samH98/LungCancerDetection
0
1import argparse2import sys3import time4import warnings5 6sys.path.append('./') # to run '$ python *.py' files in subdirectories7 8import torch9import torch.nn as nn10from torch.utils.mobile_optimizer import optimize_for_mobile11 12import models13from models.experimental import attempt_load, End2End14from utils.activations import Hardswish, SiLU15from utils.general import set_logging, check_img_size16from utils.torch_utils import select_device17from utils.add_nms import RegisterNMS18 19if __name__ == '__main__':20 parser = argparse.ArgumentParser()21 parser.add_argument('--weights', type=str, default='./yolor-csp-c.pt', help='weights path')22 parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size') # height, width23 parser.add_argument('--batch-size', type=int, default=1, help='batch size')24 parser.add_argument('--dynamic', action='store_true', help='dynamic ONNX axes')25 parser.add_argument('--dynamic-batch', action='store_true', help='dynamic batch onnx for tensorrt and onnx-runtime')26 parser.add_argument('--grid', action='store_true', help='export Detect() layer grid')27 parser.add_argument('--end2end', action='store_true', help='export end2end onnx')28 parser.add_argument('--max-wh', type=int, default=None, help='None for tensorrt nms, int value for onnx-runtime nms')29 parser.add_argument('--topk-all', type=int, default=100, help='topk objects for every images')30 parser.add_argument('--iou-thres', type=float, default=0.45, help='iou threshold for NMS')31 parser.add_argument('--conf-thres', type=float, default=0.25, help='conf threshold for NMS')32 parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')33 parser.add_argument('--simplify', action='store_true', help='simplify onnx model')34 parser.add_argument('--include-nms', action='store_true', help='export end2end onnx')35 parser.add_argument('--fp16', action='store_true', help='CoreML FP16 half-precision export')36 parser.add_argument('--int8', action='store_true', help='CoreML INT8 quantization')37 opt = parser.parse_args()38 opt.img_size *= 2 if len(opt.img_size) == 1 else 1 # expand39 opt.dynamic = opt.dynamic and not opt.end2end40 opt.dynamic = False if opt.dynamic_batch else opt.dynamic41 print(opt)42 set_logging()43 t = time.time()44 45 # Load PyTorch model46 device = select_device(opt.device)47 model = attempt_load(opt.weights, map_location=device) # load FP32 model48 labels = model.names49 50 # Checks51 gs = int(max(model.stride)) # grid size (max stride)52 opt.img_size = [check_img_size(x, gs) for x in opt.img_size] # verify img_size are gs-multiples53 54 # Input55 img = torch.zeros(opt.batch_size, 3, *opt.img_size).to(device) # image size(1,3,320,192) iDetection56 57 # Update model58 for k, m in model.named_modules():59 m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility60 if isinstance(m, models.common.Conv): # assign export-friendly activations61 if isinstance(m.act, nn.Hardswish):62 m.act = Hardswish()63 elif isinstance(m.act, nn.SiLU):64 m.act = SiLU()65 # elif isinstance(m, models.yolo.Detect):66 # m.forward = m.forward_export # assign forward (optional)67 model.model[-1].export = not opt.grid # set Detect() layer grid export68 y = model(img) # dry run69 if opt.include_nms:70 model.model[-1].include_nms = True71 y = None72 73 # TorchScript export74 try:75 print('\nStarting TorchScript export with torch %s...' % torch.__version__)76 f = opt.weights.replace('.pt', '.torchscript.pt') # filename77 ts = torch.jit.trace(model, img, strict=False)78 ts.save(f)79 print('TorchScript export success, saved as %s' % f)80 except Exception as e:81 print('TorchScript export failure: %s' % e)82 83 # CoreML export84 try:85 import coremltools as ct86 87 print('\nStarting CoreML export with coremltools %s...' % ct.