Pytorch模型转Tensorflow

1.将pytorch pt模型转onnx

def pttoonnx():
    model = torch.jit.load("./weights/crowdhuman_yolov5m.pt")
    img = torch.rand((1,3,640,640))
    torch.onnx.export(model, img, "./weights/yolov5.onnx")

2.将onnx模型转Tensorflow格式pb

def onnxtopb():
    onnx_model = onnx.load("./weights/yolov5.onnx")
    tf_rep = prepare(onnx_model, device="CPU")
    tf_rep.export_graph("./weights/yolov5")

yolov5模型onnx转pb的时候，出现一个错误，原因是转换算子版本太高，onnx_tf版本不支持，需要降低本版，torch.onnx.export(model, img, "./weights/yolov5.onnx", opset_version=12)，将opset_version设置为12

BackendIsNotSupposedToImplementIt: Unsqueeze version 13 is not implemented.

导出onnx模型时，降低版本

def pttoonnx():
    model = torch.jit.load("./weights/crowdhuman_yolov5m.pt")
    img = torch.rand((1,3,640,640))
    torch.onnx.export(model, img, "./weights/yolov5.onnx", opset_version=12)