=int,
        help="input image height"
        )
    argparser.add_argument(
        '-o',
        '--output_file',
        help="output file name"
        )
    argparser.add_argument(
        '-s_w',
        '--width',
        type=int,
        help="output width for npu input",
        )
    argparser.add_argument(
        '-s_h',
        '--height',
        type=int,
        help="output height for npu input",
        )
    argparser.add_argument(
        '-c_f',
        '--crop_first',
        help="crop first True or False",
        )
    argparser.add_argument(
        '-x',
        '--x_pos',
        type=int,
        help="left up coordinate x",
        )
    argparser.add_argument(
        '-y',
        '--y_pos',
        type=int,
        help="left up coordinate y",
        )
    argparser.add_argument(
        '-c_w',
        '--crop_w',
        type=int,
        help="crop width",
        )
    argparser.add_argument(
        '-c_h',
        '--crop_h',
        type=int,
        help="crop height",
        )
    argparser.add_argument(
        '-p_m',
        '--pad_mode',
        type=int,
        help=" 0: pad 2 sides, 1: pad 1 side, 2: no pad.",
        )
    argparser.add_argument(
        '-n_m',
        '--norm_mode',
        help="normalizaton mode: yolo, kneron, tf."
        )
    argparser.add_argument(
        '-r_m',
        '--rotate_mode',
        type=int,
        help="rotate mode:0,1,2"
        )
    argparser.add_argument(
        '-bw',
        '--bitwidth',
        type=int,
        help="Int for bitwidth"
        )
    argparser.add_argument(
        '-r',
        '--radix',
        type=int,
        help="Int for radix"
        )
    args = argparser.parse_args()
    main_(args)
--- a/kneron/preprocessing/Flow.py
+++ b/kneron/preprocessing/Flow.py
--- a/kneron/preprocessing/init.py
+++ b/kneron/preprocessing/init.py
@ -0,0 +1,2 @@
 from .Flow import *
 from .API import *
--- a/kneron/preprocessing/funcs/ColorConversion.py
+++ b/kneron/preprocessing/funcs/ColorConversion.py
@ -0,0 +1,285 @@
 import numpy as np
 from PIL import Image
 from .utils import signed_rounding, clip, str2bool
 format_bit = 10
 c00_yuv = 1
 c02_yuv = 1436
 c10_yuv = 1
 c11_yuv = -354
 c12_yuv = -732
 c20_yuv = 1
 c21_yuv = 1814
 c00_ycbcr = 1192
 c02_ycbcr = 1634
 c10_ycbcr = 1192
 c11_ycbcr = -401
 c12_ycbcr = -833
 c20_ycbcr = 1192
 c21_ycbcr = 2065
 Matrix_ycbcr_to_rgb888 = np.array(
    [[1.16438356e+00,  1.16438356e+00,  1.16438356e+00],
     [2.99747219e-07, - 3.91762529e-01,  2.01723263e+00],
     [1.59602686e+00, - 8.12968294e-01,  3.04059479e-06]])
 Matrix_rgb888_to_ycbcr = np.array(
    [[0.25678824, - 0.14822353,  0.43921569],
     [0.50412941, - 0.29099216, - 0.36778824],
     [0.09790588,  0.43921569, - 0.07142745]])
 Matrix_rgb888_to_yuv = np.array(
    [[ 0.29899106, -0.16877996,  0.49988381],
    [ 0.5865453,  -0.33110385, -0.41826072],
    [ 0.11446364,  0.49988381, -0.08162309]])
 # Matrix_rgb888_to_yuv = np.array(
 #     [[0.299, - 0.147,   0.615],
 #      [0.587, - 0.289, - 0.515],
 #      [0.114,   0.436, - 0.100]])
 # Matrix_yuv_to_rgb888 = np.array(
 #     [[1.000,   1.000,  1.000],
 #      [0.000, - 0.394,  2.032],
 #      [1.140, - 0.581,  0.000]])
 class runner(object):
    def __init__(self):
        self.set = {
            'print_info':'no',
            'model_size':[0,0],
            'numerical_type':'floating',
            "source_format": "rgb888",
            "out_format": "rgb888",
            "options": {
                "simulation": "no",
                "simulation_format": "rgb888"
            }
        }
    def update(self, **kwargs):
        #
        self.set.update(kwargs)
        ## simulation
        self.funs = []
        if str2bool(self.set['options']['simulation']) and self.set['source_format'].lower() in ['RGB888', 'rgb888', 'RGB', 'rgb']:
            if self.set['options']['simulation_format'].lower() in ['YUV422', 'yuv422', 'YUV', 'yuv']:
                self.funs.append(self._ColorConversion_RGB888_to_YUV422)
                self.set['source_format'] = 'YUV422'
            elif self.set['options']['simulation_format'].lower() in ['YCBCR422', 'YCbCr422', 'ycbcr422', 'YCBCR', 'YCbCr', 'ycbcr']:
                self.funs.append(self._ColorConversion_RGB888_to_YCbCr422)
                self.set['source_format'] = 'YCbCr422'
            elif self.set['options']['simulation_format'].lower() in['RGB565', 'rgb565']:
                self.funs.append(self._ColorConversion_RGB888_to_RGB565)
                self.set['source_format'] = 'RGB565'
        ## to rgb888
        if self.set['source_format'].lower() in ['YUV444', 'yuv444','YUV422', 'yuv422', 'YUV', 'yuv']:
            self.funs.append(self._ColorConversion_YUV_to_RGB888)
        elif self.set['source_format'].lower() in ['YCBCR444', 'YCbCr444', 'ycbcr444','YCBCR422', 'YCbCr422', 'ycbcr422', 'YCBCR', 'YCbCr', 'ycbcr']:
            self.funs.append(self._ColorConversion_YCbCr_to_RGB888)
        elif self.set['source_format'].lower() in ['RGB565', 'rgb565']:
            self.funs.append(self._ColorConversion_RGB565_to_RGB888)
        elif self.set['source_format'].lower() in ['l', 'L' , 'nir', 'NIR']:
            self.funs.append(self._ColorConversion_L_to_RGB888)
        elif self.set['source_format'].lower() in ['RGBA8888', 'rgba8888' , 'RGBA', 'rgba']:
            self.funs.append(self._ColorConversion_RGBA8888_to_RGB888)
        ## output format
        if self.set['out_format'].lower() in ['L', 'l']:
            self.funs.append(self._ColorConversion_RGB888_to_L)
        elif self.set['out_format'].lower() in['RGB565', 'rgb565']:
            self.funs.append(self._ColorConversion_RGB888_to_RGB565)
        elif self.set['out_format'].lower() in['RGBA', 'RGBA8888','rgba','rgba8888']:
            self.funs.append(self._ColorConversion_RGB888_to_RGBA8888)
        elif self.set['out_format'].lower() in['YUV', 'YUV444','yuv','yuv444']:
            self.funs.append(self._ColorConversion_RGB888_to_YUV444)
        elif self.set['out_format'].lower() in['YUV422','yuv422']:
            self.funs.append(self._ColorConversion_RGB888_to_YUV422)
        elif self.set['out_format'].lower() in['YCBCR', 'YCBCR444','YCbCr','YCbCr444','ycbcr','ycbcr444']:
            self.funs.append(self._ColorConversion_RGB888_to_YCbCr444)
        elif self.set['out_format'].lower() in['YCBCR422','YCbCr422','ycbcr422']:
            self.funs.append(self._ColorConversion_RGB888_to_YCbCr422)
    def print_info(self):
        print("<colorConversion>",
              "source_format:", self.set['source_format'],
              ', out_format:', self.set['out_format'],
              ', simulation:', self.set['options']['simulation'],
              ', simulation_format:', self.set['options']['simulation_format'])
    def run(self, image_data):
        assert isinstance(image_data, np.ndarray)
        # print info
        if str2bool(self.set['print_info']):
            self.print_info()
        # color
        for _, f in enumerate(self.funs):
            image_data = f(image_data)
        # output
        info = {}
        return image_data, info
    def _ColorConversion_RGB888_to_YUV444(self, image):
        ## floating
        image = image.astype('float')
        image = (image @ Matrix_rgb888_to_yuv + 0.5).astype('uint8')
        return image
    def _ColorConversion_RGB888_to_YUV422(self, image):
        # rgb888 to yuv444
        image = self._ColorConversion_RGB888_to_YUV444(image)
        # yuv444 to yuv422
        u2 = image[:, 0::2, 1]
        u4 = np.repeat(u2, 2, axis=1)
        v2 = image[:, 1::2, 2]
        v4 = np.repeat(v2, 2, axis=1)
        image[..., 1] = u4
        image[..., 2] = v4
        return image
    def _ColorConversion_YUV_to_RGB888(self, image):
        ## fixed
        h, w, c = image.shape
        image_f = image.reshape((h * w, c))
        image_rgb_f = np.zeros(image_f.shape, dtype=np.uint8)
        for i in range(h * w):
            image_y = image_f[i, 0] *1024
            if image_f[i, 1] > 127:
                image_u = -((~(image_f[i, 1] - 1)) & 0xFF)
            else:
                image_u = image_f[i, 1]
            if image_f[i, 2] > 127:
                image_v = -((~(image_f[i, 2] - 1)) & 0xFF)
            else:
                image_v = image_f[i, 2]
            image_r = c00_yuv * image_y + c02_yuv * image_v
            image_g = c10_yuv * image_y + c11_yuv * image_u + c12_yuv * image_v
            image_b = c20_yuv * image_y + c21_yuv * image_u
            image_r = signed_rounding(image_r, format_bit)
            image_g = signed_rounding(image_g, format_bit)
            image_b = signed_rounding(image_b, format_bit)
            image_r = image_r >> format_bit
            image_g = image_g >> format_bit
            image_b = image_b >> format_bit
            image_rgb_f[i, 0] = clip(image_r, 0, 255)
            image_rgb_f[i, 1] = clip(image_g, 0, 255)
            image_rgb_f[i, 2] = clip(image_b, 0, 255)
        image_rgb = image_rgb_f.reshape((h, w, c))
        return image_rgb
    def _ColorConversion_RGB888_to_YCbCr444(self, image):
        ## floating
        image = image.astype('float')
        image = (image @ Matrix_rgb888_to_ycbcr + 0.5).astype('uint8')
        image[:, :, 0] += 16
        image[:, :, 1] += 128
        image[:, :, 2] += 128
        return image
    def _ColorConversion_RGB888_to_YCbCr422(self, image):
        # rgb888 to ycbcr444
        image = self._ColorConversion_RGB888_to_YCbCr444(image)
        # ycbcr444 to ycbcr422
        cb2 = image[:, 0::2, 1]
        cb4 = np.repeat(cb2, 2, axis=1)
        cr2 = image[:, 1::2, 2]
        cr4 = np.repeat(cr2, 2, axis=1)
        image[..., 1] = cb4
        image[..., 2] = cr4
        return image
    def _ColorConversion_YCbCr_to_RGB888(self, image):
        ## floating
        if (self.set['numerical_type'] == 'floating'):
            image = image.astype('float')
            image[:, :, 0] -= 16
            image[:, :, 1] -= 128
            image[:, :, 2] -= 128
            image = ((image @ Matrix_ycbcr_to_rgb888) + 0.5).astype('uint8')
            return image
        ## fixed
        h, w, c = image.shape
        image_f = image.reshape((h * w, c))
        image_rgb_f = np.zeros(image_f.shape, dtype=np.uint8)
        for i in range(h * w):
            image_y = (image_f[i, 0] - 16) * c00_ycbcr
            image_cb = image_f[i, 1] - 128
            image_cr = image_f[i, 2] - 128
            image_r = image_y + c02_ycbcr * image_cr
            image_g = image_y + c11_ycbcr * image_cb + c12_ycbcr * image_cr
            image_b = image_y + c21_ycbcr * image_cb
            image_r = signed_rounding(image_r, format_bit)
            image_g = signed_rounding(image_g, format_bit)
            image_b = signed_rounding(image_b, format_bit)
            image_r = image_r >> format_bit
            image_g = image_g >> format_bit
            image_b = image_b >> format_bit
            image_rgb_f[i, 0] = clip(image_r, 0, 255)
            image_rgb_f[i, 1] = clip(image_g, 0, 255)
            image_rgb_f[i, 2] = clip(image_b, 0, 255)
        image_rgb = image_rgb_f.reshape((h, w, c))
        return image_rgb
    def _ColorConversion_RGB888_to_RGB565(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]>=3)
        image_rgb565 = np.zeros(image.shape, dtype=np.uint8)
        image_rgb = image.astype('uint8')
        image_rgb565[:, :, 0] = image_rgb[:, :, 0] >> 3
        image_rgb565[:, :, 1] = image_rgb[:, :, 1] >> 2
        image_rgb565[:, :, 2] = image_rgb[:, :, 2] >> 3
        return image_rgb565
    def _ColorConversion_RGB565_to_RGB888(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]==3)
        image_rgb = np.zeros(image.shape, dtype=np.uint8)
        image_rgb[:, :, 0] = image[:, :, 0] << 3
        image_rgb[:, :, 1] = image[:, :, 1] << 2
        image_rgb[:, :, 2] = image[:, :, 2] << 3
        return image_rgb
    def _ColorConversion_L_to_RGB888(self, image):
        image_L = image.astype('uint8')
        img = Image.fromarray(image_L).convert('RGB')
        image_data = np.array(img).astype('uint8')
        return image_data
    def _ColorConversion_RGB888_to_L(self, image):
        image_rgb = image.astype('uint8')
        img = Image.fromarray(image_rgb).convert('L')
        image_data = np.array(img).astype('uint8')
        return image_data
    def _ColorConversion_RGBA8888_to_RGB888(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]==4)
        return image[:,:,:3]
    def _ColorConversion_RGB888_to_RGBA8888(self, image):
        assert (len(image.shape)==3)
        assert (image.shape[2]==3)
        imageA = np.concatenate((image, np.zeros((image.shape[0], image.shape[1], 1), dtype=np.uint8) ), axis=2)
        return imageA
--- a/kneron/preprocessing/funcs/Crop.py
+++ b/kneron/preprocessing/funcs/Crop.py
@ -0,0 +1,145 @@
 import numpy as np
 from PIL import Image
 from .utils import str2int, str2float, str2bool, pad_square_to_4
 from .utils_520 import round_up_n
 from .Runner_base import Runner_base, Param_base
 class General(Param_base):
    type = 'center'
    align_w_to_4 = False
    pad_square_to_4 = False
    rounding_type = 0
    crop_w = 0
    crop_h = 0
    start_x = 0.
