Upload Previous files
This commit is contained in:
parent
5f8e1d37ac
commit
095addb553
BIN
Jay Chou.npy
Normal file
BIN
Jay Chou.npy
Normal file
Binary file not shown.
BIN
R34_G369K.nef
Normal file
BIN
R34_G369K.nef
Normal file
Binary file not shown.
BIN
alignment_visualization.jpg
Normal file
BIN
alignment_visualization.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 48 KiB |
321
face_recognition.py
Normal file
321
face_recognition.py
Normal file
@ -0,0 +1,321 @@
|
|||||||
|
import os
|
||||||
|
import sys
|
||||||
|
import argparse
|
||||||
|
import numpy as np
|
||||||
|
import json
|
||||||
|
import glob
|
||||||
|
import time
|
||||||
|
import cv2
|
||||||
|
from mtcnn.mtcnn import MTCNN
|
||||||
|
import kp
|
||||||
|
|
||||||
|
# 引入之前的功能
|
||||||
|
from nef_test import (
|
||||||
|
load_image_safe, landmarks, affine_matrix, extract_vector_data,
|
||||||
|
load_vector, cosine_similarity, SCPU_FW_PATH, NCPU_FW_PATH, visualize_alignment
|
||||||
|
)
|
||||||
|
|
||||||
|
# 預設參數
|
||||||
|
MODEL_FILE_PATH = 'R34_G369K.nef'
|
||||||
|
VECTOR_DATABASE_DIR = 'face_vectors'
|
||||||
|
SIMILARITY_THRESHOLD = 0.5 # 相似度閾值，可根據需要調整
|
||||||
|
|
||||||
|
def get_face_vector(device_group, model_nef_descriptor, image_path):
|
||||||
|
"""從圖像中擷取人臉向量"""
|
||||||
|
|
||||||
|
# 創建MTCNN檢測器
|
||||||
|
detector = MTCNN(device="CPU:0")
|
||||||
|
|
||||||
|
# 載入圖像
|
||||||
|
try:
|
||||||
|
img_rgb, img_bgr = load_image_safe(image_path)
|
||||||
|
print(f" - 已載入圖像: {image_path}")
|
||||||
|
except Exception as e:
|
||||||
|
print(f"錯誤: {str(e)}")
|
||||||
|
return None
|
||||||
|
|
||||||
|
# 獲取人臉特徵點並計算仿射變換矩陣
|
||||||
|
try:
|
||||||
|
lmks = landmarks(detector, img_rgb)
|
||||||
|
mat, size = affine_matrix(lmks)
|
||||||
|
print(" - 已檢測到人臉特徵點")
|
||||||
|
except Exception as e:
|
||||||
|
print(f"錯誤: {str(e)}")
|
||||||
|
return None
|
||||||
|
|
||||||
|
# 應用仿射變換
|
||||||
|
aligned_img = cv2.warpAffine(img_bgr, mat, size)
|
||||||
|
|
||||||
|
# visualize_alignment(img_bgr, lmks, aligned_img)
|
||||||
|
|
||||||
|
# 轉換為BGR565格式並調整大小
|
||||||
|
aligned_img_bgr565 = cv2.cvtColor(aligned_img, cv2.COLOR_BGR2BGR565)
|
||||||
|
img_bgr565 = cv2.resize(aligned_img_bgr565, (112, 112), interpolation=cv2.INTER_LINEAR)
|
||||||
|
print(" - 已對齊圖像並格式化")
|
||||||
|
|
||||||
|
# 準備通用圖像推理輸入描述符
|
||||||
|
generic_inference_input_descriptor = kp.GenericImageInferenceDescriptor(
|
||||||
|
model_id=model_nef_descriptor.models[0].id,
|
||||||
|
inference_number=0,
|
||||||
|
input_node_image_list=[
|
||||||
|
kp.GenericInputNodeImage(
|
||||||
|
image=img_bgr565,
|
||||||
|
image_format=kp.ImageFormat.KP_IMAGE_FORMAT_RGB565,
|
||||||
|
resize_mode=kp.ResizeMode.KP_RESIZE_ENABLE,
|
||||||
|
padding_mode=kp.PaddingMode.KP_PADDING_CORNER,
|
||||||
|
normalize_mode=kp.NormalizeMode.KP_NORMALIZE_KNERON
|
||||||
|
)
|
||||||
|
]
|
||||||
|
)
|
||||||
|
|
||||||
|
# 開始推理工作
|
||||||
|
try:
|
||||||
|
kp.inference.generic_image_inference_send(
|
||||||
|
device_group=device_group,
|
||||||
|
generic_inference_input_descriptor=generic_inference_input_descriptor
|
||||||
|
)
|
||||||
|
|
||||||
|
generic_raw_result = kp.inference.generic_image_inference_receive(device_group=device_group)
|
||||||
|
print(" - 推理成功完成")
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print(f' - 錯誤: 推理失敗，錯誤 = {exception}')
|
||||||
|
return None
|
||||||
|
|
||||||
|
# 獲取推理節點輸出
|
||||||
|
inf_node_output_list = []
|
||||||
|
for node_idx in range(generic_raw_result.header.num_output_node):
|
||||||
|
inference_float_node_output = kp.inference.generic_inference_retrieve_float_node(
|
||||||
|
node_idx=node_idx,
|
||||||
|
generic_raw_result=generic_raw_result,
|
||||||
|
channels_ordering=kp.ChannelOrdering.KP_CHANNEL_ORDERING_CHW
|
||||||
|
)
|
||||||
|
inf_node_output_list.append(inference_float_node_output)
|
||||||
|
|
||||||
|
# 獲取人臉向量
|
||||||
|
if len(inf_node_output_list) > 0:
|
||||||
|
face_vector = inf_node_output_list[0]
|
||||||
|
# 轉換向量為標準NumPy數組
|
||||||
|
face_vector_np = extract_vector_data(face_vector)
|
||||||
|
|
||||||
|
# Ensure the vector is flattened to 1D
|
||||||
|
face_vector_np = face_vector_np.flatten()
|
||||||
|
|
||||||
|
# Print shape for debugging
|
||||||
|
print(f" - Face vector shape: {face_vector_np.shape}")
|
||||||
|
|
||||||
|
return face_vector_np
|
||||||
|
else:
|
||||||
|
print("錯誤: 推理結果中沒有找到輸出節點")
|
||||||
|
return None
|
||||||
|
|
||||||
|
def load_face_database(database_dir):
|
||||||
|
"""載入人臉資料庫"""
|
||||||
|
face_db = []
|
||||||
|
|
||||||
|
# 確保資料庫目錄存在
|
||||||
|
if not os.path.exists(database_dir):
|
||||||
|
print(f"警告: 人臉資料庫目錄 '{database_dir}' 不存在，將創建該目錄")
|
||||||
|
os.makedirs(database_dir)
|
||||||
|
return face_db
|
||||||
|
|
||||||
|
# 查找所有支持的向量文件
|
||||||
|
vector_files = []
|
||||||
|
vector_files.extend(glob.glob(os.path.join(database_dir, "*.npy")))
|
||||||
|
vector_files.extend(glob.glob(os.path.join(database_dir, "*.json")))
|
||||||
|
vector_files.extend(glob.glob(os.path.join(database_dir, "*.pkl")))
|
||||||
|
|
||||||
|
# 載入每個向量文件
|
||||||
|
for vector_file in vector_files:
|
||||||
|
try:
|
||||||
|
# 根據文件擴展名自動確定格式
|
||||||
|
file_ext = os.path.splitext(vector_file)[1].lower()
|
||||||
|
if file_ext == '.npy':
|
||||||
|
vector = load_vector(vector_file, format='numpy')
|
||||||
|
# 嘗試從文件名中提取標識信息
|
||||||
|
name = os.path.basename(vector_file).replace('.npy', '')
|
||||||
|
metadata = {'name': name}
|
||||||
|
face_db.append({
|
||||||
|
'vector': vector,
|
||||||
|
'metadata': metadata,
|
||||||
|
'file_path': vector_file
|
||||||
|
})
|
||||||
|
elif file_ext == '.json':
|
||||||
|
vector, metadata = load_vector(vector_file, format='json')
|
||||||
|
face_db.append({
|
||||||
|
'vector': vector,
|
||||||
|
'metadata': metadata or {'name': os.path.basename(vector_file).replace('.json', '')},
|
||||||
|
'file_path': vector_file
|
||||||
|
})
|
||||||
|
elif file_ext == '.pkl':
|
||||||
|
vector = load_vector(vector_file, format='pickle')
|
||||||
|
# 嘗試從文件名中提取標識信息
|
||||||
|
name = os.path.basename(vector_file).replace('.pkl', '')
|
||||||
|
metadata = {'name': name}
|
||||||
|
face_db.append({
|
||||||
|
'vector': vector,
|
||||||
|
'metadata': metadata,
|
||||||
|
'file_path': vector_file
|
||||||
|
})
|
||||||
|
|
||||||
|
