这是官方开发者手册少数几个onnx相关的demo之一。 环境: ubuntu 16.04 python 3.5 torch 1.0.1 GTX1080 torchvision0.2.2 借助这个demo体会一下onnx的自定义层的定义/使用方式。看了readme之后发现，yolov3用到的自定义层，比如upsample层是在pytorch转onnx的时候就定义了解析方式的，在onnx转TensorRT的engine过程中并没有什么特殊的操作，所以决定来详细理解一下这个代码。尤其是yolov3转onnx的部分。
main函数先
def main(): if sys.version_info[0] > 2: raise Exception("This script is only compatible with python2, please re-run this script with python2. The rest of this sample can be run with either version of python.") cfg_file_path = download_file( 'yolov3.cfg', 'https://raw.githubusercontent.com/pjreddie/darknet/f86901f6177dfc6116360a13cc06ab680e0c86b0/cfg/yolov3.cfg', 'b969a43a848bbf26901643b833cfb96c') supported_layers = ['net', 'convolutional', 'shortcut', 'route', 'upsample'] parser = DarkNetParser(supported_layers) layer_configs = parser.parse_cfg_file(cfg_file_path) del parser output_tensor_dims = OrderedDict() output_tensor_dims['082_convolutional'] = [255, 19, 19] output_tensor_dims['094_convolutional'] = [255, 38, 38] output_tensor_dims['106_convolutional'] = [255, 76, 76] builder = GraphBuilderONNX(output_tensor_dims) weights_file_path = download_file( 'yolov3.weights', 'https://pjreddie.com/media/files/yolov3.weights', 'c84e5b99d0e52cd466ae710cadf6d84c') yolov3_model_def = builder.build_onnx_graph( layer_configs=layer_configs, weights_file_path=weights_file_path, verbose=True) del builder onnx.checker.check_model(yolov3_model_def) output_file_path = 'yolov3.onnx' onnx.save(yolov3_model_def, output_file_path)