    start_y = 0.
    end_x = 0.
    end_y = 0.
    def update(self, **dic):
        self.type = dic['type']
        self.align_w_to_4 = str2bool(dic['align_w_to_4'])
        self.rounding_type = str2int(dic['rounding_type'])
        self.crop_w = str2int(dic['crop_w'])
        self.crop_h = str2int(dic['crop_h'])
        self.start_x = str2float(dic['start_x'])
        self.start_y = str2float(dic['start_y'])
        self.end_x = str2float(dic['end_x'])
        self.end_y = str2float(dic['end_y'])
    def __str__(self):
        str_out = [
            ', type:',str(self.type),
            ', align_w_to_4:',str(self.align_w_to_4),
            ', pad_square_to_4:',str(self.pad_square_to_4),
            ', crop_w:',str(self.crop_w),
            ', crop_h:',str(self.crop_h),
            ', start_x:',str(self.start_x),
            ', start_y:',str(self.start_y),
            ', end_x:',str(self.end_x),
            ', end_y:',str(self.end_y)]
        return(' '.join(str_out))
 class runner(Runner_base):
    ## overwrite the class in Runner_base
    general = General()
    def __str__(self):
        return('<Crop>')
    def update(self, **kwargs):
        ##
        super().update(**kwargs)
        ##
        if (self.general.start_x != self.general.end_x) and (self.general.start_y != self.general.end_y):
            self.general.type = 'specific'
        elif(self.general.type != 'specific'):
            if self.general.crop_w == 0 or self.general.crop_h == 0:
                self.general.crop_w = self.common.model_size[0]
                self.general.crop_h = self.common.model_size[1]
            assert(self.general.crop_w > 0)
            assert(self.general.crop_h > 0)
            assert(self.general.type.lower() in ['CENTER', 'Center', 'center', 'CORNER', 'Corner', 'corner'])
        else:
            assert(self.general.type == 'specific')
    def run(self, image_data):
        ## init
        img = Image.fromarray(image_data)
        w, h = img.size
        ## get range
        if self.general.type.lower() in ['CENTER', 'Center', 'center']:
            x1, y1, x2, y2 = self._calcuate_xy_center(w, h)
        elif self.general.type.lower() in ['CORNER', 'Corner', 'corner']:
            x1, y1, x2, y2 = self._calcuate_xy_corner(w, h)
        else:
            x1 = self.general.start_x
            y1 = self.general.start_y
            x2 = self.general.end_x
            y2 = self.general.end_y
            assert( ((x1 != x2) and (y1 != y2)) )
        ## rounding
        if self.general.rounding_type == 0:
            x1 = int(np.floor(x1))
            y1 = int(np.floor(y1))
            x2 = int(np.ceil(x2))
            y2 = int(np.ceil(y2))
        else:
            x1 = int(round(x1))
            y1 = int(round(y1))
            x2 = int(round(x2))
            y2 = int(round(y2))
        if self.general.align_w_to_4:
            # x1 = (x1+1) &(~3)  #//+2
            # x2 = (x2+2) &(~3)  #//+1
            x1 = (x1+3) &(~3)  #//+2
            left = w - x2
            left = (left+3) &(~3)
            x2 = w - left
        ## pad_square_to_4
        if str2bool(self.general.pad_square_to_4):
            x1,x2,y1,y2 = pad_square_to_4(x1,x2,y1,y2)
        # do crop
        box = (x1,y1,x2,y2)
        img = img.crop(box)
        # print info
        if str2bool(self.common.print_info):
            self.general.start_x = x1
            self.general.start_y = y1
            self.general.end_x = x2
            self.general.end_y = y2
            self.general.crop_w = x2 - x1
            self.general.crop_h = y2 - y1
            self.print_info()
        # output
        image_data = np.array(img)
        info = {}
        info['box'] = box
        return image_data, info
    ## protect fun
    def _calcuate_xy_center(self, w, h):
        x1 = w/2 - self.general.crop_w / 2
        y1 = h/2 - self.general.crop_h / 2
        x2 = w/2 + self.general.crop_w / 2
        y2 = h/2 + self.general.crop_h / 2
        return x1, y1, x2, y2
    def _calcuate_xy_corner(self, _1, _2):
        x1 = 0
        y1 = 0
        x2 = self.general.crop_w
        y2 = self.general.crop_h
        return x1, y1, x2, y2
    def do_crop(self, image_data, startW, startH, endW, endH):
        return image_data[startH:endH, startW:endW, :]
--- a/kneron/preprocessing/funcs/Normalize.py
+++ b/kneron/preprocessing/funcs/Normalize.py
@ -0,0 +1,186 @@
 import numpy as np
 from .utils import str2bool, str2int, str2float, clip_ary
 class runner(object):
    def __init__(self):
        self.set = {
            'general': {
                'print_info':'no',
                'model_size':[0,0],
                'numerical_type':'floating',
                'type': 'kneron'
            },
            'floating':{
                "scale": 1,
                "bias": 0,
                "mean": "",
                "std": "",
            },
            'hw':{
                "radix":8,
                "shift":"",
                "sub":""
            }
        }
        return
    def update(self, **kwargs):
        #
        self.set.update(kwargs)
        #
        if self.set['general']['numerical_type'] == '520':
            if self.set['general']['type'].lower() in ['TF', 'Tf', 'tf']:
                self.fun_normalize = self._chen_520
                self.shift = 7 - self.set['hw']['radix']
                self.sub = 128
            elif self.set['general']['type'].lower() in ['YOLO', 'Yolo', 'yolo']:
                self.fun_normalize = self._chen_520
                self.shift = 8 - self.set['hw']['radix']
                self.sub = 0
            elif self.set['general']['type'].lower() in ['KNERON', 'Kneron', 'kneron']:
                self.fun_normalize = self._chen_520
                self.shift = 8 - self.set['hw']['radix']
                self.sub = 128
            else:
                self.fun_normalize = self._chen_520
                self.shift = 0
                self.sub = 0      
        elif self.set['general']['numerical_type'] == '720':
                self.fun_normalize = self._chen_720
                self.shift = 0
                self.sub = 0                   
        else:
            if self.set['general']['type'].lower() in ['TORCH', 'Torch', 'torch']:
                self.fun_normalize = self._normalize_torch
                self.set['floating']['scale'] = 255.
                self.set['floating']['mean'] = [0.485, 0.456, 0.406]
                self.set['floating']['std'] = [0.229, 0.224, 0.225]
            elif self.set['general']['type'].lower() in ['TF', 'Tf', 'tf']:
                self.fun_normalize = self._normalize_tf
                self.set['floating']['scale'] = 127.5
                self.set['floating']['bias'] = -1.
            elif self.set['general']['type'].lower() in ['CAFFE', 'Caffe', 'caffe']:
                self.fun_normalize = self._normalize_caffe
                self.set['floating']['mean'] = [103.939, 116.779, 123.68]
            elif self.set['general']['type'].lower() in ['YOLO', 'Yolo', 'yolo']:
                self.fun_normalize = self._normalize_yolo
                self.set['floating']['scale'] = 255.
            elif self.set['general']['type'].lower() in ['KNERON', 'Kneron', 'kneron']:
                self.fun_normalize = self._normalize_kneron
                self.set['floating']['scale'] = 256.