print(f"已載入人臉向量: {vector_file}")
|
||||||
|
except Exception as e:
|
||||||
|
print(f"警告: 無法載入向量文件 '{vector_file}': {str(e)}")
|
||||||
|
|
||||||
|
print(f"成功載入 {len(face_db)} 個人臉向量")
|
||||||
|
return face_db
|
||||||
|
|
||||||
|
def recognize_face(new_face_vector, face_database, threshold=SIMILARITY_THRESHOLD):
|
||||||
|
"""識別人臉，將新的人臉向量與資料庫中的向量進行比較"""
|
||||||
|
if not face_database:
|
||||||
|
return None, 0.0
|
||||||
|
|
||||||
|
max_similarity = 0.0
|
||||||
|
best_match = None
|
||||||
|
|
||||||
|
for face_entry in face_database:
|
||||||
|
stored_vector = face_entry['vector']
|
||||||
|
similarity = cosine_similarity(new_face_vector, stored_vector)
|
||||||
|
|
||||||
|
if similarity > max_similarity:
|
||||||
|
max_similarity = similarity
|
||||||
|
best_match = face_entry
|
||||||
|
|
||||||
|
# 如果最高相似度超過閾值，則返回匹配結果
|
||||||
|
if max_similarity >= threshold:
|
||||||
|
return best_match, max_similarity
|
||||||
|
else:
|
||||||
|
return None, max_similarity
|
||||||
|
|
||||||
|
def add_face_to_database(face_vector, database_dir, name=None, image_path=None, format='json'):
|
||||||
|
"""添加新的人臉向量到資料庫"""
|
||||||
|
if not os.path.exists(database_dir):
|
||||||
|
os.makedirs(database_dir)
|
||||||
|
|
||||||
|
# 創建唯一的文件名
|
||||||
|
timestamp = time.strftime("%Y%m%d_%H%M%S")
|
||||||
|
if name:
|
||||||
|
file_name = f"{name}_{timestamp}"
|
||||||
|
else:
|
||||||
|
file_name = f"unknown_{timestamp}"
|
||||||
|
|
||||||
|
# 創建元數據
|
||||||
|
metadata = {
|
||||||
|
'name': name or 'unknown',
|
||||||
|
'timestamp': time.strftime("%Y-%m-%d %H:%M:%S"),
|
||||||
|
'image_path': image_path
|
||||||
|
}
|
||||||
|
|
||||||
|
# 根據格式保存向量
|
||||||
|
if format == 'numpy':
|
||||||
|
file_path = os.path.join(database_dir, f"{file_name}.npy")
|
||||||
|
np.save(file_path, face_vector)
|
||||||
|
elif format == 'pickle':
|
||||||
|
import pickle
|
||||||
|
file_path = os.path.join(database_dir, f"{file_name}.pkl")
|
||||||
|
with open(file_path, 'wb') as f:
|
||||||
|
pickle.dump(face_vector, f)
|
||||||
|
else: # 預設使用JSON格式
|
||||||
|
file_path = os.path.join(database_dir, f"{file_name}.json")
|
||||||
|
data = {
|
||||||
|
'vector': face_vector.tolist(),
|
||||||
|
'metadata': metadata
|
||||||
|
}
|
||||||
|
with open(file_path, 'w') as f:
|
||||||
|
json.dump(data, f)
|
||||||
|
|
||||||
|
print(f"已將新的人臉向量保存到: {file_path}")
|
||||||
|
return file_path
|
||||||
|
|
||||||
|
def main():
|
||||||
|
parser = argparse.ArgumentParser(description='人臉識別和比對系統')
|
||||||
|
parser.add_argument('-p', '--port_id', help='使用指定的端口ID連接設備', default=28, type=int)
|
||||||
|
parser.add_argument('-m', '--model', help=f'模型文件路徑 (.nef) (預設: {MODEL_FILE_PATH})',
|
||||||
|
default=MODEL_FILE_PATH, type=str)
|
||||||
|
parser.add_argument('-i', '--img', help='待識別的圖像文件路徑', required=True, type=str)
|
||||||
|
parser.add_argument('-d', '--database', help=f'人臉向量資料庫目錄 (預設: {VECTOR_DATABASE_DIR})',
|
||||||
|
default=VECTOR_DATABASE_DIR, type=str)
|
||||||
|
parser.add_argument('-t', '--threshold', help=f'相似度閾值 (預設: {SIMILARITY_THRESHOLD})',
|
||||||
|
default=SIMILARITY_THRESHOLD, type=float)
|
||||||
|
parser.add_argument('-a', '--add', help='將新人臉添加到資料庫 (如果未識別)', action='store_true')
|
||||||
|
parser.add_argument('-n', '--name', help='新人臉的名稱 (與 --add 一起使用)', default=None, type=str)
|
||||||
|
parser.add_argument('-f', '--format', help='向量儲存格式: numpy, pickle, 或 json (預設: json)',
|
||||||
|
default='json', choices=['numpy', 'pickle', 'json'], type=str)
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
# 連接設備
|
||||||
|
try:
|
||||||
|
print('[連接設備]')
|
||||||
|
device_group = kp.core.connect_devices(usb_port_ids=[args.port_id])
|
||||||
|
print(' - 成功')
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print(f'錯誤: 連接設備失敗, 端口ID = \'{args.port_id}\'，錯誤信息: [{str(exception)}]')
|
||||||
|
return
|
||||||
|
|
||||||
|
# 設置USB通信超時
|
||||||
|
print('[設置設備超時]')
|
||||||
|
kp.core.set_timeout(device_group=device_group, milliseconds=5000)
|
||||||
|
print(' - 成功')
|
||||||
|
|
||||||
|
# 上傳固件到設備
|
||||||
|
try:
|
||||||
|
print('[上傳固件]')
|
||||||
|
kp.core.load_firmware_from_file(device_group=device_group,
|
||||||
|
scpu_fw_path=SCPU_FW_PATH,
|
||||||
|
ncpu_fw_path=NCPU_FW_PATH)
|
||||||
|
print(' - 成功')
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print(f'錯誤: 上傳固件失敗，錯誤 = \'{str(exception)}\'')
|
||||||
|
return
|
||||||
|
|
||||||
|
# 上傳模型到設備
|
||||||
|
try:
|
||||||
|
print('[上傳模型]')
|
||||||
|
model_nef_descriptor = kp.core.load_model_from_file(device_group=device_group,
|
||||||
|
file_path=args.model)
|
||||||
|
print(' - 成功')
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print(f'錯誤: 上傳模型失敗，錯誤 = \'{str(exception)}\'')
|
||||||
|
return
|
||||||
|
|
||||||
|
# 載入人臉資料庫
|
||||||
|
print('[載入人臉資料庫]')
|
||||||
|
face_database = load_face_database(args.database)
|
||||||
|
|
||||||
|
# 從圖像中提取人臉向量
|
||||||
|
print('[處理輸入圖像]')
|
||||||
|
new_face_vector = get_face_vector(device_group, model_nef_descriptor, args.img)
|
||||||
|
|
||||||
|
if new_face_vector is not None:
|
||||||
|
# 識別人臉
|
||||||
|
print('[識別人臉]')
|
||||||
|
match, similarity = recognize_face(new_face_vector, face_database, args.threshold)
|
||||||
|
|
||||||
|
if match:
|
||||||
|
metadata = match['metadata']
|
||||||
|
name = metadata.get('name', '未知')
|
||||||
|
print(f"[結果] 識別為: {name}")
|
||||||
|
print(f" - 相似度: {similarity:.4f}")
|
||||||
|
print(f" - 來源文件: {match['file_path']}")
|
||||||
|
if 'timestamp' in metadata:
|
||||||
|
print(f" - 記錄時間: {metadata['timestamp']}")
|
||||||
|
else:
|
||||||
|
print(f"[結果] 未能識別人臉 (最高相似度: {similarity:.4f}, 閾值: {args.threshold})")
|
||||||
|
|
||||||
|
# 如果指定了--add參數，則將新的人臉添加到資料庫
|
||||||
|
if args.add:
|
||||||
|
print('[添加新人臉到資料庫]')
|
||||||
|
file_path = add_face_to_database(
|
||||||
|
new_face_vector,
|
||||||
|
args.database,
|
||||||
|
name=args.name,
|
||||||
|
image_path=args.img,
|
||||||
|
format=args.format
|
||||||
|
)
|
||||||
|
print(f"[添加完成] 已添加新人臉: {args.name or '未命名'}")
|
||||||
|
|
||||||
|
# 清理
|
||||||
|
kp.core.disconnect_devices(device_group=device_group)
|
||||||
|
print("[清理] 設備已斷開連接")
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
main()
|
||||||
BIN
face_vectors/Jay.npy
Normal file
BIN
face_vectors/Jay.npy
Normal file
Binary file not shown.
