import numpy as np from PIL import Image import argparse import os # import matplotlib.pyplot as plt """ ATTENTION: Current supported preprocessing includes: 2 resize styles: resize and resize with padding right and bottom. 4 normalization modes: kneron, tf, caffe, torch """ def preprocess(image, clrspace, pos, crop_dim, resize_dim, keep_aspect_ratio, mode, b_crop, outFile, bSaveCrop, bcvtOnly): """ image: PIL Image object clrspace: one of "RGB", "BGR", "L" size: (width, height) keep_aspect_ratio: True or False mode: one of "kneron", "caffe", "tf", "torch" """ if clrspace == "BGR": image = image.convert("RGB") image = Image.fromarray(np.array(image)[...,::-1]) else: image = image.convert(clrspace) if bcvtOnly == 0: if(b_crop): image = do_crop(image, pos, crop_dim, outFile) img_data = resize(image, resize_dim, keep_aspect_ratio, outFile, bSaveCrop, clrspace) else: img_data = np.array(image) img_data = normalization(img_data, mode) return img_data def resize(x, size, keep_aspect_ratio, oFile, bSave, clrspace): if not keep_aspect_ratio: return np.array(x.resize(size, Image.BILINEAR)) else: x.thumbnail(size, Image.ANTIALIAS) #if bSave == True: # x.save(oFile + "_resize_out.jpg") x = x.crop((0, 0, size[0], size[1])) #if bSave == True: # x.save(oFile + "_final_rgba.jpg") iw, ih = x.size w, h = size scale = min(w/iw, h/ih) nw = int(iw*scale) nh = int(ih*scale) x = x.resize((nw,nh), Image.BILINEAR) new_image = Image.new('RGB', size, (0,0,0)) # new_image.paste(image, ((w-nw)//2, (h-nh)//2)) # keras setting new_image.paste(x, (0,0)) # npu setting final = np.array(new_image) if clrspace == "L" or clrspace == "l": final = final[:,:,0] return final def do_crop(x, pos, size, oFile): x = x.crop((pos[0], pos[1], size[0], size[1])) # x.save(oFile + "_crop_out.jpg") return x def normalization(x, mode, **kwargs): """Normalize a Numpy array encoding a batch of images. # Arguments x: Input array, 3D or 4D. mode: One of "kneron","caffe", "tf" or "torch". - caffe: will convert the images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. - tf: will scale pixels between -1 and 1, sample-wise. - torch: will scale pixels between 0 and 1 and then will normalize each channel with respect to the ImageNet dataset. # Returns Preprocessed Numpy array. """ x = x.astype(float) if mode == 'none': return x if mode == 'yolo': # 0 -1 8-7 x /= 255. return x if mode == 'kneron': #-0.5～0.5 x /= 256. x -= 0.5 return x if mode == 'tensorflow': # -1 ～ 1 8-7 x /= 127.5 x -= 1. return x if mode == 'pytorch': #-2.5～2.5 8-6 x /= 255. mean = [0.485, 0.456, 0.406] std = [0.229, 0.224, 0.225] x[..., 0] -= mean[0] x[..., 1] -= mean[1] x[..., 2] -= mean[2] if std is not None: x[..., 0] /= std[0] x[..., 1] /= std[1] x[..., 2] /= std[2] return x if mode == 'caffe': #-123 - 123 8-0 ### mean is for BGR format mean = [103.939, 116.779, 123.68] std = None x[..., 0] -= mean[0] x[..., 1] -= mean[1] x[..., 2] -= mean[2] if std is not None: x[..., 0] /= std[0] x[..., 1] /= std[1] x[..., 2] /= std[2] return x #this is the customized part if mode == 'customized': #-123 - 123 8-0 print("customized") y = (x - 127.5) * 0.0078125 return y def img2txt_bin(input_file : str, input_is_dir : bool, output_folder : str, out_h : int, out_w : int, clrspace : str, mode : str, task : str, kdp_version : int, radix : int, bitwidth : int = 0, x_pos : int = 0, y_pos : int = 0, crop_h : int = 0, crop_w : int = 0, bCvtOnly : int = 0, enable_crop : bool = False, keep_aspect_ratio : bool = False ): # Prepare inputs if input_is_dir: input_list = [os.path.join(input_file, file_name) for file_name in os.listdir(input_file)] else: input_list = [input_file] # Prepare output folder if not os.path.exists(output_folder): os.mkdir(output_folder) # For each input file for in_file in input_list: x = x_pos y = y_pos w = x + crop_w h = y + crop_h crop_size = w,h resize_size = out_w, out_h if enable_crop: pos = x, y bCrop = False else: pos = 0, 0 bCrop = False # print(in_file) if os.path.exists(in_file) != True: assert(0); imagefile = in_file; image = Image.open(imagefile) if task == "img2bin": bNeedSaveCrop = True; else: bNeedSaveCrop = False; img_data = preprocess(image, clrspace, pos, crop_size, resize_size, keep_aspect_ratio, mode, bCrop, "temp.txt", bNeedSaveCrop, bCvtOnly) if task == "img2txt": out_file = os.path.join(output_folder, in_file.split("/")[-1].split(".")[0] + ".txt") np.savetxt(out_file, img_data.reshape((-1, 1)), fmt="%.8f") elif task == "img2bin": col, row = resize_size if kdp_version == 520: col_output = int(np.ceil(col*1.0/16)*16) else: col_output = col img_data = np.clip(img_data * (2**radix) , a_min=-1 * 2**(bitwidth-1), a_max=2**(bitwidth-1) - 1) img_data = np.around(img_data).astype("int") out_file = os.path.join(output_folder, in_file.split("/")[-1].split(".")[0] + ".bin") binfile = out_file img_data_output = np.zeros((row, col_output, 4)) if len(clrspace) == 3: img_data_output[:,:col,:3] = img_data else: img_data_output[:,:col,0] = img_data img_data_output[:,:col,1] = img_data img_data_output[:,:col,2] = img_data img_data_output.astype("int8").tofile(binfile) def main_(args): img2txt_bin(args.input_folder, True, args.output_folder, args.s_h, args.s_w, args.color_space, args.mode, args.task, args.kdp_version, args.radix, args.bitwidth) return if __name__ == "__main__": argparser = argparse.ArgumentParser( description="convert img to RGB, resize or crop, etc" ) argparser.add_argument( '-t', '--task', help=" 'img2bin', 'img2txt'" ) argparser.add_argument( '-i', '--input_file', help="input file name" ) argparser.add_argument( '-if', '--input_folder', help="input folder path" ) argparser.add_argument( '-o', '--output_file', help="output file name" ) argparser.add_argument( '-of', '--output_folder', help="output folder path" ) argparser.add_argument( '-c', '--color_space', help="L, RGB or BGR" ) 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( '-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( '-a', '--keep_aspect_ratio', help="True/False to indicate maintain aspect_ratio or not" ) argparser.add_argument( '-m', '--mode', help="normalizaton mode: yolo, kneron, caffe, tf , torch." ) argparser.add_argument( '-bw', '--bitwidth', type=int, help="Int for bitwidth" ) argparser.add_argument( '-r', '--radix', type=int, help="Int for radix" ) argparser.add_argument( '-crop', '--enCrop', type=int, help="indicate if this task will act Crop" ) argparser.add_argument( '-cvtOnly', '--convertOnly', type=int, help="indicate if just do format convert" ) argparser.add_argument( '--kdp_version', type=int, default=520, help="kdp version" ) args = argparser.parse_args() main_(args)