__version__)88 # convert model from torchscript and apply pixel scaling as per detect.py89 ct_model = ct.convert(ts, inputs=[ct.ImageType('image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])])90 bits, mode = (8, 'kmeans_lut') if opt.int8 else (16, 'linear') if opt.fp16 else (32, None)91 if bits < 32:92 if sys.platform.lower() == 'darwin': # quantization only supported on macOS93 with warnings.catch_warnings():94 warnings.filterwarnings("ignore", category=DeprecationWarning) # suppress numpy==1.20 float warning95 ct_model = ct.models.neural_network.quantization_utils.quantize_weights(ct_model, bits, mode)96 else:97 print('quantization only supported on macOS, skipping...')98 99 f = opt.weights.replace('.pt', '.mlmodel') # filename100 ct_model.save(f)101 print('CoreML export success, saved as %s' % f)102 except Exception as e:103 print('CoreML export failure: %s' % e)104 105 # TorchScript-Lite export106 try:107 print('\nStarting TorchScript-Lite export with torch %s...' % torch.__version__)108 f = opt.weights.replace('.pt', '.torchscript.ptl') # filename109 tsl = torch.jit.trace(model, img, strict=False)110 tsl = optimize_for_mobile(tsl)111 tsl._save_for_lite_interpreter(f)112 print('TorchScript-Lite export success, saved as %s' % f)113 except Exception as e:114 print('TorchScript-Lite export failure: %s' % e)115 116 # ONNX export117 try:118 import onnx119 120 print('\nStarting ONNX export with onnx %s...' % onnx.__version__)121 f = opt.weights.replace('.pt', '.onnx') # filename122 model.eval()123 output_names = ['classes', 'boxes'] if y is None else ['output']124 dynamic_axes = None125 if opt.dynamic:126 dynamic_axes = {'images': {0: 'batch', 2: 'height', 3: 'width'}, # size(1,3,640,640)127 'output': {0: 'batch', 2: 'y', 3: 'x'}}128 if opt.dynamic_batch:129 opt.batch_size = 'batch'130 dynamic_axes = {131 'images': {132 0: 'batch',133 }, }134 if opt.end2end and opt.max_wh is None:135 output_axes = {136 'num_dets': {0: 'batch'},137 'det_boxes': {0: 'batch'},138 'det_scores': {0: 'batch'},139 'det_classes': {0: 'batch'},140 }141 else:142 output_axes = {143 'output': {0: 'batch'},144 }145 dynamic_axes.update(output_axes)146 if opt.grid:147 if opt.end2end:148 print('\nStarting export end2end onnx model for %s...' % 'TensorRT' if opt.max_wh is None else 'onnxruntime')149 model = End2End(model,opt.topk_all,opt.iou_thres,opt.conf_thres,opt.max_wh,device,len(labels))150 if opt.end2end and opt.max_wh is None:151 output_names = ['num_dets', 'det_boxes', 'det_scores', 'det_classes']152 shapes = [opt.batch_size, 1, opt.batch_size, opt.topk_all, 4,153 opt.batch_size, opt.topk_all, opt.batch_size, opt.topk_all]154 else:155 output_names = ['output']156 else:157 model.model[-1].concat = True158 159 torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'],160 output_names=output_names,161 dynamic_axes=dynamic_axes)162 163 # Checks164 onnx_model = onnx.load(f) # load onnx model165 onnx.checker.check_model(onnx_model) # check onnx model166 167 if opt.end2end and opt.max_wh is None:168 for i in onnx_model.graph.output:169 for j in i.type.tensor_type.shape.dim:170 j.dim_param = str(shapes.pop(0))171 172 # print(onnx.helper.printable_graph(onnx_model.graph)) # print a human readable model173 174 # # Metadata175 # d = {'stride': int(max(model.stride))}176 # for k, v in d.items():177 # meta = onnx_model.metadata_props.add()178 # meta.key, meta.value = k, str(v)179 # onnx.save(onnx_model, f)180 181 if opt.simplify:182 try:183 import onnxsim184 185 print('\nStarting to simplify ONNX...')186 onnx_model, check = onnxsim.simplify(onnx_model)187 assert check, 'assert check failed'188 except Exception as e:189 print(f'Simplifier failure: {e}')190 191 # print(onnx.helper.printable_graph(onnx_model.graph)) # print a human readable model192 onnx.save(onnx_model,f)193 print('ONNX export success, saved as %s' % f)194 195 if opt.include_nms:196 print('Registering NMS plugin for ONNX...')197 mo = RegisterNMS(f)198 mo.register_nms()199 mo.save(f)200 201 except Exception as e:202 print('ONNX export failure: %s' % e)203 204 # Finish205 print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t))206