BIN
face_vectors/Will Smith.npy
Normal file
BIN
face_vectors/Will Smith.npy
Normal file
Binary file not shown.
BIN
face_vectors/Zhiyu.npy
Normal file
BIN
face_vectors/Zhiyu.npy
Normal file
Binary file not shown.
BIN
face_vectors/山下智久.npy
Normal file
BIN
face_vectors/山下智久.npy
Normal file
Binary file not shown.
8
fw_info.txt
Normal file
8
fw_info.txt
Normal file
@ -0,0 +1,8 @@
|
|||||||
|
Total [1] models:
|
||||||
|
[input]
|
||||||
|
id: [32769], version: [0x8b28]
|
||||||
|
size: input [0xc400], output [0x2000], buf [0x134000], cmd [0x1c14], weight [0x1839c00], fw_code [0xa0]
|
||||||
|
addr: input [0x60000000], output [0x6000c400], buf [0x6000e400], cmd [0x60142400], weight [0x60144020], fw_code [0x6197dc20]
|
||||||
|
dram_addr_end [0x6197dcc0], total bin size: [0x183b8c0]
|
||||||
|
checksum: all_models.bin [0xf7c5b2d6]
|
||||||
|
0 [0xf7c5b2d6],
|
||||||
BIN
input.kdp520.decomposed.onnx
Normal file
BIN
input.kdp520.decomposed.onnx
Normal file
Binary file not shown.
BIN
input.kdp520.scaled.bie
Normal file
BIN
input.kdp520.scaled.bie
Normal file
Binary file not shown.
BIN
input.onnx
Normal file
BIN
input.onnx
Normal file
Binary file not shown.
727
model_fx_report.html
Normal file
727
model_fx_report.html
Normal file
@ -0,0 +1,727 @@
|
|||||||
|
<h1>Summary</h1><br><hr><table border="2" class="dataframe">
|
||||||
|
<thead>
|
||||||
|
<tr style="text-align: right;">
|
||||||
|
<th></th>
|
||||||
|
<th>info</th>
|
||||||
|
</tr>
|
||||||
|
</thead>
|
||||||
|
<tbody>
|
||||||
|
<tr>
|
||||||
|
<th>docker_version</th>
|
||||||
|
<td>kneron/toolchain:v0.29.0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>comments</th>
|
||||||
|
<td></td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/input bitwidth</th>
|
||||||
|
<td>int8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/output bitwidth</th>
|
||||||
|
<td>int8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/cpu bitwidth</th>
|
||||||
|
<td>int8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/datapath bitwidth</th>
|
||||||
|
<td>int8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/weight bitwidth</th>
|
||||||
|
<td>int8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/ip_eval/fps</th>
|
||||||
|
<td>13.5583</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/ip_eval/ITC(ms)</th>
|
||||||
|
<td>73.7554 ms</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/ip_eval/RDMA bandwidth GB/s</th>
|
||||||
|
<td>0.8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/ip_eval/WDMA bandwidth GB/s</th>
|
||||||
|
<td>0.8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/ip_eval/GETW bandwidth GB/s</th>
|
||||||
|
<td>0.8</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/ip_eval/cpu_node</th>
|
||||||
|
<td>N/A</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/bie</th>
|
||||||
|
<td>input.kdp520.scaled.bie</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/onnx</th>
|
||||||
|
<td>input.kdp520.decomposed.onnx</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>kdp520/nef</th>
|
||||||
|
<td>models_520.nef</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>gen fx model report</th>
|
||||||
|
<td>model_fx_report.html</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>gen fx model json</th>
|
||||||
|
<td>model_fx_report.json</td>
|
||||||
|
</tr>
|
||||||
|
</tbody>
|
||||||
|
</table><br><hr><h2>kdp520</h2><br><hr><table border="1" class="dataframe">
|
||||||
|
<thead>
|
||||||
|
<tr style="text-align: right;">
|
||||||
|
<th></th>
|
||||||
|
<th>node</th>
|
||||||
|
<th>node origin</th>
|
||||||
|
<th>type</th>
|
||||||
|
<th>bw in</th>
|
||||||
|
<th>bw out</th>
|
||||||
|
<th>bw weight</th>
|
||||||
|
<th>node backend</th>
|
||||||
|
</tr>
|
||||||
|
</thead>
|
||||||
|
<tbody>
|
||||||
|
<tr>
|
||||||
|
<th>0</th>
|
||||||
|
<td>input</td>
|
||||||
|
<td>None</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>N/A</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>N/A</td>
|
||||||
|
<td>None</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>1</th>
|
||||||
|
<td>/conv1/Conv</td>
|
||||||
|
<td>/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/conv1/Conv_/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>2</th>
|
||||||
|
<td>/prelu/PRelu</td>
|
||||||
|
<td>/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>3</th>
|
||||||
|
<td>/layer1/layer1.0/bn1/BatchNormalization_KNOPT_bn_sep_1</td>
|
||||||
|
<td>/layer1/layer1.0/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer1/layer1.0/bn1/BatchNormalization_KNOPT_bn_sep_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>4</th>
|
||||||
|
<td>/layer1/layer1.0/bn1/BatchNormalization_KNOPT_bn_sep_2</td>
|
||||||
|
<td>/layer1/layer1.0/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer1/layer1.0/bn1/BatchNormalization_KNOPT_bn_sep_2</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>5</th>
|
||||||
|
<td>/layer1/layer1.0/conv1/Conv</td>
|
||||||
|
<td>/layer1/layer1.0/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer1/layer1.0/conv1/Conv_/layer1/layer1.0/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>6</th>
|
||||||
|
<td>/layer1/layer1.0/prelu/PRelu</td>
|
||||||
|
<td>/layer1/layer1.0/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>7</th>
|
||||||
|
<td>/layer1/layer1.0/conv2/Conv_KNOPT_/layer1/layer1.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer1/layer1.0/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer1/layer1.0/conv2/Conv_KNOPT_/layer1/layer1.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>8</th>
|
||||||
|
<td>/layer1/layer1.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>/layer1/layer1.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer1/layer1.0/downsample/downsample.0/Conv_/layer1/layer1.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>9</th>
|
||||||
|
<td>/layer1/layer1.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer1/layer1.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>10</th>
|
||||||
|
<td>/layer1/layer1.0/Add</td>
|
||||||
|
<td>/layer1/layer1.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer1/layer1.0/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>11</th>
|
||||||
|
<td>/layer1/layer1.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
<td>/layer1/layer1.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer1/layer1.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>12</th>
|
||||||
|
<td>/layer1/layer1.1/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer1/layer1.1/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer1/layer1.1/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>13</th>
|
||||||
|
<td>/layer1/layer1.1/conv1/Conv</td>
|
||||||
|
<td>/layer1/layer1.1/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer1/layer1.1/conv1/Conv_/layer1/layer1.1/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>14</th>
|
||||||
|
<td>/layer1/layer1.1/prelu/PRelu</td>
|
||||||
|
<td>/layer1/layer1.1/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>15</th>
|
||||||
|
<td>/layer1/layer1.1/conv2/Conv_KNOPT_/layer1/layer1.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer1/layer1.1/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer1/layer1.1/conv2/Conv_KNOPT_/layer1/layer1.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>16</th>
|
||||||
|
<td>/layer1/layer1.1/Add</td>
|
||||||
|
<td>/layer1/layer1.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer1/layer1.1/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>17</th>
|
||||||
|
<td>/layer1/layer1.1/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer2/layer2.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer1/layer1.1/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>18</th>
|
||||||
|
<td>/layer2/layer2.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>/layer2/layer2.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer2/layer2.0/downsample/downsample.0/Conv_/layer2/layer2.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>19</th>
|
||||||
|
<td>/layer2/layer2.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer2/layer2.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>20</th>
|
||||||
|
<td>/layer2/layer2.0/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer2/layer2.0/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer2/layer2.0/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>21</th>
|
||||||
|
<td>/layer2/layer2.0/conv1/Conv</td>
|
||||||
|
<td>/layer2/layer2.0/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer2/layer2.0/conv1/Conv_/layer2/layer2.0/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>22</th>
|
||||||
|
<td>/layer2/layer2.0/prelu/PRelu</td>
|
||||||
|
<td>/layer2/layer2.0/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>23</th>
|
||||||
|
<td>/layer2/layer2.0/conv2/Conv_KNOPT_/layer2/layer2.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer2/layer2.0/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer2/layer2.0/conv2/Conv_KNOPT_/layer2/layer2.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>24</th>
|
||||||
|
<td>/layer2/layer2.0/Add</td>
|
||||||
|
<td>/layer2/layer2.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer2/layer2.0/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>25</th>
|
||||||
|
<td>/layer2/layer2.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
<td>/layer2/layer2.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer2/layer2.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>26</th>
|