                self.set['floating']['bias'] = -0.5
            else:
                self.fun_normalize = self._normalize_customized
                self.set['floating']['scale'] = str2float(self.set['floating']['scale'])
                self.set['floating']['bias'] = str2float(self.set['floating']['bias'])
                if self.set['floating']['mean'] != None:
                    if len(self.set['floating']['mean']) != 3:
                        self.set['floating']['mean'] = None
                if self.set['floating']['std'] != None:
                    if len(self.set['floating']['std']) != 3:
                        self.set['floating']['std'] = None
    def print_info(self):
        if self.set['general']['numerical_type'] == '520':
            print("<normalize>",
            'numerical_type', self.set['general']['numerical_type'],
            ", type:", self.set['general']['type'],
            ', shift:',self.shift, 
            ', sub:', self.sub)
        else:
            print("<normalize>",
            'numerical_type', self.set['general']['numerical_type'],
            ", type:", self.set['general']['type'],
            ', scale:',self.set['floating']['scale'], 
            ', bias:', self.set['floating']['bias'],
            ', mean:', self.set['floating']['mean'],
            ', std:',self.set['floating']['std'])
    def run(self, image_data):
        # print info
        if str2bool(self.set['general']['print_info']):
            self.print_info()
        # norm
        image_data = self.fun_normalize(image_data)
        # output
        info = {}
        return image_data, info
    def _normalize_torch(self, x):
        if len(x.shape) != 3:
            return x
        x = x.astype('float')
        x = x / self.set['floating']['scale']
        x[..., 0] -= self.set['floating']['mean'][0]
        x[..., 1] -= self.set['floating']['mean'][1]
        x[..., 2] -= self.set['floating']['mean'][2]
        x[..., 0] /= self.set['floating']['std'][0]
        x[..., 1] /= self.set['floating']['std'][1]
        x[..., 2] /= self.set['floating']['std'][2]
        return x
    def _normalize_tf(self, x):
        # print('_normalize_tf')
        x = x.astype('float')
        x = x / self.set['floating']['scale']
        x = x + self.set['floating']['bias']
        return x
    def _normalize_caffe(self, x):
        if len(x.shape) != 3:
            return x
        x = x.astype('float')
        x = x[..., ::-1]
        x[..., 0] -= self.set['floating']['mean'][0]
        x[..., 1] -= self.set['floating']['mean'][1]
        x[..., 2] -= self.set['floating']['mean'][2]
        return x
    def _normalize_yolo(self, x):
        # print('_normalize_yolo')
        x = x.astype('float')
        x = x / self.set['floating']['scale']
        return x
    def _normalize_kneron(self, x):
        # print('_normalize_kneron')
        x = x.astype('float')
        x = x/self.set['floating']['scale']
        x = x + self.set['floating']['bias']
        return x
    def _normalize_customized(self, x):
        # print('_normalize_customized')
        x = x.astype('float')
        if  self.set['floating']['scale'] != 0:
            x = x/ self.set['floating']['scale'] 
        x = x + self.set['floating']['bias'] 
        if self.set['floating']['mean'] is not None:
            x[..., 0] -= self.set['floating']['mean'][0]
            x[..., 1] -= self.set['floating']['mean'][1]
            x[..., 2] -= self.set['floating']['mean'][2]
        if self.set['floating']['std'] is not None:
            x[..., 0] /= self.set['floating']['std'][0]
            x[..., 1] /= self.set['floating']['std'][1]
            x[..., 2] /= self.set['floating']['std'][2]
        return x
    def _chen_520(self, x):
        # print('_chen_520')
        x = (x - self.sub).astype('uint8')
        x = (np.right_shift(x,self.shift))
        x=x.astype('uint8')
        return x
    def _chen_720(self, x):
        # print('_chen_720')
        if self.shift == 1:
            x = x + np.array([[self.sub], [self.sub], [self.sub]])
        else:
            x = x + np.array([[self.sub], [self.sub], [self.sub]])
        return x
--- a/kneron/preprocessing/funcs/Padding.py
+++ b/kneron/preprocessing/funcs/Padding.py
@ -0,0 +1,187 @@
 import numpy as np
 from PIL import Image
 from .utils import str2bool, str2int, str2float
 from .Runner_base import Runner_base, Param_base
 class General(Param_base):
    type = ''
    pad_val = ''
    padded_w = ''
    padded_h = ''
    pad_l = ''
    pad_r = ''
    pad_t = ''
    pad_b = ''
    padding_ch = 3
    padding_ch_type = 'RGB'
    def update(self, **dic):
        self.type = dic['type']
        self.pad_val = dic['pad_val']
        self.padded_w = str2int(dic['padded_w'])
        self.padded_h = str2int(dic['padded_h'])
        self.pad_l = str2int(dic['pad_l'])
        self.pad_r = str2int(dic['pad_r'])
        self.pad_t = str2int(dic['pad_t'])
        self.pad_b = str2int(dic['pad_b'])
    def __str__(self):
        str_out = [
            ', type:',str(self.type),
            ', pad_val:',str(self.pad_val),
            ', pad_l:',str(self.pad_l),
            ', pad_r:',str(self.pad_r),
            ', pad_r:',str(self.pad_t),
            ', pad_b:',str(self.pad_b),
            ', padding_ch:',str(self.padding_ch)]
        return(' '.join(str_out))
 class Hw(Param_base):
    radix = 8
    normalize_type = 'floating'
    def update(self, **dic):
        self.radix = dic['radix']
        self.normalize_type = dic['normalize_type']
    def __str__(self):
        str_out = [
            ', radix:', str(self.radix),
            ', normalize_type:',str(self.normalize_type)]
        return(' '.join(str_out))
 class runner(Runner_base):
    ## overwrite the class in Runner_base
    general = General()
    hw = Hw()
    def __str__(self):
        return('<Padding>')
    def update(self, **kwargs):
        super().update(**kwargs)
        ## update pad type & pad length
        if (self.general.pad_l != 0) or (self.general.pad_r != 0) or (self.general.pad_t != 0) or (self.general.pad_b != 0):
            self.general.type = 'specific'
            assert(self.general.pad_l >= 0)
            assert(self.general.pad_r >= 0)
            assert(self.general.pad_t >= 0)
            assert(self.general.pad_b >= 0)
        elif(self.general.type != 'specific'):
            if self.general.padded_w == 0 or self.general.padded_h == 0:
                self.general.padded_w = self.common.model_size[0]
                self.general.padded_h = self.common.model_size[1]
            assert(self.general.padded_w > 0)
            assert(self.general.padded_h > 0)
            assert(self.general.type.lower() in ['CENTER', 'Center', 'center', 'CORNER', 'Corner', 'corner'])
        else:
            assert(self.general.type == 'specific')
        ## decide pad_val & padding ch
        # if numerical_type is floating
        if (self.common.numerical_type == 'floating'):
            if self.general.pad_val != 'edge':
                self.general.pad_val = str2float(self.general.pad_val)
            self.general.padding_ch = 3
            self.general.padding_ch_type = 'RGB'
        # if numerical_type is 520 or 720
        else: 
            if self.general.pad_val == '':
                if self.hw.normalize_type.lower() in ['TF', 'Tf', 'tf']:
                    self.general.pad_val = np.uint8(-128 >> (7 - self.hw.radix))
                elif self.hw.normalize_type.lower() in ['YOLO', 'Yolo', 'yolo']:
                    self.general.pad_val = np.uint8(0 >> (8 - self.hw.radix))
                elif self.hw.normalize_type.lower() in ['KNERON', 'Kneron', 'kneron']:
                    self.general.pad_val = np.uint8(-128 >> (8 - self.hw.radix))
                else:
                    self.general.pad_val = np.uint8(0 >> (8 - self.hw.radix))
            else:
                self.general.pad_val = str2int(self.general.pad_val)
            self.general.padding_ch = 4
            self.general.padding_ch_type = 'RGBA'
    def run(self, image_data):
        # init
        shape = image_data.shape
        w = shape[1]
        h = shape[0]
        if len(shape) < 3:
            self.general.padding_ch = 1
            self.general.padding_ch_type = 'L'
        else:
            if shape[2] == 3 and self.general.padding_ch == 4:
                image_data = np.concatenate((image_data, np.zeros((h, w, 1), dtype=np.uint8) ), axis=2)
        ## padding
        if self.general.type.lower() in ['CENTER',  'Center',  'center']:
            img_pad = self._padding_center(image_data, w, h)
        elif self.general.type.lower() in ['CORNER',  'Corner',  'corner']:
            img_pad = self._padding_corner(image_data, w, h)
        else:
            img_pad = self._padding_sp(image_data, w, h)
        # print info
        if str2bool(self.common.print_info):
            self.print_info()
        # output
        info = {}
        return img_pad, info
    ## protect fun
    def _padding_center(self, img, ori_w, ori_h):
        # img_pad = Image.new(self.general.padding_ch_type, (self.general.padded_w, self.general.padded_h), int(self.general.pad_val[0]))
        # img = Image.fromarray(img)
        # img_pad.paste(img, ((self.general.padded_w-ori_w)//2, (self.general.padded_h-ori_h)//2))
        # return img_pad
        padH = self.general.padded_h - ori_h
        padW = self.general.padded_w - ori_w
        self.general.pad_t = padH // 2
        self.general.pad_b = (padH // 2) + (padH % 2)
        self.general.pad_l = padW // 2
        self.general.pad_r = (padW // 2) + (padW % 2)
        if self.general.pad_l < 0 or self.general.pad_r <0 or self.general.pad_t <0 or self.general.pad_b<0:
            return img
        img_pad = self._padding_sp(img,ori_w,ori_h)
        return img_pad
    def _padding_corner(self, img, ori_w, ori_h):
        # img_pad = Image.new(self.general.padding_ch_type, (self.general.padded_w, self.general.padded_h), self.general.pad_val)
        # img_pad.paste(img, (0, 0))
        self.general.pad_l = 0
        self.general.pad_r = self.general.padded_w - ori_w
        self.general.pad_t = 0
        self.general.pad_b = self.general.padded_h - ori_h
        if self.general.pad_l < 0 or self.general.pad_r <0 or self.general.pad_t <0 or self.general.pad_b<0:
            return img
        img_pad = self._padding_sp(img,ori_w,ori_h)
        return img_pad
    def _padding_sp(self, img, ori_w, ori_h):
        # block_t = np.zeros((self.general.pad_t, self.general.pad_l + self.general.pad_r + ori_w, self.general.padding_ch), dtype=np.float)
        # block_l = np.zeros((ori_h, self.general.pad_l, self.general.padding_ch), dtype=np.float)
        # block_r = np.zeros((ori_h, self.general.pad_r, self.general.padding_ch), dtype=np.float)
        # block_b = np.zeros((self.general.pad_b, self.general.pad_l + self.general.pad_r + ori_w, self.general.padding_ch), dtype=np.float)
        # for i in range(self.general.padding_ch):
        #     block_t[:, :, i] = np.ones(block_t[:, :, i].shape, dtype=np.float) * self.general.pad_val
        #     block_l[:, :, i] = np.ones(block_l[:, :, i].shape, dtype=np.float) * self.general.pad_val
        #     block_r[:, :, i] = np.ones(block_r[:, :, i].shape, dtype=np.float) * self.general.pad_val
        #     block_b[:, :, i] = np.ones(block_b[:, :, i].shape, dtype=np.float) * self.general.pad_val
        # padded_image_hor = np.concatenate((block_l, img, block_r), axis=1)