||||||
|
<td>/layer2/layer2.1/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer2/layer2.1/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer2/layer2.1/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>27</th>
|
||||||
|
<td>/layer2/layer2.1/conv1/Conv</td>
|
||||||
|
<td>/layer2/layer2.1/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer2/layer2.1/conv1/Conv_/layer2/layer2.1/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>28</th>
|
||||||
|
<td>/layer2/layer2.1/prelu/PRelu</td>
|
||||||
|
<td>/layer2/layer2.1/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>29</th>
|
||||||
|
<td>/layer2/layer2.1/conv2/Conv_KNOPT_/layer2/layer2.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer2/layer2.1/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer2/layer2.1/conv2/Conv_KNOPT_/layer2/layer2.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>30</th>
|
||||||
|
<td>/layer2/layer2.1/Add</td>
|
||||||
|
<td>/layer2/layer2.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer2/layer2.1/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>31</th>
|
||||||
|
<td>/layer2/layer2.1/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer3/layer3.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer2/layer2.1/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>32</th>
|
||||||
|
<td>/layer3/layer3.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>/layer3/layer3.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer3/layer3.0/downsample/downsample.0/Conv_/layer3/layer3.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>33</th>
|
||||||
|
<td>/layer3/layer3.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer3/layer3.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>34</th>
|
||||||
|
<td>/layer3/layer3.0/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer3/layer3.0/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer3/layer3.0/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>35</th>
|
||||||
|
<td>/layer3/layer3.0/conv1/Conv</td>
|
||||||
|
<td>/layer3/layer3.0/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer3/layer3.0/conv1/Conv_/layer3/layer3.0/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>36</th>
|
||||||
|
<td>/layer3/layer3.0/prelu/PRelu</td>
|
||||||
|
<td>/layer3/layer3.0/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>37</th>
|
||||||
|
<td>/layer3/layer3.0/conv2/Conv_KNOPT_/layer3/layer3.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer3/layer3.0/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer3/layer3.0/conv2/Conv_KNOPT_/layer3/layer3.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>38</th>
|
||||||
|
<td>/layer3/layer3.0/Add</td>
|
||||||
|
<td>/layer3/layer3.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer3/layer3.0/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>39</th>
|
||||||
|
<td>/layer3/layer3.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
<td>/layer3/layer3.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer3/layer3.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>40</th>
|
||||||
|
<td>/layer3/layer3.1/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer3/layer3.1/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer3/layer3.1/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>41</th>
|
||||||
|
<td>/layer3/layer3.1/conv1/Conv</td>
|
||||||
|
<td>/layer3/layer3.1/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer3/layer3.1/conv1/Conv_/layer3/layer3.1/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>42</th>
|
||||||
|
<td>/layer3/layer3.1/prelu/PRelu</td>
|
||||||
|
<td>/layer3/layer3.1/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>43</th>
|
||||||
|
<td>/layer3/layer3.1/conv2/Conv_KNOPT_/layer3/layer3.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer3/layer3.1/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer3/layer3.1/conv2/Conv_KNOPT_/layer3/layer3.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>44</th>
|
||||||
|
<td>/layer3/layer3.1/Add</td>
|
||||||
|
<td>/layer3/layer3.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer3/layer3.1/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>45</th>
|
||||||
|
<td>/layer3/layer3.1/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer4/layer4.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer3/layer3.1/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>46</th>
|
||||||
|
<td>/layer4/layer4.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>/layer4/layer4.0/downsample/downsample.0/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer4/layer4.0/downsample/downsample.0/Conv_/layer4/layer4.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>47</th>
|
||||||
|
<td>/layer4/layer4.0/Add_KNOPT_dummy_bn_1</td>
|
||||||
|
<td>/layer4/layer4.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>48</th>
|
||||||
|
<td>/layer4/layer4.0/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer4/layer4.0/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer4/layer4.0/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>49</th>
|
||||||
|
<td>/layer4/layer4.0/conv1/Conv</td>
|
||||||
|
<td>/layer4/layer4.0/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer4/layer4.0/conv1/Conv_/layer4/layer4.0/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>50</th>
|
||||||
|
<td>/layer4/layer4.0/prelu/PRelu</td>
|
||||||
|
<td>/layer4/layer4.0/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>51</th>
|
||||||
|
<td>/layer4/layer4.0/conv2/Conv_KNOPT_/layer4/layer4.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer4/layer4.0/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer4/layer4.0/conv2/Conv_KNOPT_/layer4/layer4.0/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>52</th>
|
||||||
|
<td>/layer4/layer4.0/Add</td>
|
||||||
|
<td>/layer4/layer4.0/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer4/layer4.0/Add</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>53</th>
|
||||||
|
<td>/layer4/layer4.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
<td>/layer4/layer4.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer4/layer4.0/Add_KNOPT_dummy_bn_2</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>54</th>
|
||||||
|
<td>/layer4/layer4.1/bn1/BatchNormalization</td>
|
||||||
|
<td>/layer4/layer4.1/bn1/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/layer4/layer4.1/bn1/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>55</th>
|
||||||
|
<td>/layer4/layer4.1/conv1/Conv</td>
|
||||||
|
<td>/layer4/layer4.1/conv1/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer4/layer4.1/conv1/Conv_/layer4/layer4.1/prelu/PRelu</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>56</th>
|
||||||
|
<td>/layer4/layer4.1/prelu/PRelu</td>
|
||||||
|
<td>/layer4/layer4.1/prelu/PRelu</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>57</th>
|
||||||
|
<td>/layer4/layer4.1/conv2/Conv_KNOPT_/layer4/layer4.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
<td>/layer4/layer4.1/conv2/Conv</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>/layer4/layer4.1/conv2/Conv_KNOPT_/layer4/layer4.1/Add_KNOPT_dummy_bn_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>58</th>
|
||||||
|
<td>/layer4/layer4.1/Add</td>
|
||||||
|
<td>/layer4/layer4.1/Add</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8, 8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer4/layer4.1/Add_/bn2/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>59</th>
|
||||||
|
<td>/bn2/BatchNormalization</td>
|
||||||
|
<td>/bn2/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>↑</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>60</th>
|
||||||
|
<td>/Flatten</td>
|
||||||
|
<td>/Flatten</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>None</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>61</th>
|
||||||
|
<td>/fc/Gemm</td>
|
||||||
|
<td>/fc/Gemm</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>npu_fusion_node_/layer4/layer4.1/Add_/bn2/BatchNormalization_KNERON_REFORMAT_next_0</td>
|
||||||
|
</tr>
|
||||||
|
<tr>
|
||||||
|
<th>62</th>
|
||||||
|
<td>/features/BatchNormalization</td>
|
||||||
|
<td>/features/BatchNormalization</td>
|
||||||
|
<td>NPU</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[8]</td>
|
||||||
|
<td>[16]</td>
|
||||||
|
<td>/features/BatchNormalization</td>
|
||||||
|
</tr>
|
||||||
|
</tbody>
|
||||||
|
</table><br><hr>
|
||||||
20
model_fx_report.json
Normal file
20
model_fx_report.json
Normal file
@ -0,0 +1,20 @@
|
|||||||
|
{
|
||||||
|
"docker_version": "kneron/toolchain:v0.29.0",
|
||||||
|
"comments": "",
|
||||||
|
"kdp520/input bitwidth": "int8",
|
||||||
|
"kdp520/output bitwidth": "int8",
|
||||||
|
"kdp520/cpu bitwidth": "int8",
|
||||||
|
"kdp520/datapath bitwidth": "int8",
|
||||||
|
"kdp520/weight bitwidth": "int8",
|
||||||
|
"kdp520/ip_eval/fps": "13.5583",
|
||||||
|
"kdp520/ip_eval/ITC(ms)": "73.7554 ms",
|
||||||
|
"kdp520/ip_eval/RDMA bandwidth GB/s": 0.8,
|
||||||
|
"kdp520/ip_eval/WDMA bandwidth GB/s": 0.8,
|
||||||
|
"kdp520/ip_eval/GETW bandwidth GB/s": 0.8,
|
||||||
|
"kdp520/ip_eval/cpu_node": "N/A",
|
||||||
|
"kdp520/bie": "input.kdp520.scaled.bie",
|
||||||
|
"kdp520/onnx": "input.kdp520.decomposed.onnx",
|
||||||
|
"kdp520/nef": "models_520.nef",
|
||||||
|
"gen fx model report": "model_fx_report.html",
|
||||||
|
"gen fx model json": "model_fx_report.json"
|
||||||
|
}
|
||||||
BIN
models_520.nef
Normal file
BIN
models_520.nef
Normal file
Binary file not shown.
518
nef_test.py
Normal file
518
nef_test.py
Normal file
@ -0,0 +1,518 @@
|
|||||||
|
import os
|
||||||
|
import sys
|
||||||
|
import argparse
|
||||||
|
import kp
|
||||||
|
import cv2
|
||||||
|
import numpy as np
|
||||||
|
from mtcnn.mtcnn import MTCNN
|
||||||
|
import time
|
||||||
|
import pickle
|
||||||
|
import json
|
||||||
|
|
||||||
|
SCPU_FW_PATH = r"C:\Users\mason\AppData\Local\Kneron_Academy\firmware\KL520\fw_scpu.bin"
|
||||||
|
NCPU_FW_PATH = r"C:\Users\mason\AppData\Local\Kneron_Academy\firmware\KL520\fw_ncpu.bin"
|
||||||
|
MODEL_FILE_PATH = 'R34_G369K.nef'
|
||||||
|
IMAGE_FILE_PATH = 'Chou1.jpg'
|
||||||
|
|
||||||
|
def load_image_safe(image_path):
|
||||||
|
if not os.path.isfile(image_path):
|
||||||
|
raise FileNotFoundError(f"[Error] Image file '{image_path}' not found.")
|
||||||
|
img = cv2.imread(image_path)
|
||||||
|
if img is None:
|
||||||
|
raise ValueError(f"[Error] Failed to load image '{image_path}'. Check the file format (must be jpg, png, etc.).")
|
||||||
|
img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
||||||
|
return img_rgb, img
|
||||||
|
|
||||||
|
def landmarks(detector, img_rgb):
|
||||||
|
faces = detector.detect_faces(img_rgb)
|
||||||
|
if len(faces) == 0:
|
||||||
|
raise ValueError("[Error] No faces detected in the image.")
|
||||||
|
face = max(faces, key=lambda x: x['confidence'])
|
||||||
|
return face['keypoints']
|
||||||
|
|
||||||
|
def affine_matrix(lmks, scale=2.5):
|
||||||
|
nose = np.array(lmks['nose'], dtype=np.float32)
|
||||||
|
left_eye = np.array(lmks['left_eye'], dtype=np.float32)
|
||||||
|
right_eye = np.array(lmks['right_eye'], dtype=np.float32)
|
||||||
|
eye_width = right_eye - left_eye
|
||||||
|
angle = np.arctan2(eye_width[1], eye_width[0])
|
||||||
|
center = nose
|
||||||
|
alpha = np.cos(angle)
|
||||||
|
beta = np.sin(angle)
|
||||||
|
w = np.sqrt(np.sum(eye_width**2)) * scale
|
||||||
|
m = [[alpha, beta, -alpha * center[0] - beta * center[1] + w * 0.5],
|
||||||
|
[-beta, alpha, beta * center[0] - alpha * center[1] + w * 0.5]]
|
||||||
|
return np.array(m), (int(w), int(w))
|
||||||
|
|
||||||
|
def extract_vector_data(vector):
|
||||||
|
"""
|
||||||
|
Extract data from the inference output object and convert to a standard numpy array.
|
||||||
|
Always returns a flattened 1D array regardless of input shape.
|
||||||
|
"""
|
||||||
|
try:
|
||||||
|
# For Kneron InferenceFloatNodeOutput specifically
|
||||||
|
if 'InferenceFloatNodeOutput' in str(type(vector)):
|
||||||
|
# Try to access the data directly
|
||||||
|
if hasattr(vector, 'ndarray'):
|
||||||
|
data = vector.ndarray
|
||||||
|
elif hasattr(vector, 'data'):
|
||||||
|
data = vector.data
|
||||||
|
elif hasattr(vector, 'content'):
|
||||||
|
data = vector.content
|
||||||
|
elif hasattr(vector, 'output'):
|
||||||
|
data = vector.output
|
||||||
|
else:
|
||||||
|
# If no direct data attribute, try to access via shape and indexing
|
||||||
|
if hasattr(vector, 'shape'):
|
||||||
|
shape = vector.shape
|
||||||
|
if len(shape) == 4 and shape[0] == 1 and shape[2] == 1 and shape[3] == 1:
|
||||||
|
# Common shape for CNN feature vectors: [1, features, 1, 1]
|
||||||
|
# We need to extract each value manually
|
||||||
|
try:
|
||||||
|
data = np.array([vector[0, i, 0, 0] for i in range(shape[1])], dtype=np.float32)
|
||||||
|
return data # Return early as this is already flat
|
||||||
|
except Exception as e:
|
||||||
|
print(f"Manual extraction failed: {e}")
|
||||||
|
# Continue to other methods
|
||||||
|
|
||||||
|
# Last resort - try numpy conversion
|
||||||
|
data = np.array(vector, dtype=np.float32)
|
||||||
|
|
||||||
|
# Convert to numpy and flatten
|
||||||
|
array_data = np.array(data, dtype=np.float32)
|
||||||
|
return array_data.flatten() # Ensure 1D output
|
||||||
|
|
||||||
|
# For regular numpy arrays or similar
|
||||||
|
if isinstance(vector, np.ndarray):
|
||||||
|
return vector.flatten() # Ensure 1D output
|
||||||
|
|
||||||
|
# For list-like objects
|
||||||
|
if hasattr(vector, 'tolist'):
|
||||||
|
return np.array(vector.tolist(), dtype=np.float32).flatten()
|
||||||
|
|
||||||
|
# Generic conversion
|
||||||
|
return np.array(vector, dtype=np.float32).flatten()
|
||||||
|
|
||||||
|
except Exception as e:
|
||||||
|
print(f"Warning: Error converting vector: {e}")
|
||||||
|
print(f"Vector type: {type(vector)}")
|
||||||
|
if hasattr(vector, 'shape'):
|
||||||
|
print(f"Vector shape: {vector.shape}")
|
||||||
|
|
||||||
|
# Return a properly shaped array of zeros as fallback
|
||||||
|
if hasattr(vector, 'shape') and len(vector.shape) > 0:
|
||||||
|
# Find the largest dimension which is likely the feature dimension
|
||||||
|
max_dim = max(vector.shape)
|
||||||
|
if max_dim > 10: # Reasonable size for a feature vector
|
||||||
|
return np.zeros(max_dim, dtype=np.float32)
|
||||||
|
|
||||||
|
# Default fallback size
|
||||||
|
return np.zeros(512, dtype=np.float32)
|
||||||
|
|
||||||
|
def save_vector(vector, file_path, format='numpy', metadata=None):
|
||||||
|
"""Save a face vector to file using specified format"""
|
||||||
|
directory = os.path.dirname(file_path)
|
||||||
|
if directory and not os.path.exists(directory):
|
||||||
|
os.makedirs(directory)
|
||||||
|
|
||||||
|
# Don't print the entire raw vector - it might be a complex object
|
||||||
|
print(f"Saving vector of type: {type(vector)}")
|
||||||
|
|
||||||
|
# First, try to extract the data into a standard numpy array
|
||||||
|
vector_np = extract_vector_data(vector)
|
||||||
|
|
||||||
|
# Check for all-zeros vector which indicates extraction failed
|
||||||
|
if np.all(vector_np == 0):
|
||||||
|
print("WARNING: Extracted vector contains all zeros - extraction likely failed!")