        # padded_image = np.concatenate((block_t, padded_image_hor, block_b), axis=0)
        # return padded_image
        if self.general.padding_ch == 1:
            pad_range = ( (self.general.pad_t, self.general.pad_b),(self.general.pad_l, self.general.pad_r) )
        else:
            pad_range = ((self.general.pad_t, self.general.pad_b),(self.general.pad_l, self.general.pad_r),(0,0))
        if isinstance(self.general.pad_val, str):
            if self.general.pad_val == 'edge':
                padded_image = np.pad(img, pad_range, mode="edge")
            else:
                padded_image = np.pad(img, pad_range, mode="constant",constant_values=0)
        else:
            padded_image = np.pad(img, pad_range, mode="constant",constant_values=self.general.pad_val)
        return padded_image
--- a/kneron/preprocessing/funcs/Resize.py
+++ b/kneron/preprocessing/funcs/Resize.py
@ -0,0 +1,237 @@
 import numpy as np
 import cv2
 from PIL import Image
 from .utils import str2bool, str2int
 from ctypes import c_float
 from .Runner_base import Runner_base, Param_base
 class General(Param_base):
    type = 'bilinear'
    keep_ratio = True
    zoom = True
    calculate_ratio_using_CSim = True
    resize_w = 0
    resize_h = 0
    resized_w = 0
    resized_h = 0
    def update(self, **dic):
        self.type = dic['type']
        self.keep_ratio = str2bool(dic['keep_ratio'])
        self.zoom = str2bool(dic['zoom'])
        self.calculate_ratio_using_CSim = str2bool(dic['calculate_ratio_using_CSim'])
        self.resize_w = str2int(dic['resize_w'])
        self.resize_h = str2int(dic['resize_h'])
    def __str__(self):
        str_out = [
            ', type:',str(self.type),
            ', keep_ratio:',str(self.keep_ratio),
            ', zoom:',str(self.zoom),
            ', calculate_ratio_using_CSim:',str(self.calculate_ratio_using_CSim),
            ', resize_w:',str(self.resize_w),
            ', resize_h:',str(self.resize_h),
            ', resized_w:',str(self.resized_w),
            ', resized_h:',str(self.resized_h)]
        return(' '.join(str_out))
 class Hw(Param_base):
    resize_bit = 12
    def update(self, **dic):
        pass
    def __str__(self):
        str_out = [
            ', resize_bit:',str(self.resize_bit)]
        return(' '.join(str_out))
 class runner(Runner_base):
    ## overwrite the class in Runner_base
    general = General()
    hw = Hw()
    def __str__(self):
        return('<Resize>')
    def update(self, **kwargs):
        super().update(**kwargs)
        ## if resize size has not been assigned, then it will take model size as resize size
        if self.general.resize_w == 0 or self.general.resize_h == 0:
            self.general.resize_w = self.common.model_size[0]
            self.general.resize_h = self.common.model_size[1]
        assert(self.general.resize_w > 0)
        assert(self.general.resize_h > 0)
        ##
        if self.common.numerical_type == '520':
            self.general.type = 'fixed_520'
        elif self.common.numerical_type == '720':
            self.general.type = 'fixed_720'
        assert(self.general.type.lower() in ['BILINEAR',  'Bilinear',  'bilinear', 'BICUBIC',  'Bicubic',  'bicubic', 'FIXED',  'Fixed', 'fixed', 'FIXED_520',  'Fixed_520',  'fixed_520', 'FIXED_720', 'Fixed_720', 'fixed_720','CV', 'cv', 'opencv', 'OpenCV', 'CV2', 'cv2'])
    def run(self, image_data):
        ## init
        ori_w = image_data.shape[1]
        ori_h = image_data.shape[0]
        info = {}
        ##
        if self.general.keep_ratio:
            self.general.resized_w, self.general.resized_h = self.calcuate_scale_keep_ratio(self.general.resize_w,self.general.resize_h, ori_w, ori_h, self.general.calculate_ratio_using_CSim)
        else:
            self.general.resized_w = int(self.general.resize_w)
            self.general.resized_h = int(self.general.resize_h)
        assert(self.general.resized_w > 0)
        assert(self.general.resized_h > 0)
        ##
        if (self.general.resized_w > ori_w) or (self.general.resized_h > ori_h):
            if not self.general.zoom: 
                info['size'] = (ori_w,ori_h)
                if str2bool(self.common.print_info):
                    print('no resize')
                    self.print_info()
                return image_data, info
        ## resize
        if self.general.type.lower() in ['BILINEAR',  'Bilinear',  'bilinear']:
            image_data = self.do_resize_bilinear(image_data, self.general.resized_w, self.general.resized_h)
        elif self.general.type.lower() in ['BICUBIC',  'Bicubic',  'bicubic']:
            image_data = self.do_resize_bicubic(image_data, self.general.resized_w, self.general.resized_h)
        elif self.general.type.lower() in ['CV',  'cv',  'opencv', 'OpenCV',  'CV2',  'cv2']:
            image_data = self.do_resize_cv2(image_data, self.general.resized_w, self.general.resized_h)
        elif self.general.type.lower() in ['FIXED',  'Fixed',  'fixed', 'FIXED_520',  'Fixed_520',  'fixed_520', 'FIXED_720', 'Fixed_720', 'fixed_720']:
            image_data = self.do_resize_fixed(image_data, self.general.resized_w, self.general.resized_h, self.hw.resize_bit, self.general.type)
        # output
        info['size'] = (self.general.resized_w, self.general.resized_h)
        # print info
        if str2bool(self.common.print_info):
            self.print_info()
        return image_data, info
    def calcuate_scale_keep_ratio(self, tar_w, tar_h, ori_w, ori_h, calculate_ratio_using_CSim):
        if not calculate_ratio_using_CSim:
            scale_w = tar_w * 1.0 / ori_w*1.0
            scale_h = tar_h * 1.0 / ori_h*1.0
            scale = scale_w if scale_w < scale_h else scale_h
            new_w = int(round(ori_w * scale))
            new_h = int(round(ori_h * scale))
            return new_w, new_h
        ## calculate_ratio_using_CSim
        scale_w = c_float(tar_w * 1.0 / (ori_w * 1.0)).value
        scale_h = c_float(tar_h * 1.0 / (ori_h * 1.0)).value
        scale_ratio = 0.0
        scale_target_w = 0
        scale_target_h = 0
        padH = 0
        padW = 0
        bScaleW = True if scale_w < scale_h else False
        if bScaleW:
            scale_ratio = scale_w
            scale_target_w = int(c_float(scale_ratio * ori_w + 0.5).value)
            scale_target_h = int(c_float(scale_ratio * ori_h + 0.5).value)
            assert (abs(scale_target_w - tar_w) <= 1), "Error: scale down width cannot meet expectation\n"
            padH = tar_h - scale_target_h
            padW = 0
            assert (padH >= 0), "Error: padH shouldn't be less than zero\n"
        else:
            scale_ratio = scale_h 
            scale_target_w = int(c_float(scale_ratio * ori_w + 0.5).value)
            scale_target_h = int(c_float(scale_ratio * ori_h + 0.5).value)
            assert (abs(scale_target_h - tar_h) <= 1), "Error: scale down height cannot meet expectation\n"
            padW = tar_w - scale_target_w
            padH = 0
            assert (padW >= 0), "Error: padW shouldn't be less than zero\n"
        new_w = tar_w - padW
        new_h = tar_h - padH
        return new_w, new_h
    def do_resize_bilinear(self, image_data, resized_w, resized_h):
        img = Image.fromarray(image_data)
        img = img.resize((resized_w, resized_h), Image.BILINEAR)
        image_data = np.array(img).astype('uint8')
        return image_data        
    def do_resize_bicubic(self, image_data, resized_w, resized_h):
        img = Image.fromarray(image_data)
        img = img.resize((resized_w, resized_h), Image.BICUBIC)
        image_data = np.array(img).astype('uint8')
        return image_data
    def do_resize_cv2(self, image_data, resized_w, resized_h):
        image_data = cv2.resize(image_data, (resized_w, resized_h))
        image_data = np.array(image_data)
        # image_data = np.array(image_data).astype('uint8')
        return image_data
    def do_resize_fixed(self, image_data, resized_w, resized_h, resize_bit, type):
        if len(image_data.shape) < 3:
            m, n = image_data.shape
            tmp = np.zeros((m,n,3), dtype=np.uint8)
            tmp[:,:,0] = image_data
            image_data = tmp
            c = 3
            gray = True
        else:
            m, n, c = image_data.shape
            gray = False
        resolution = 1 << resize_bit
        # Width
        ratio = int(((n - 1) << resize_bit) / (resized_w - 1))
        ratio_cnt = 0
        src_x = 0
        resized_image_w = np.zeros((m, resized_w, c), dtype=np.uint8)
        for dst_x in range(resized_w):
            while ratio_cnt > resolution:
                ratio_cnt = ratio_cnt - resolution
                src_x = src_x + 1
            mul1 = np.ones((m, c)) * (resolution - ratio_cnt)
            mul2 = np.ones((m, c)) * ratio_cnt
            resized_image_w[:, dst_x, :] = np.multiply(np.multiply(
                image_data[:, src_x, :], mul1) + np.multiply(image_data[:, src_x + 1, :], mul2), 1/resolution)
            ratio_cnt = ratio_cnt + ratio
        # Height
        ratio = int(((m - 1) << resize_bit) / (resized_h - 1))
        ## NPU HW special case 2 , only on 520
        if type.lower() in ['FIXED_520',  'Fixed_520',  'fixed_520']:
            if (((ratio * (resized_h - 1)) % 4096 == 0) and ratio != 4096):
                ratio -= 1
        ratio_cnt = 0
        src_x = 0
        resized_image = np.zeros(
            (resized_h, resized_w, c), dtype=np.uint8)
        for dst_x in range(resized_h):
            while ratio_cnt > resolution:
                ratio_cnt = ratio_cnt - resolution
                src_x = src_x + 1
            mul1 = np.ones((resized_w, c)) * (resolution - ratio_cnt)
            mul2 = np.ones((resized_w, c)) * ratio_cnt
            ## NPU HW special case 1 , both on 520 / 720
            if (((dst_x > 0) and ratio_cnt == resolution) and (ratio != resolution)):
                if type.lower() in ['FIXED_520',  'Fixed_520',  'fixed_520','FIXED_720',  'Fixed_720',  'fixed_720' ]:
                    resized_image[dst_x, :, :] = np.multiply(np.multiply(
                        resized_image_w[src_x+1, :, :], mul1) + np.multiply(resized_image_w[src_x + 2, :, :], mul2), 1/resolution)
            else:
                resized_image[dst_x, :, :] = np.multiply(np.multiply(
                    resized_image_w[src_x, :, :], mul1) + np.multiply(resized_image_w[src_x + 1, :, :], mul2), 1/resolution)
            ratio_cnt = ratio_cnt + ratio
        if gray:
            resized_image = resized_image[:,:,0]
        return resized_image
--- a/kneron/preprocessing/funcs/Rotate.py
+++ b/kneron/preprocessing/funcs/Rotate.py
@ -0,0 +1,45 @@
 import numpy as np
 from .utils import str2bool, str2int
 class runner(object):
    def __init__(self, *args, **kwargs):
        self.set = {
            'operator': '',
            "rotate_direction": 0,
        }
        self.update(*args, **kwargs)
    def update(self, *args, **kwargs):
        self.set.update(kwargs)
        self.rotate_direction = str2int(self.set['rotate_direction'])
        # print info
        if str2bool(self.set['b_print']):
            self.print_info()
    def print_info(self):
        print("<rotate>",
            'rotate_direction', self.rotate_direction,)
    def run(self, image_data):
        image_data = self._rotate(image_data)
        return image_data
    def _rotate(self,img):
        if self.rotate_direction == 1 or self.rotate_direction == 2:
            col, row, unit = img.shape
            pInBuf = img.reshape((-1,1))
            pOutBufTemp = np.zeros((col* row* unit))
            for r in range(row):
                for c in range(col):
                    for u in range(unit):
                        if self.rotate_direction == 1:
                            pOutBufTemp[unit * (c * row + (row - r - 1))+u] = pInBuf[unit * (r * col + c)+u]
                        elif self.rotate_direction == 2:
                            pOutBufTemp[unit * (row * (col - c - 1) + r)+u] = pInBuf[unit * (r * col + c)+u]
            img = pOutBufTemp.reshape((col,row,unit))
        return img
--- a/kneron/preprocessing/funcs/Runner_base.py
+++ b/kneron/preprocessing/funcs/Runner_base.py
@ -0,0 +1,59 @@
 from abc import ABCMeta, abstractmethod
 class Param_base(object):
    @abstractmethod
    def update(self,**dic):
        raise NotImplementedError("Must override")
    def load_dic(self, key, **dic):
        if key in dic:
            param = eval('self.'+key)
            param = dic[key]
    def __str__(self):
        str_out = []
        return(' '.join(str_out))
 class Common(Param_base):
    print_info = False
    model_size = [0,0]
    numerical_type = 'floating'
    def update(self, **dic):
        self.print_info = dic['print_info']
        self.model_size = dic['model_size']
        self.numerical_type = dic['numerical_type']
    def __str__(self):
        str_out = ['numerical_type:',str(self.numerical_type)]
        return(' '.join(str_out))
 class Runner_base(metaclass=ABCMeta):
    common = Common()
    general = Param_base()
    floating = Param_base()
    hw = Param_base()
    def update(self, **kwargs):
        ## update param
        self.common.update(**kwargs['common'])
        self.general.update(**kwargs['general'])
        assert(self.common.numerical_type.lower() in ['floating', '520', '720'])
        if (self.common.numerical_type == 'floating'):
            if (self.floating.__class__.__name__ != 'Param_base'):
                self.floating.update(**kwargs['floating'])
        else:
            if (self.hw.__class__.__name__ != 'Param_base'):
                self.hw.update(**kwargs['hw'])
    def print_info(self):
        if (self.common.numerical_type == 'floating'):
            print(self, self.common, self.general, self.floating)
        else:
            print(self, self.common, self.general, self.hw)
--- a/kneron/preprocessing/funcs/init.py
+++ b/kneron/preprocessing/funcs/init.py
@ -0,0 +1,2 @@
 from . import ColorConversion, Padding, Resize, Crop, Normalize, Rotate
--- a/kneron/preprocessing/funcs/utils.py
+++ b/kneron/preprocessing/funcs/utils.py
@ -0,0 +1,372 @@
 import numpy as np
 from PIL import Image
 import struct
 def pad_square_to_4(x_start, x_end, y_start, y_end):
    w_int = x_end - x_start 
    h_int = y_end - y_start
    pad = w_int - h_int
    if pad > 0:
        pad_s = (pad >> 1) &(~3)
        pad_e = pad - pad_s
        y_start -= pad_s
        y_end += pad_e
    else:#//pad <=0
        pad_s = -(((pad) >> 1) &(~3))
        pad_e = (-pad) - pad_s
        x_start -= pad_s
        x_end += pad_e
    return x_start, x_end, y_start, y_end
 def str_fill(value):
    if len(value) == 1:
        value = "0" + value
    elif len(value) == 0:
        value = "00"
    return value
 def clip_ary(value):
    list_v = []
    for i in range(len(value)):
        v = value[i] % 256
        list_v.append(v)
    return list_v
 def str2bool(v):
    if isinstance(v,bool):
        return v
    return v.lower() in ('TRUE', 'True', 'true', '1', 'T', 't', 'Y', 'YES', 'y', 'yes')
 def str2int(s):
    if s == "":
        s = 0
    s = int(s)
    return s
 def str2float(s):
    if s == "":
        s = 0
    s = float(s)
    return s
 def clip(value, mini, maxi):
    if value < mini:
        result = mini
    elif value > maxi:
        result = maxi
    else:
        result = value
    return result
 def clip_ary(value):
    list_v = []
    for i in range(len(value)):
        v = value[i] % 256
        list_v.append(v)
    return list_v
 def signed_rounding(value, bit):
    if value < 0:
        value = value - (1 << (bit - 1))
    else:
        value = value + (1 << (bit - 1))
    return value
 def hex_loader(data_folder,**kwargs):
    format_mode = kwargs['raw_img_fmt']
    src_h = kwargs['img_in_height']
    src_w = kwargs['img_in_width']
    if format_mode in ['YUV444', 'yuv444', 'YCBCR444', 'YCbCr444', 'ycbcr444']:
        output = hex_yuv444(data_folder,src_h,src_w)
    elif format_mode in ['RGB565', 'rgb565']:
        output = hex_rgb565(data_folder,src_h,src_w)
    elif format_mode in ['YUV422', 'yuv422', 'YCBCR422', 'YCbCr422', 'ycbcr422']:
        output = hex_yuv422(data_folder,src_h,src_w)
    return output
 def hex_rgb565(hex_folder,src_h,src_w):
    pix_per_line = 8
    byte_per_line = 16
    f = open(hex_folder)
    pixel_r = []
    pixel_g = []
    pixel_b = []
    # Ignore the first line
    f.readline()
    input_line = int((src_h * src_w)/pix_per_line)
    for i in range(input_line):
        readline = f.readline()
        for j in range(int(byte_per_line/2)-1, -1, -1):
            data1 = int(readline[(j * 4 + 0):(j * 4 + 2)], 16)
            data0 = int(readline[(j * 4 + 2):(j * 4 + 4)], 16)
            r = ((data1 & 0xf8) >> 3)
            g = (((data0 & 0xe0) >> 5) + ((data1 & 0x7) << 3))
            b = (data0 & 0x1f)
            pixel_r.append(r)
            pixel_g.append(g)
            pixel_b.append(b)
    ary_r = np.array(pixel_r, dtype=np.uint8)
    ary_g = np.array(pixel_g, dtype=np.uint8)
    ary_b = np.array(pixel_b, dtype=np.uint8)
    output = np.concatenate((ary_r[:, None], ary_g[:, None], ary_b[:, None]), axis=1)
    output = output.reshape((src_h, src_w, 3))
    return output
 def hex_yuv444(hex_folder,src_h,src_w):
    pix_per_line = 4
    byte_per_line = 16
    f = open(hex_folder)
    byte0 = []
    byte1 = []
    byte2 = []
    byte3 = []
    # Ignore the first line
    f.readline()
    input_line = int((src_h * src_w)/pix_per_line)
    for i in range(input_line):
        readline = f.readline()
        for j in range(byte_per_line-1, -1, -1):
            data = int(readline[(j*2):(j*2+2)], 16)
            if (j+1) % 4 == 0:
                byte0.append(data)
            elif (j+2) % 4 == 0:
                byte1.append(data)
            elif (j+3) % 4 == 0:
                byte2.append(data)
            elif (j+4) % 4 == 0:
                byte3.append(data)
    # ary_a = np.array(byte0, dtype=np.uint8)
    ary_v = np.array(byte1, dtype=np.uint8)
    ary_u = np.array(byte2, dtype=np.uint8)
    ary_y = np.array(byte3, dtype=np.uint8)
    output = np.concatenate((ary_y[:, None], ary_u[:, None], ary_v[:, None]), axis=1)
    output = output.reshape((src_h, src_w, 3))
    return output
 def hex_yuv422(hex_folder,src_h,src_w):
    pix_per_line = 8
    byte_per_line = 16
    f = open(hex_folder)
    pixel_y = []
    pixel_u = []
    pixel_v = []
    # Ignore the first line
    f.readline()
    input_line = int((src_h * src_w)/pix_per_line)
    for i in range(input_line):
        readline = f.readline()
        for j in range(int(byte_per_line/4)-1, -1, -1):
            data3 = int(readline[(j * 8 + 0):(j * 8 + 2)], 16)
            data2 = int(readline[(j * 8 + 2):(j * 8 + 4)], 16)
            data1 = int(readline[(j * 8 + 4):(j * 8 + 6)], 16)
            data0 = int(readline[(j * 8 + 6):(j * 8 + 8)], 16)
            pixel_y.append(data3)
            pixel_y.append(data1)
            pixel_u.append(data2)
            pixel_u.append(data2)
            pixel_v.append(data0)
            pixel_v.append(data0)
    ary_y = np.array(pixel_y, dtype=np.uint8)
    ary_u = np.array(pixel_u, dtype=np.uint8)
    ary_v = np.array(pixel_v, dtype=np.uint8)
    output = np.concatenate((ary_y[:, None], ary_u[:, None], ary_v[:, None]), axis=1)
    output = output.reshape((src_h, src_w, 3))
    return output
 def bin_loader(data_folder,**kwargs):
    format_mode = kwargs['raw_img_fmt']
    src_h = kwargs['img_in_height']
    src_w = kwargs['img_in_width']
    if format_mode in ['YUV','yuv','YUV444', 'yuv444', 'YCBCR','YCbCr','ycbcr','YCBCR444', 'YCbCr444', 'ycbcr444']:
        output = bin_yuv444(data_folder,src_h,src_w)
    elif format_mode in ['RGB565', 'rgb565']:
        output = bin_rgb565(data_folder,src_h,src_w)
    elif format_mode in ['NIR', 'nir','NIR888', 'nir888']:
        output = bin_nir(data_folder,src_h,src_w)
    elif format_mode in ['YUV422', 'yuv422', 'YCBCR422', 'YCbCr422', 'ycbcr422']:
        output = bin_yuv422(data_folder,src_h,src_w)
    elif format_mode in ['RGB888','rgb888']:
        output = np.fromfile(data_folder, dtype='uint8')
        output = output.reshape(src_h,src_w,3)
    elif format_mode in ['RGBA8888','rgba8888', 'RGBA' , 'rgba']:
        output_temp = np.fromfile(data_folder, dtype='uint8')
        output_temp = output_temp.reshape(src_h,src_w,4)
        output = output_temp[:,:,0:3]
    return output
 def bin_yuv444(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    row = src_h
    col = src_w
    pixels = row*col
    raw = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            raw.append(s[0])
    raw = raw[:pixels*4]
    #
    output = np.zeros((pixels * 3), dtype=np.uint8)
    cnt = 0
    for i in range(0, pixels*4, 4):
        #Y
        output[cnt] = raw[i+3]
        #U
        cnt += 1
        output[cnt] = raw[i+2]
        #V
        cnt += 1
        output[cnt] = raw[i+1]
        cnt += 1          
    output = output.reshape((src_h,src_w,3))
    return output
 def bin_yuv422(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    row = src_h
    col = src_w
    pixels = row*col
    raw = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            raw.append(s[0])
    raw = raw[:pixels*2]
    #
    output = np.zeros((pixels * 3), dtype=np.uint8)
    cnt = 0
    for i in range(0, pixels*2, 4):
        #Y0
        output[cnt] = raw[i+3]
        #U0
        cnt += 1
        output[cnt] = raw[i+2]
        #V0
        cnt += 1
        output[cnt] = raw[i]
        #Y1
        cnt += 1
        output[cnt] = raw[i+1]
        #U1
        cnt += 1
        output[cnt] = raw[i+2]
        #V1
        cnt += 1
        output[cnt] = raw[i]
        cnt += 1          
    output = output.reshape((src_h,src_w,3))
    return output
 def bin_rgb565(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    row = src_h
    col = src_w
    pixels = row*col
    rgba565 = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            rgba565.append(s[0])
    rgba565 = rgba565[:pixels*2]
    # rgb565_bin to numpy_array
    output = np.zeros((pixels * 3), dtype=np.uint8)
    cnt = 0
    for i in range(0, pixels*2, 2):
        temp = rgba565[i]
        temp2 = rgba565[i+1]
        #R-5
        output[cnt] = (temp2 >>3)
        #G-6
        cnt += 1
        output[cnt] = ((temp & 0xe0) >> 5) + ((temp2 & 0x07) << 3)
        #B-5
        cnt += 1
        output[cnt] = (temp & 0x1f)
        cnt += 1          
    output = output.reshape((src_h,src_w,3))
    return output
 def bin_nir(in_img_path,src_h,src_w):
    # load bin
    struct_fmt = '1B' 
    struct_len = struct.calcsize(struct_fmt)
    struct_unpack = struct.Struct(struct_fmt).unpack_from
    nir = []
    with open(in_img_path, "rb") as f:
        while True:
            data = f.read(struct_len)
            if not data: break
            s = struct_unpack(data)