|
||||||
|
|
||||||
|
# Add extra debugging before giving up
|
||||||
|
print("Attempting emergency extraction methods...")
|
||||||
|
|
||||||
|
# Last-ditch effort - try direct attribute access with common names
|
||||||
|
for attr_name in ['data', 'array', 'values', 'tensor', 'vector', 'features']:
|
||||||
|
if hasattr(vector, attr_name):
|
||||||
|
try:
|
||||||
|
data = getattr(vector, attr_name)
|
||||||
|
vector_np = np.array(data, dtype=np.float32)
|
||||||
|
print(f"Emergency extraction via '{attr_name}' attribute succeeded!")
|
||||||
|
break
|
||||||
|
except:
|
||||||
|
continue
|
||||||
|
|
||||||
|
# Debug information
|
||||||
|
print(f"Extracted vector type: {type(vector_np)}")
|
||||||
|
print(f"Extracted vector shape: {vector_np.shape}")
|
||||||
|
print(f"Sample values: {vector_np[:5]} ... {vector_np[-5:] if len(vector_np) > 5 else []}")
|
||||||
|
|
||||||
|
# Check if the shape needs to be adjusted
|
||||||
|
if len(vector_np.shape) > 1:
|
||||||
|
vector_np = vector_np.squeeze()
|
||||||
|
print(f"Squeezed vector shape: {vector_np.shape}")
|
||||||
|
|
||||||
|
# Save according to format
|
||||||
|
try:
|
||||||
|
if format == 'numpy':
|
||||||
|
np.save(file_path, vector_np)
|
||||||
|
elif format == 'pickle':
|
||||||
|
with open(file_path, 'wb') as f:
|
||||||
|
pickle.dump(vector_np, f)
|
||||||
|
elif format == 'json':
|
||||||
|
data = {
|
||||||
|
'vector': vector_np.tolist(),
|
||||||
|
'metadata': metadata or {}
|
||||||
|
}
|
||||||
|
with open(file_path, 'w') as f:
|
||||||
|
json.dump(data, f)
|
||||||
|
else:
|
||||||
|
raise ValueError(f"Unsupported format: {format}")
|
||||||
|
|
||||||
|
print(f"Vector saved to {file_path}")
|
||||||
|
return file_path
|
||||||
|
except Exception as e:
|
||||||
|
print(f"Error saving vector: {e}")
|
||||||
|
|
||||||
|
# Alternative save method if standard methods fail
|
||||||
|
if format == 'numpy' or format == 'pickle':
|
||||||
|
# Try saving as JSON as a fallback
|
||||||
|
try:
|
||||||
|
alt_path = file_path + '.json'
|
||||||
|
data = {
|
||||||
|
'vector': vector_np.tolist(),
|
||||||
|
'metadata': metadata or {}
|
||||||
|
}
|
||||||
|
with open(alt_path, 'w') as f:
|
||||||
|
json.dump(data, f)
|
||||||
|
print(f"Vector saved using alternative method to {alt_path}")
|
||||||
|
return alt_path
|
||||||
|
except Exception as e2:
|
||||||
|
print(f"Alternative save method also failed: {e2}")
|
||||||
|
return None
|
||||||
|
return None
|
||||||
|
|
||||||
|
def load_vector(file_path, format='numpy'):
|
||||||
|
"""Load a face vector from file using specified format"""
|
||||||
|
if not os.path.isfile(file_path):
|
||||||
|
raise FileNotFoundError(f"Vector file '{file_path}' not found.")
|
||||||
|
|
||||||
|
if format == 'numpy' or file_path.endswith('.npy'):
|
||||||
|
return np.load(file_path)
|
||||||
|
elif format == 'pickle' or file_path.endswith('.pkl'):
|
||||||
|
with open(file_path, 'rb') as f:
|
||||||
|
return pickle.load(f)
|
||||||
|
elif format == 'json' or file_path.endswith('.json'):
|
||||||
|
with open(file_path, 'r') as f:
|
||||||
|
data = json.load(f)
|
||||||
|
return np.array(data['vector']), data.get('metadata')
|
||||||
|
else:
|
||||||
|
raise ValueError(f"Unsupported format: {format}")
|
||||||
|
|
||||||
|
def visualize_alignment(original_img, lmks, aligned_img, save_path='alignment_visualization.jpg'):
|
||||||
|
"""
|
||||||
|
視覺化MTCNN檢測到的特徵點和對齊後的結果
|
||||||
|
|
||||||
|
Args:
|
||||||
|
original_img: 原始BGR圖像
|
||||||
|
lmks: MTCNN檢測到的特徵點字典 ('left_eye', 'right_eye', 'nose', etc.)
|
||||||
|
aligned_img: 對齊後的人臉圖像
|
||||||
|
save_path: 保存視覺化結果的路徑
|
||||||
|
"""
|
||||||
|
import matplotlib.pyplot as plt
|
||||||
|
|
||||||
|
# 創建原始圖像的副本用於繪製
|
||||||
|
img_vis = original_img.copy()
|
||||||
|
|
||||||
|
# 在原始圖像上繪製特徵點
|
||||||
|
cv2.circle(img_vis, tuple(map(int, lmks['left_eye'])), 5, (0, 255, 0), -1)
|
||||||
|
cv2.circle(img_vis, tuple(map(int, lmks['right_eye'])), 5, (0, 255, 0), -1)
|
||||||
|
cv2.circle(img_vis, tuple(map(int, lmks['nose'])), 5, (0, 255, 0), -1)
|
||||||
|
cv2.circle(img_vis, tuple(map(int, lmks['mouth_left'])), 5, (0, 255, 0), -1)
|
||||||
|
cv2.circle(img_vis, tuple(map(int, lmks['mouth_right'])), 5, (0, 255, 0), -1)
|
||||||
|
|
||||||
|
# 在眼睛之間繪製線條，顯示對齊參考線
|
||||||
|
cv2.line(img_vis,
|
||||||
|
tuple(map(int, lmks['left_eye'])),
|
||||||
|
tuple(map(int, lmks['right_eye'])),
|
||||||
|
(255, 0, 0), 2)
|
||||||
|
|
||||||
|
# 創建一個圖形來顯示兩張圖像
|
||||||
|
plt.figure(figsize=(12, 6))
|
||||||
|
|
||||||
|
# 顯示帶有特徵點的原始圖像
|
||||||
|
plt.subplot(1, 2, 1)
|
||||||
|
plt.imshow(cv2.cvtColor(img_vis, cv2.COLOR_BGR2RGB))
|
||||||
|
plt.title('原始圖像與特徵點')
|
||||||
|
plt.axis('off')
|
||||||
|
|
||||||
|
# 顯示對齊後的圖像
|
||||||
|
plt.subplot(1, 2, 2)
|
||||||
|
plt.imshow(cv2.cvtColor(aligned_img, cv2.COLOR_BGR2RGB))
|
||||||
|
plt.title('對齊後的人臉')
|
||||||
|
plt.axis('off')
|
||||||
|
|
||||||
|
plt.tight_layout()
|
||||||
|
plt.savefig(save_path)
|
||||||
|
plt.show()
|
||||||
|
|
||||||
|
print(f"視覺化結果已保存到 '{save_path}'")
|
||||||
|
|
||||||
|
|
||||||
|
def cosine_similarity(vec1, vec2):
|
||||||
|
"""
|
||||||
|
Calculate cosine similarity between two vectors.
|
||||||
|
Handles different input shapes by flattening both vectors.
|
||||||
|
"""
|
||||||
|
# Ensure both vectors are numpy arrays
|
||||||
|
vec1 = np.array(vec1, dtype=np.float32)
|
||||||
|
vec2 = np.array(vec2, dtype=np.float32)
|
||||||
|
|
||||||
|
# Flatten both vectors to ensure 1D
|
||||||
|
vec1 = vec1.flatten()
|
||||||
|
vec2 = vec2.flatten()
|
||||||
|
|
||||||
|
# Check if vectors have compatible sizes
|
||||||
|
if vec1.size != vec2.size:
|
||||||
|
print(f"Warning: Vector size mismatch: {vec1.size} vs {vec2.size}")
|
||||||
|
# Resize shorter vector or truncate longer vector
|
||||||
|
if vec1.size < vec2.size:
|
||||||
|
vec2 = vec2[:vec1.size]
|
||||||
|
else:
|
||||||
|
vec1 = vec1[:vec2.size]
|
||||||
|
|
||||||
|
# Calculate cosine similarity
|
||||||
|
dot_product = np.dot(vec1, vec2)
|
||||||
|
norm_a = np.linalg.norm(vec1)
|
||||||
|
norm_b = np.linalg.norm(vec2)
|
||||||
|
|
||||||
|
# Handle zero division
|
||||||
|
if norm_a < 1e-10 or norm_b < 1e-10:
|
||||||
|
print("Warning: Vector with near-zero magnitude detected")
|
||||||
|
return 0.0
|
||||||
|
|
||||||
|
return dot_product / (norm_a * norm_b)
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
parser = argparse.ArgumentParser(description='KL520 ResNet18 model image inference implementation')
|
||||||
|
parser.add_argument('-p',
|
||||||
|
'--port_id',
|
||||||
|
help='Using specified port ID for connecting device (Default: port ID of first scanned Kneron device)',
|
||||||
|
default=28,
|
||||||
|
type=int)
|
||||||
|
parser.add_argument('-m',
|
||||||
|
'--model',
|
||||||
|
help='Model file path (.nef) (Default: {})'.format(MODEL_FILE_PATH),
|
||||||
|
default=MODEL_FILE_PATH,
|
||||||
|
type=str)
|
||||||
|
parser.add_argument('-i',
|
||||||
|
'--img',
|
||||||
|
help='Image file path (Default: {})'.format(IMAGE_FILE_PATH),
|
||||||
|
default=IMAGE_FILE_PATH,
|
||||||
|
type=str)
|
||||||
|
parser.add_argument('-o',
|
||||||
|
'--output',
|
||||||
|
help='Output vector file path (Default: output.npy)',
|
||||||
|
default='face_vectors\\output.npy',
|
||||||
|
type=str)
|
||||||
|
parser.add_argument('-f',
|
||||||
|
'--format',
|
||||||
|
help='Output format: numpy, pickle, or json (Default: numpy)',
|
||||||
|
default='numpy',
|
||||||
|
choices=['numpy', 'pickle', 'json'],
|
||||||
|
type=str)
|
||||||
|
parser.add_argument('-n',
|
||||||
|
'--name',
|
||||||
|
help='Person name for the face vector (for metadata)',
|
||||||
|
default=None,
|
||||||
|
type=str)
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
usb_port_id = args.port_id
|
||||||
|
MODEL_FILE_PATH = args.model
|
||||||
|
IMAGE_FILE_PATH = args.img
|
||||||
|
|
||||||
|
"""
|
||||||
|
connect the device
|
||||||
|
"""
|
||||||
|
try:
|
||||||
|
print('[Connect Device]')
|
||||||
|
device_group = kp.core.connect_devices(usb_port_ids=[usb_port_id])
|
||||||
|
print(' - Success')
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print('Error: connect device fail, port ID = \'{}\', error msg: [{}]'.format(usb_port_id,
|
||||||
|
str(exception)))
|
||||||
|
exit(0)
|
||||||
|
|
||||||
|
"""
|
||||||
|
setting timeout of the usb communication with the device
|
||||||
|
"""
|
||||||
|
print('[Set Device Timeout]')
|
||||||
|
kp.core.set_timeout(device_group=device_group, milliseconds=5000)
|
||||||
|
print(' - Success')
|
||||||
|
|
||||||
|
"""
|
||||||
|
upload firmware to device
|
||||||
|
"""
|
||||||
|
try:
|
||||||
|
print('[Upload Firmware]')
|
||||||
|
kp.core.load_firmware_from_file(device_group=device_group,
|
||||||
|
scpu_fw_path=SCPU_FW_PATH,
|
||||||
|
ncpu_fw_path=NCPU_FW_PATH)
|
||||||
|
print(' - Success')
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print('Error: upload firmware failed, error = \'{}\''.format(str(exception)))
|
||||||
|
exit(0)
|
||||||
|
|
||||||
|
"""
|
||||||
|
upload model to device
|
||||||
|
"""
|
||||||
|
try:
|
||||||
|
print('[Upload Model]')
|
||||||
|
model_nef_descriptor = kp.core.load_model_from_file(device_group=device_group,
|
||||||
|
file_path=MODEL_FILE_PATH)
|
||||||
|
print(' - Success')
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print('Error: upload model failed, error = \'{}\''.format(str(exception)))
|
||||||
|
exit(0)
|
||||||
|
|
||||||
|
"""
|
||||||
|
MTCNN Part
|
||||||
|
"""
|
||||||
|
print('[Process MTCNN]')
|
||||||
|
start = time.time()
|
||||||
|
# Create MTCNN detector
|
||||||
|
detector = MTCNN(device="CPU:0")
|
||||||
|
|
||||||
|
# Load image
|
||||||
|
try:
|
||||||
|
img_rgb, img_bgr = load_image_safe(IMAGE_FILE_PATH)
|
||||||
|
print(f" - Image loaded: {IMAGE_FILE_PATH}")
|
||||||
|
except Exception as e:
|
||||||
|
print(str(e))
|
||||||
|
exit(0)
|
||||||
|
|
||||||
|
# Get landmarks and calculate affine matrix
|
||||||
|
try:
|
||||||
|
lmks = landmarks(detector, img_rgb)
|
||||||
|
mat, size = affine_matrix(lmks)
|
||||||
|
print(" - Face landmarks detected")
|
||||||
|
except Exception as e:
|
||||||
|
print(str(e))
|
||||||
|
exit(0)
|
||||||
|
|
||||||
|
# Apply affine transformation
|
||||||
|
aligned_img = cv2.warpAffine(img_bgr, mat, size)
|
||||||
|
|
||||||
|
end = time.time()
|
||||||
|
print(f" - MTCNN processing time: {end - start:.2f} seconds")
|
||||||
|
|
||||||
|
# 添加此行以可視化對齊過程
|
||||||
|
visualize_alignment(img_bgr, lmks, aligned_img)
|
||||||
|
|
||||||
|
# Convert aligned_img to BGR565 format and resize
|
||||||
|
aligned_img_bgr565 = cv2.cvtColor(aligned_img, cv2.COLOR_BGR2BGR565)
|
||||||
|
img_bgr565 = cv2.resize(aligned_img_bgr565, (112, 112), interpolation=cv2.INTER_LINEAR)
|
||||||
|
print(" - Image aligned and formatted for inference")
|
||||||
|
|
||||||
|
"""
|
||||||
|
prepare generic image inference input descriptor
|
||||||
|
"""
|
||||||
|
generic_inference_input_descriptor = kp.GenericImageInferenceDescriptor(
|
||||||
|
model_id=model_nef_descriptor.models[0].id,
|
||||||
|
inference_number=0,
|
||||||
|
input_node_image_list=[
|
||||||
|
kp.GenericInputNodeImage(
|
||||||
|
image=img_bgr565,
|
||||||
|
image_format=kp.ImageFormat.KP_IMAGE_FORMAT_RGB565,
|
||||||
|
resize_mode=kp.ResizeMode.KP_RESIZE_ENABLE,
|
||||||
|
padding_mode=kp.PaddingMode.KP_PADDING_CORNER,
|
||||||
|
normalize_mode=kp.NormalizeMode.KP_NORMALIZE_KNERON
|
||||||
|
)
|
||||||
|
]
|
||||||
|
)
|
||||||
|
|
||||||
|
"""
|
||||||
|
starting inference work
|
||||||
|
"""
|
||||||
|
print('[Starting Inference Work]')
|
||||||
|
try:
|
||||||
|
kp.inference.generic_image_inference_send(device_group=device_group,
|
||||||
|
generic_inference_input_descriptor=generic_inference_input_descriptor)
|
||||||
|
|
||||||
|