            nir.append(s[0])
    nir = nir[:src_h*src_w]
    pixels = len(nir)
    # nir_bin to numpy_array
    output = np.zeros((len(nir) * 3), dtype=np.uint8)
    for i in range(0, pixels):
        output[i*3]=nir[i]
        output[i*3+1]=nir[i]
        output[i*3+2]=nir[i]
    output = output.reshape((src_h,src_w,3))
    return output
--- a/kneron/preprocessing/funcs/utils_520.py
+++ b/kneron/preprocessing/funcs/utils_520.py
@ -0,0 +1,50 @@
 import math
 def round_up_16(num):
    return ((num + (16 - 1)) & ~(16 - 1))
 def round_up_n(num, n):
    if (num > 0):
        temp = float(num) / n
        return math.ceil(temp) * n
    else:
        return -math.ceil(float(-num) / n) * n
 def cal_img_row_offset(crop_num, pad_num, start_row, out_row, orig_row):
    scaled_img_row = int(out_row - (pad_num[1] + pad_num[3]))
    if ((start_row - pad_num[1]) > 0):
        img_str_row = int((start_row - pad_num[1]))
    else:
        img_str_row = 0
    valid_row = int(orig_row - (crop_num[1] + crop_num[3]))
    img_str_row = int(valid_row * img_str_row / scaled_img_row)
    return int(img_str_row + crop_num[1])
 def get_pad_num(pad_num_orig, left, up, right, bottom):
    pad_num = [0]*4
    for i in range(0,4):
        pad_num[i] = pad_num_orig[i]
    if not (left):
        pad_num[0] = 0
    if not (up):
        pad_num[1] = 0
    if not (right):
        pad_num[2] = 0
    if not (bottom):
        pad_num[3] = 0
    return pad_num
 def get_byte_per_pixel(raw_fmt):
    if raw_fmt.lower() in ['RGB888', 'rgb888', 'RGB', 'rgb888']:
        return 4
    elif raw_fmt.lower() in ['YUV', 'yuv', 'YUV422', 'yuv422']:
        return 2
    elif raw_fmt.lower() in ['RGB565', 'rgb565']:
        return 2
    elif raw_fmt.lower() in ['NIR888', 'nir888', 'NIR', 'nir']:
        return 1
    else:
        return -1
--- a/kneron/preprocessing/funcs/utils_720.py
+++ b/kneron/preprocessing/funcs/utils_720.py
@ -0,0 +1,42 @@
 import numpy as np
 from PIL import Image
 def twos_complement(value):
    value = int(value)
    # msb = (value & 0x8000) * (1/np.power(2, 15))
    msb = (value & 0x8000) >> 15
    if msb == 1:
        if (((~value) & 0xFFFF) + 1) >= 0xFFFF:
            result = ((~value) & 0xFFFF)
        else:
            result = (((~value) & 0xFFFF) + 1)
        result = result * (-1)
    else:
        result = value
    return result
 def twos_complement_pix(value):
    h, _ = value.shape
    for i in range(h):
        value[i, 0] = twos_complement(value[i, 0])
    return value
 def clip(value, mini, maxi):
    if value < mini:
        result = mini
    elif value > maxi:
        result = maxi
    else:
        result = value
    return result
 def clip_pix(value, mini, maxi):
    h, _ = value.shape
    for i in range(h):
        value[i, 0] = clip(value[i, 0], mini, maxi)
    return value
--- a/kneron/quantize_yolov5.py
+++ b/kneron/quantize_yolov5.py
@ -0,0 +1,66 @@
 import os
 import numpy as np
 import torch
 import ktc  # Kneron Toolchain
 import onnx
 from yolov5_preprocess import Yolov5_preprocess
 import kneron_preprocessing
 # 設定裝置
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 # 設定圖片大小
 imgsz_h, imgsz_w = 640, 640
 # 設定正確的數據目錄
 data_path = "/workspace/yolov5/data50"
 # 確保 data50 內有圖片
 files_found = [f for _, _, files in os.walk(data_path) for f in files if f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp"))]
 if not files_found:
    raise FileNotFoundError(f"❌ Error: No images found in {data_path}! Please check your dataset.")
 print(f"✅ Found {len(files_found)} images in {data_path}")
 # **獲取 ONNX 模型的輸入名稱**
 onnx_model_path = "/workspace/yolov5/runs/train/exp24/weights/latest.opt.onnx"
 m = onnx.load(onnx_model_path)
 input_name = m.graph.input[0].name  # 確保 key 與 ONNX input name 一致
 km = ktc.ModelConfig(20008, "0001", "720", onnx_model=onnx_model_path)
 # 存儲預處理後的圖片數據
 img_list = []
 # 遍歷 data50 並進行預處理
 for root, _, files in os.walk(data_path):
    for f in files:
        fullpath = os.path.join(root, f)
        # **只處理圖片文件**
        if not f.lower().endswith((".jpg", ".jpeg", ".png", ".bmp")):
            print(f"⚠️ Skipping non-image file: {fullpath}")
            continue
        # **嘗試處理圖片**
        try:
            img_data, _ = Yolov5_preprocess(fullpath, device, imgsz_h, imgsz_w)
            img_data = img_data.cpu().numpy()
            print(f"✅ Processed: {fullpath}")
            img_list.append(img_data)
        except Exception as e:
            print(f"❌ Failed to process {fullpath}: {e}")
 # **確保 img_list 不是空的**
 if not img_list:
    raise ValueError("❌ Error: No valid images were processed! Please check the image paths and formats.")
 # **執行 BIE 量化**
 bie_model_path = km.analysis({input_name: img_list})
 # **確認 BIE 模型是否生成**
 if not os.path.exists(bie_model_path):
    raise RuntimeError(f"❌ Error: BIE model was not generated! Please check your quantization process.")
 # 顯示成功訊息
 print("\n✅ Fixed-point analysis done! BIE model saved to:", bie_model_path)
--- a/kneron/removenode.py
+++ b/kneron/removenode.py
@ -0,0 +1,34 @@
 import onnx
 import ktc.onnx_optimizer as kneron_opt
 from onnx import helper
 def replace_sigmoid_with_identity(model):
    """
    Replaces all Sigmoid nodes with Identity nodes to maintain model integrity.
    """
    for node in model.graph.node:
        if node.op_type == "Sigmoid":
            print(f"Replacing {node.name} with Identity")
            identity_node = helper.make_node(
                "Identity",
                inputs=node.input,
                outputs=node.output,
                name=node.name + "_identity"
            )
            model.graph.node.extend([identity_node])
            model.graph.node.remove(node)
    return model
 def process_onnx(input_onnx_path, output_onnx_path):
    """ Replaces Sigmoid with Identity and saves the new model """
    model = onnx.load(input_onnx_path)
    model = replace_sigmoid_with_identity(model)
    onnx.save(model, output_onnx_path)
    print(f"Modified ONNX model saved to: {output_onnx_path}")
 # 使用 Docker 掛載的路徑
 input_onnx = "/workspace/yolov5/runs/train/exp24/weights/best_simplified.onnx"
 output_onnx = "/workspace/yolov5/runs/train/exp24/weights/best_no_sigmoid.onnx"
 process_onnx(input_onnx, output_onnx)
--- a/kneron/yolov5_export.py
+++ b/kneron/yolov5_export.py
@ -0,0 +1,80 @@
 import os
 import torch
 import sys
 import yaml
 import argparse
 from yolov5_runner import Yolov5Runner
 def save_weight(num_classes): 
    current_path=os.getcwd()
    par_path = os.path.dirname(current_path)
    sys.path.append(os.path.join(par_path, 'yolov5'))
    from models.yolo import Model  
    num_classes = num_classes 
    device=torch.device('cpu')
    ckpt = torch.load(path, map_location=device)
    model = Model(yaml_path, nc=num_classes)
    ckpt['model'] = {k: v for k, v in ckpt['model'].float().state_dict().items() if k in model.state_dict() and model.state_dict()[k].shape == v.shape}
    model.load_state_dict(ckpt['model'])
    torch.save(model.state_dict(),pt_path,_use_new_zipfile_serialization=False)
 def export_onnx(input_h, input_w, num_classes):
    onnx_batch_size, onnx_img_h, onnx_img_w = 1, input_h, input_w
    yolov5_model = Yolov5Runner(model_path=pt_path, yaml_path=yaml_path, grid20_path=grid20_path, grid40_path=grid40_path, grid80_path=grid80_path, num_classes=num_classes, imgsz_h=onnx_img_h, imgsz_w=onnx_img_w, conf_thres=0.001, iou_thres=0.65, top_k_num=3000, vanish_point=0.0) 
    # Input
    img = torch.zeros((onnx_batch_size, 3, onnx_img_h, onnx_img_w))  
    # img = img.type(torch.cuda.FloatTensor)
    # Load PyTorch model
    model = yolov5_model.yolov5_model
    model.eval()
    model.model[-1].export = True  # set Detect() layer export=True
    y = model(img)  # dry run
    # ONNX export
    try:
        import onnx
        print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
        print('****onnx file****',onnx_export_file)
        torch.onnx.export(model, img, onnx_export_file, verbose=False, opset_version=11, keep_initializers_as_inputs=True, input_names=['images'], output_names=['classes', 'boxes'] if y is None else ['output'])
        # Checks
        onnx_model = onnx.load(onnx_export_file)  # load onnx model
        onnx.checker.check_model(onnx_model)  # check onnx model
        print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
        print('ONNX export success, saved as %s' % onnx_export_file)
    except Exception as e:
        print('ONNX export failure: %s' % e)
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--data', type=str, default='../yolov5/data/pretrained_paths_520.yaml', help='the path to pretrained model paths yaml file')
    args = parser.parse_args()
    with open(args.data) as f:
        data_dict = yaml.load(f, Loader=yaml.FullLoader)  # data dict
    os.environ["CUDA_VISIBLE_DEVICES"] = '0'
    num_classes = data_dict['nc']
    input_w = data_dict['input_w']
    input_h = data_dict['input_h']
    grid_dir = data_dict['grid_dir']
    grid20_path = data_dict['grid20_path']
    grid40_path = data_dict['grid40_path']
    grid80_path = data_dict['grid80_path']
    path = data_dict['path']
    pt_path=data_dict['pt_path']
    yaml_path=data_dict['yaml_path']
    onnx_export_file = data_dict['onnx_export_file']
    save_weight(num_classes)
    export_onnx(input_h, input_w, num_classes)
--- a/kneron/yolov5_preprocess.py
+++ b/kneron/yolov5_preprocess.py
@ -0,0 +1,161 @@
 # coding: utf-8
 import torch
 import cv2
 import numpy as np
 import math
 import time
 import kneron_preprocessing
 kneron_preprocessing.API.set_default_as_520()
 torch.backends.cudnn.deterministic = True
 img_formats = ['.bmp', '.jpg', '.jpeg', '.png', '.tif', '.tiff', '.dng']
 def make_divisible(x, divisor):
    # Returns x evenly divisble by divisor
    return math.ceil(x / divisor) * divisor
 def check_img_size(img_size, s=32):
    # Verify img_size is a multiple of stride s
    new_size = make_divisible(img_size, int(s))  # ceil gs-multiple
    if new_size != img_size:
        print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size))
    return new_size
 def letterbox_ori(img, new_shape=(640, 640), color=(0, 0, 0), auto=True, scaleFill=False, scaleup=True):
    # Resize image to a 32-pixel-multiple rectangle https://github.com/ultralytics/yolov3/issues/232
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)
    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)
    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) # width, height 
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    dw /= 2  # divide padding into 2 sides
    dh /= 2
    if shape[::-1] != new_unpad:  # resize
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
        #img = kneron_preprocessing.API.resize(img,size=new_unpad, keep_ratio = False)
    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    # top, bottom = int(0), int(round(dh + 0.1))
    # left, right = int(0), int(round(dw + 0.1))    
    img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    #img = kneron_preprocessing.API.pad(img, left, right, top, bottom, 0)
    return img, ratio, (dw, dh)
 def letterbox(img, new_shape=(640, 640), color=(0, 0, 0), auto=True, scaleFill=False, scaleup=True):
    # Resize image to a 32-pixel-multiple rectangle https://github.com/ultralytics/yolov3/issues/232
    shape = img.shape[:2]  # current shape [height, width]
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)