generic_raw_result = kp.inference.generic_image_inference_receive(device_group=device_group)
|
||||||
|
print(" - Inference completed successfully")
|
||||||
|
except kp.ApiKPException as exception:
|
||||||
|
print(' - Error: inference failed, error = {}'.format(exception))
|
||||||
|
exit(0)
|
||||||
|
|
||||||
|
"""
|
||||||
|
retrieve inference node output
|
||||||
|
"""
|
||||||
|
print('[Retrieve Inference Node Output]')
|
||||||
|
inf_node_output_list = []
|
||||||
|
for node_idx in range(generic_raw_result.header.num_output_node):
|
||||||
|
inference_float_node_output = kp.inference.generic_inference_retrieve_float_node(
|
||||||
|
node_idx=node_idx,
|
||||||
|
generic_raw_result=generic_raw_result,
|
||||||
|
channels_ordering=kp.ChannelOrdering.KP_CHANNEL_ORDERING_CHW
|
||||||
|
)
|
||||||
|
inf_node_output_list.append(inference_float_node_output)
|
||||||
|
print(' - Success')
|
||||||
|
|
||||||
|
"""
|
||||||
|
Process and save the face embedding vector
|
||||||
|
"""
|
||||||
|
# For face recognition models, typically the output is a feature vector
|
||||||
|
# Usually, the feature vector is in the first output node
|
||||||
|
if len(inf_node_output_list) > 0:
|
||||||
|
face_vector = inf_node_output_list[0]
|
||||||
|
# print(face_vector)
|
||||||
|
print(f"[Face Vector] Original type: {type(face_vector)}")
|
||||||
|
print(f"[Face Vector] Shape: {face_vector.shape}")
|
||||||
|
|
||||||
|
# Try to examine the vector object
|
||||||
|
print("[Face Vector] Available attributes:", [attr for attr in dir(face_vector) if not attr.startswith('__')])
|
||||||
|
|
||||||
|
# Create metadata if name is provided
|
||||||
|
metadata = None
|
||||||
|
if args.name:
|
||||||
|
metadata = {
|
||||||
|
'name': args.name,
|
||||||
|
'image_path': IMAGE_FILE_PATH,
|
||||||
|
'timestamp': time.strftime("%Y-%m-%d %H:%M:%S"),
|
||||||
|
'dimensions': 512,
|
||||||
|
}
|
||||||
|
|
||||||
|
output_format = args.format
|
||||||
|
output_path = args.output
|
||||||
|
# If saving numpy and the file name doesn’t end with .npy, add it
|
||||||
|
if output_format == 'numpy' and not output_path.endswith('.npy'):
|
||||||
|
output_path += '.npy'
|
||||||
|
# Likewise for pickle
|
||||||
|
if output_format == 'pickle' and not (output_path.endswith('.pkl') or output_path.endswith('.pickle')):
|
||||||
|
output_path += '.pkl'
|
||||||
|
|
||||||
|
|
||||||
|
# Save the vector
|
||||||
|
output_path = save_vector(
|
||||||
|
face_vector,
|
||||||
|
output_path,
|
||||||
|
format=output_format,
|
||||||
|
metadata=metadata
|
||||||
|
)
|
||||||
|
|
||||||
|
if output_path:
|
||||||
|
print(f"[Result] Face embedding vector saved to: {output_path}")
|
||||||
|
print(f" - Format: {output_format}")
|
||||||
|
if metadata:
|
||||||
|
print(f" - Metadata: {metadata}")
|
||||||
|
else:
|
||||||
|
print("[Error] Failed to save the vector")
|
||||||
|
else:
|
||||||
|
print("[Error] No output nodes found in the inference result")
|
||||||
|
|
||||||
|
# Clean up
|
||||||
|
kp.core.disconnect_devices(device_group=device_group)
|
||||||
|
print("[Cleanup] Device disconnected")
|
||||||
|
|
||||||
|
if output_format == 'numpy':
|
||||||
|
loaded = np.load(output_path)
|
||||||
|
print("Reload check:", loaded.shape, loaded.dtype)
|
||||||
BIN
output.npy
Normal file
BIN
output.npy
Normal file
Binary file not shown.
47
test_npy.py
Normal file
47
test_npy.py
Normal file
@ -0,0 +1,47 @@
|
|||||||
|
import numpy as np
|
||||||
|
import numpy.lib.format as fmt
|
||||||
|
import argparse
|
||||||
|
import os
|
||||||
|
import sys
|
||||||
|
|
||||||
|
def main():
|
||||||
|
parser = argparse.ArgumentParser(description="讀取 .npy 檔並輸出內容 (診斷模式)")
|
||||||
|
parser.add_argument("npy_path", help="要讀取的 .npy 檔案路徑")
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
npy_path = args.npy_path
|
||||||
|
if not os.path.isfile(npy_path):
|
||||||
|
print(f"Error: 檔案不存在：{npy_path}")
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
size = os.path.getsize(npy_path)
|
||||||
|
print(f"File size: {size} bytes")
|
||||||
|
if size == 0:
|
||||||
|
print("Error: 檔案是空的，請確認產生並儲存時沒有錯誤。")
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
# 嘗試只讀 header
|
||||||
|
with open(npy_path, 'rb') as f:
|
||||||
|
try:
|
||||||
|
version = fmt.read_magic(f)
|
||||||
|
header = fmt._read_array_header(f, version)
|
||||||
|
print("成功讀取 header:", header)
|
||||||
|
except Exception as e:
|
||||||
|
print("讀 header 失敗：", e)
|
||||||
|
|
||||||
|
try:
|
||||||
|
data = np.load(npy_path, allow_pickle=True)
|
||||||
|
except Exception as e:
|
||||||
|
print(f"載入失敗：{e}")
|
||||||
|
sys.exit(1)
|
||||||
|
|
||||||
|
print(f"Loaded data from: {npy_path}")
|
||||||
|
print(f"Data type: {type(data)}")
|
||||||
|
if isinstance(data, np.ndarray):
|
||||||
|
print(f"Array shape: {data.shape}")
|
||||||
|
print(f"Array dtype: {data.dtype}")
|
||||||
|
print("Contents:")
|
||||||
|
print(data)
|
||||||
|
|
||||||
|
if __name__ == "__main__":
|
||||||
|
main()
|
||||||
BIN
test_picture/Chou1.jpg
Normal file
BIN
test_picture/Chou1.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 28 KiB |
BIN
test_picture/Chou2.jpg
Normal file
BIN
test_picture/Chou2.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 145 KiB |
BIN
test_picture/Chou3.jpg
Normal file
BIN
test_picture/Chou3.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 29 KiB |
BIN
test_picture/Chou4.jpg
Normal file
BIN
test_picture/Chou4.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 27 KiB |
BIN
test_picture/WS1.jpg
Normal file
BIN
test_picture/WS1.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 186 KiB |
BIN
test_picture/WS2.jpg
Normal file
BIN
test_picture/WS2.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 220 KiB |
BIN
test_picture/Will_Smith.jpg
Normal file
BIN
test_picture/Will_Smith.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 324 KiB |
BIN
test_picture/t.jpg
Normal file
BIN
test_picture/t.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 38 KiB |
BIN
test_picture/t2.jpg
Normal file
BIN
test_picture/t2.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 20 KiB |
BIN
test_picture/test.jpg
Normal file
BIN
test_picture/test.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 78 KiB |
BIN
test_picture/ziyu.jpg
Normal file
BIN
test_picture/ziyu.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 109 KiB |
BIN
test_picture/ziyu2.jpg
Normal file
BIN
test_picture/ziyu2.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 30 KiB |
BIN
test_picture/ziyu4.jpg
Normal file
BIN
test_picture/ziyu4.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 44 KiB |
Loading…
x
Reference in New Issue
Block a user