    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not scaleup:  # only scale down, do not scale up (for better test mAP)
        r = min(r, 1.0)
    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) # width, height 
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    # dw /= 2  # divide padding into 2 sides
    # dh /= 2
    if shape[::-1] != new_unpad:  # resize
        #img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
        img = kneron_preprocessing.API.resize(img,size=new_unpad, keep_ratio = False)
    # top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
    # left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
    top, bottom = int(0), int(round(dh + 0.1))
    left, right = int(0), int(round(dw + 0.1))    
    #img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
    img = kneron_preprocessing.API.pad(img, left, right, top, bottom, 0)
    return img, ratio, (dw, dh)
 def letterbox_test(img, new_shape=(640, 640), color=(0, 0, 0), auto=True, scaleFill=False, scaleup=True):
    ratio = 1.0, 1.0
    dw, dh = 0, 0
    img = kneron_preprocessing.API.resize(img, size=(480, 256), keep_ratio=False, type='bilinear')
    return img, ratio, (dw, dh)
 def LoadImages(path,img_size):  #_rgb # for inference
    if isinstance(path, str):
        img0 = cv2.imread(path)  # BGR       
    else:
        img0 = path  # BGR
    # Padded resize
    img = letterbox(img0, new_shape=img_size)[0]
    # Convert
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    img = np.ascontiguousarray(img)
    return img, img0
 def LoadImages_yyy(path,img_size): #_yyy # for inference
    if isinstance(path, str):
        img0 = cv2.imread(path)  # BGR       
    else:
        img0 = path  # BGR
    yvu = cv2.cvtColor(img0, cv2.COLOR_BGR2YCrCb)
    y, v, u = cv2.split(yvu)
    img0 = np.stack((y,)*3, axis=-1)
    # Padded resize
    img = letterbox(img0, new_shape=img_size)[0]
    # Convert
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    img = np.ascontiguousarray(img)
    return img, img0
 def LoadImages_yuv420(path,img_size):  #_yuv420 # for inference 
    if isinstance(path, str):
        img0 = cv2.imread(path)  # BGR       
    else:
        img0 = path  # BGR
    img_h, img_w = img0.shape[:2]
    img_h = (img_h // 2) * 2
    img_w = (img_w // 2) * 2
    img = img0[:img_h,:img_w,:]
    yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV_I420)
    img0= cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR_I420) #yuv420
    # Padded resize
    img = letterbox(img0, new_shape=img_size)[0]
    # Convert
    img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
    img = np.ascontiguousarray(img)
    return img, img0
 def Yolov5_preprocess(image_path, device, imgsz_h, imgsz_w) : 
    model_stride_max = 32
    imgsz_h = check_img_size(imgsz_h, s=model_stride_max)  # check img_size
    imgsz_w = check_img_size(imgsz_w, s=model_stride_max)  # check img_size
    img, im0 = LoadImages(image_path, img_size=(imgsz_h,imgsz_w))
    img = kneron_preprocessing.API.norm(img) #path1
    #print('img',img.shape)
    img = torch.from_numpy(img).to(device) #path1,path2
    # img = img.float()  # uint8 to fp16/32 #path2
    # img /= 255.0#256.0 - 0.5 # 0 - 255 to -0.5 - 0.5 #path2
    if img.ndimension() == 3:
        img = img.unsqueeze(0)
    return img, im0
--- a/kneron/yolov5_savingWeight.py
+++ b/kneron/yolov5_savingWeight.py
@ -0,0 +1,42 @@
 import os
 import torch
 import sys
 import argparse
 import yaml
 def save_weight(num_classes): 
    from models.yolo import Model  
    num_classes = num_classes 
    device=torch.device('cpu')
    ckpt = torch.load(path, map_location=device)
    model = Model(yaml_path, nc=num_classes)
    ckpt['model'] = {k: v for k, v in ckpt['model'].float().state_dict().items() if k in model.state_dict() and model.state_dict()[k].shape == v.shape}
    model.load_state_dict(ckpt['model'])
    torch.save(model.state_dict(),pt_path,_use_new_zipfile_serialization=False)
 if __name__ == '__main__':  
    parser = argparse.ArgumentParser()
    parser.add_argument('--data', type=str, default='data/pretrained_paths_520.yaml', help='the path to pretrained model paths yaml file')
    args = parser.parse_args()
    with open(args.data) as f:
        data_dict = yaml.load(f, Loader=yaml.FullLoader)  # data dict
    os.environ["CUDA_VISIBLE_DEVICES"] = '0'
    input_w = data_dict['input_w']
    input_h = data_dict['input_h']
    grid_dir = data_dict['grid_dir']
    grid20_path = data_dict['grid20_path']
    grid40_path = data_dict['grid40_path']
    grid80_path = data_dict['grid80_path']
    path = data_dict['path']
    pt_path=data_dict['pt_path']
    yaml_path=data_dict['yaml_path']
    save_weight(data_dict['nc']) 
--- a/models/init.py
+++ b/models/init.py
--- a/models/common.py
+++ b/models/common.py
@ -0,0 +1,246 @@
 # This file contains modules common to various models
 import math
 import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from PIL import Image, ImageDraw
 from utils.datasets import letterbox
 from utils.general import non_max_suppression, make_divisible, scale_coords, xyxy2xywh
 from utils.plots import color_list
 def autopad(k, p=None):  # kernel, padding
    # Pad to 'same'
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
    return p
 def DWConv(c1, c2, k=1, s=1, act=True):
    # Depthwise convolution
    return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
 class Conv(nn.Module):
    # Standard convolution
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Conv, self).__init__()
        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        #self.act = nn.Hardswish() if act else nn.Identity()
        self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity()
    def forward(self, x):   
        # print('x',x.size())  
        y = self.act(self.bn(self.conv(x)))
        # print('y',y.size())
        return y
    def fuseforward(self, x):
        return self.act(self.conv(x))
 class Bottleneck(nn.Module):
    # Standard bottleneck
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, shortcut, groups, expansion
        super(Bottleneck, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2
    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
 class BottleneckCSP(nn.Module):
    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super(BottleneckCSP, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
        self.cv4 = Conv(2 * c_, c2, 1, 1)
        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
        self.act = nn.LeakyReLU(0.1, inplace=True)
        self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
    def forward(self, x):
        y1 = self.cv3(self.m(self.cv1(x)))
        y2 = self.cv2(x)
        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
 class Focus(nn.Module):
    # Focus wh information into c-space
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Focus, self).__init__()
        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
 class Concat(nn.Module):
    # Concatenate a list of tensors along dimension
    def __init__(self, dimension=1):
        super(Concat, self).__init__()
        self.d = dimension
    def forward(self, x):
        # print('Concat x.size()',x.size())
        y=torch.cat(x, self.d)
        # print('Concat y.size()',y.size())
        return y
 class NMS(nn.Module):
    # Non-Maximum Suppression (NMS) module
    conf = 0.25  # confidence threshold
    iou = 0.45  # IoU threshold
    classes = None  # (optional list) filter by class
    def __init__(self):
        super(NMS, self).__init__()
    def forward(self, x):
        return non_max_suppression(x[0], conf_thres=self.conf, iou_thres=self.iou, classes=self.classes)
 class autoShape(nn.Module):
    # input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
    img_size = 640  # inference size (pixels)
    conf = 0.25  # NMS confidence threshold
    iou = 0.45  # NMS IoU threshold
    classes = None  # (optional list) filter by class
    def __init__(self, model):
        super(autoShape, self).__init__()
        self.model = model.eval()
    def forward(self, imgs, size=640, augment=False, profile=False):
        # supports inference from various sources. For height=720, width=1280, RGB images example inputs are:
        #   opencv:     imgs = cv2.imread('image.jpg')[:,:,::-1]  # HWC BGR to RGB x(720,1280,3)
        #   PIL:        imgs = Image.open('image.jpg')  # HWC x(720,1280,3)
        #   numpy:      imgs = np.zeros((720,1280,3))  # HWC
        #   torch:      imgs = torch.zeros(16,3,720,1280)  # BCHW
        #   multiple:   imgs = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...]  # list of images
        p = next(self.model.parameters())  # for device and type
        if isinstance(imgs, torch.Tensor):  # torch
            return self.model(imgs.to(p.device).type_as(p), augment, profile)  # inference
        # Pre-process
        if not isinstance(imgs, list):
            imgs = [imgs]
        shape0, shape1 = [], []  # image and inference shapes
        batch = range(len(imgs))  # batch size
        for i in batch:
            imgs[i] = np.array(imgs[i])  # to numpy
            imgs[i] = imgs[i][:, :, :3] if imgs[i].ndim == 3 else np.tile(imgs[i][:, :, None], 3)  # enforce 3ch input
            s = imgs[i].shape[:2]  # HWC
            shape0.append(s)  # image shape
            g = (size / max(s))  # gain
            shape1.append([y * g for y in s])
        shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)]  # inference shape
        x = [letterbox(imgs[i], new_shape=shape1, auto=False)[0] for i in batch]  # pad
        x = np.stack(x, 0) if batch[-1] else x[0][None]  # stack
        x = np.ascontiguousarray(x.transpose((0, 3, 1, 2)))  # BHWC to BCHW
        x = torch.from_numpy(x).to(p.device).type_as(p) / 255.  # uint8 to fp16/32
        # Inference
        with torch.no_grad():
            y = self.model(x, augment, profile)[0]  # forward
        y = non_max_suppression(y, conf_thres=self.conf, iou_thres=self.iou, classes=self.classes)  # NMS
        # Post-process
        for i in batch:
            if y[i] is not None:
                y[i][:, :4] = scale_coords(shape1, y[i][:, :4], shape0[i])
        return Detections(imgs, y, self.names)
 class Detections:
    # detections class for YOLOv5 inference results
    def __init__(self, imgs, pred, names=None):
        super(Detections, self).__init__()
        self.imgs = imgs  # list of images as numpy arrays
        self.pred = pred  # list of tensors pred[0] = (xyxy, conf, cls)
        self.names = names  # class names
        self.xyxy = pred  # xyxy pixels
        self.xywh = [xyxy2xywh(x) for x in pred]  # xywh pixels
        gn = [torch.Tensor([*[im.shape[i] for i in [1, 0, 1, 0]], 1., 1.]) for im in imgs]  # normalization gains
        self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)]  # xyxy normalized
        self.xywhn = [x / g for x, g in zip(self.xywh, gn)]  # xywh normalized
    def display(self, pprint=False, show=False, save=False):
        colors = color_list()
        for i, (img, pred) in enumerate(zip(self.imgs, self.pred)):
            str = f'Image {i + 1}/{len(self.pred)}: {img.shape[0]}x{img.shape[1]} '
            if pred is not None:
                for c in pred[:, -1].unique():
                    n = (pred[:, -1] == c).sum()  # detections per class
                    str += f'{n} {self.names[int(c)]}s, '  # add to string
                if show or save:
                    img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img  # from np
                    for *box, conf, cls in pred:  # xyxy, confidence, class
                        # str += '%s %.2f, ' % (names[int(cls)], conf)  # label
                        ImageDraw.Draw(img).rectangle(box, width=4, outline=colors[int(cls) % 10])  # plot
            if save:
                f = f'results{i}.jpg'
                str += f"saved to '{f}'"
                img.save(f)  # save
            if show:
                img.show(f'Image {i}')  # show
            if pprint:
                print(str)
    def print(self):
        self.display(pprint=True)  # print results
    def show(self):
        self.display(show=True)  # show results
    def save(self):
        self.display(save=True)  # save results
 class Flatten(nn.Module):
    # Use after nn.AdaptiveAvgPool2d(1) to remove last 2 dimensions
    @staticmethod
    def forward(x):
        return x.view(x.size(0), -1)
 class Classify(nn.Module):
    # Classification head, i.e. x(b,c1,20,20) to x(b,c2)
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1):  # ch_in, ch_out, kernel, stride, padding, groups
        super(Classify, self).__init__()
        self.aap = nn.AdaptiveAvgPool2d(1)  # to x(b,c1,1,1)
        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)  # to x(b,c2,1,1)
        self.flat = Flatten()
    def forward(self, x):
        z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1)  # cat if list
        return self.flat(self.conv(z))  # flatten to x(b,c2)
 class SPP(nn.Module):
    # Spatial pyramid pooling layer used in YOLOv3-SPP
    def __init__(self, c1, c2, k=(5, 9, 13)):
        super(SPP, self).__init__()
        c_ = c1 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
    def forward(self, x):
        x = self.cv1(x)
        return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
--- a/models/experimental.py
+++ b/models/experimental.py
@ -0,0 +1,152 @@
 # This file contains experimental modules
 import numpy as np
 import torch
 import torch.nn as nn
 from models.common import Conv, DWConv
 from utils.google_utils import attempt_download
 class CrossConv(nn.Module):
    # Cross Convolution Downsample
    def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
        # ch_in, ch_out, kernel, stride, groups, expansion, shortcut
        super(CrossConv, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, (1, k), (1, s))
        self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
        self.add = shortcut and c1 == c2
    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
 class C3(nn.Module):
    # Cross Convolution CSP
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
        super(C3, self).__init__()
        c_ = int(c2 * e)  # hidden channels
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
        self.cv4 = Conv(2 * c_, c2, 1, 1)
        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
        self.act = nn.LeakyReLU(0.1, inplace=True)
        self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
    def forward(self, x):
        y1 = self.cv3(self.m(self.cv1(x)))
        y2 = self.cv2(x)
        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
 class Sum(nn.Module):
    # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
    def __init__(self, n, weight=False):  # n: number of inputs
        super(Sum, self).__init__()
        self.weight = weight  # apply weights boolean
        self.iter = range(n - 1)  # iter object
        if weight:
            self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True)  # layer weights
    def forward(self, x):
        y = x[0]  # no weight
        if self.weight:
            w = torch.sigmoid(self.w) * 2
            for i in self.iter:
                y = y + x[i + 1] * w[i]
        else:
            for i in self.iter:
                y = y + x[i + 1]
        return y
 class GhostConv(nn.Module):
    # Ghost Convolution https://github.com/huawei-noah/ghostnet
    def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
        super(GhostConv, self).__init__()
        c_ = c2 // 2  # hidden channels
        self.cv1 = Conv(c1, c_, k, s, None, g, act)
        self.cv2 = Conv(c_, c_, 5, 1, None, c_, act)
    def forward(self, x):
        y = self.cv1(x)
        return torch.cat([y, self.cv2(y)], 1)
 class GhostBottleneck(nn.Module):
    # Ghost Bottleneck https://github.com/huawei-noah/ghostnet
    def __init__(self, c1, c2, k, s):
        super(GhostBottleneck, self).__init__()
        c_ = c2 // 2
        self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1),  # pw
                                  DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(),  # dw
                                  GhostConv(c_, c2, 1, 1, act=False))  # pw-linear
        self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
                                      Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
    def forward(self, x):
        return self.conv(x) + self.shortcut(x)
 class MixConv2d(nn.Module):
    # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595
    def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
        super(MixConv2d, self).__init__()
        groups = len(k)
        if equal_ch:  # equal c_ per group
            i = torch.linspace(0, groups - 1E-6, c2).floor()  # c2 indices
            c_ = [(i == g).sum() for g in range(groups)]  # intermediate channels
        else:  # equal weight.numel() per group
            b = [c2] + [0] * groups
            a = np.eye(groups + 1, groups, k=-1)
            a -= np.roll(a, 1, axis=1)
            a *= np.array(k) ** 2
            a[0] = 1
            c_ = np.linalg.lstsq(a, b, rcond=None)[0].round()  # solve for equal weight indices, ax = b
        self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
        self.bn = nn.BatchNorm2d(c2)
        self.act = nn.LeakyReLU(0.1, inplace=True)
    def forward(self, x):
        return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
 class Ensemble(nn.ModuleList):
    # Ensemble of models
    def __init__(self):
        super(Ensemble, self).__init__()
    def forward(self, x, augment=False):
        y = []
        for module in self:
            y.append(module(x, augment)[0])
        # y = torch.stack(y).max(0)[0]  # max ensemble
        # y = torch.cat(y, 1)  # nms ensemble
        y = torch.stack(y).mean(0)  # mean ensemble
        return y, None  # inference, train output
 def attempt_load(weights, map_location=None):
    # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
    model = Ensemble()
    for w in weights if isinstance(weights, list) else [weights]:
        ckpt = torch.load(w, map_location=map_location)
        model.append( ckpt['model'].float().fuse().eval() )  # load FP32 model
    # Compatibility updates
    for m in model.modules():
        if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
            m.inplace = True  # pytorch 1.7.0 compatibility
        elif type(m) is Conv:
            m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
    if len(model) == 1:
        return model[-1]  # return model
    else:
        print('Ensemble created with %s\n' % weights)
        for k in ['names', 'stride']:
            setattr(model, k, getattr(model[-1], k))
        return model  # return ensemble
--- a/models/export.py
+++ b/models/export.py
@ -0,0 +1,94 @@
 """Exports a YOLOv5 *.pt model to ONNX and TorchScript formats
 Usage:
    $ export PYTHONPATH="$PWD" && python models/export.py --weights ./weights/yolov5s.pt --img 640 --batch 1
 """
 import argparse
 import sys
 import time
 sys.path.append('./')  # to run '$ python *.py' files in subdirectories
 import torch
 import torch.nn as nn
 import models
 from models.experimental import attempt_load
 from utils.activations import Hardswish
 from utils.general import set_logging, check_img_size
 if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path')  # from yolov5/models/
    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size')  # height, width
    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
    opt = parser.parse_args()
    opt.img_size *= 2 if len(opt.img_size) == 1 else 1  # expand
    print(opt)
    set_logging()
    t = time.time()
    # Load PyTorch model
    model = attempt_load(opt.weights, map_location=torch.device('cpu'))  # load FP32 model
    labels = model.names
    # Checks
    gs = int(max(model.stride))  # grid size (max stride)
    opt.img_size = [check_img_size(x, gs) for x in opt.img_size]  # verify img_size are gs-multiples
    # Input
    img = torch.zeros(opt.batch_size, 3, *opt.img_size)  # image size(1,3,320,192) iDetection
    # Update model
    for k, m in model.named_modules():
        m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
        if isinstance(m, models.common.Conv) and isinstance(m.act, nn.Hardswish):
            m.act = Hardswish()  # assign activation
        # if isinstance(m, models.yolo.Detect):
        #     m.forward = m.forward_export  # assign forward (optional)
    model.model[-1].export = True  # set Detect() layer export=True
    y = model(img)  # dry run
    # TorchScript export
    try:
        print('\nStarting TorchScript export with torch %s...' % torch.__version__)
        f = opt.weights.replace('.pt', '.torchscript.pt')  # filename
        ts = torch.jit.trace(model, img)
        ts.save(f)
        print('TorchScript export success, saved as %s' % f)
    except Exception as e:
        print('TorchScript export failure: %s' % e)
    # ONNX export
    try:
        import onnx
        print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
        f = opt.weights.replace('.pt', '.onnx')  # filename
        torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'],
                          output_names=['classes', 'boxes'] if y is None else ['output'])
        # Checks
        onnx_model = onnx.load(f)  # load onnx model
        onnx.checker.check_model(onnx_model)  # check onnx model
        # print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
        print('ONNX export success, saved as %s' % f)
    except Exception as e:
        print('ONNX export failure: %s' % e)
    # CoreML export
    try:
        import coremltools as ct
        print('\nStarting CoreML export with coremltools %s...' % ct.__version__)
        # convert model from torchscript and apply pixel scaling as per detect.py
        model = ct.convert(ts, inputs=[ct.ImageType(name='image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])])
        f = opt.weights.replace('.pt', '.mlmodel')  # filename
        model.save(f)
        print('CoreML export success, saved as %s' % f)
    except Exception as e:
        print('CoreML export failure: %s' % e)
    # Finish
    print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t))
--- a/Show More
+++ b/Show More
		`@ -0,0 +1,2 @@`
							`0 0.771875 0.50859375 0.03203125 0.0703125`
							`0 0.71484375 0.46171875 0.05 0.0515625`
		`@ -0,0 +1,2 @@`
							`0 0.0203125 0.92109375 0.040625 0.1578125`
							`0 0.1046875 0.73046875 0.0609375 0.07421875`
		`@ -0,0 +1,2 @@`
							`1 0.590625 0.2671875 0.0546875 0.03125`
							`1 0.6546875 0.28359375 0.06875 0.03046875`
		`@ -0,0 +1,2 @@`
							`1 0.67265625 0.31640625 0.1453125 0.09765625`
							`1 0.81484375 0.38359375 0.12421875 0.11015625`
		`@ -0,0 +1,2 @@`
							`from . import ColorConversion, Padding, Resize, Crop, Normalize, Rotate`