How to convert from pytorch to tensorflow==2.4.1

[chenyuqi990215]

I try to convert model from torch==1.6.0 to tensorflow==2.4.1 using onnx-tf==1.7.0.

My code is as follow:

    def torchToTensorflow(self, model):
        input = torch.randn([100, *image_size(DATASET_NAME)])
        # 设置输入张量名，多个输入就是多个名
        input_names = ["input"]
        # 设置输出张量名
        output_names = ["output"]
        # 自定义onnx文件名和路径
        onnx_filename = "model.onnx"
        # 执行转化和保存
        torch.onnx.export(model, input, onnx_filename, verbose=True, input_names=input_names,
                          output_names=output_names)
        onnx_model = onnx.load("model.onnx")  # load onnx model
        tf_exp = prepare(onnx_model)  # prepare tf representation
        tf_exp.export_graph("model.pb")  # export the model
        print(tf_exp.tensor_dict)
        with tf.Graph().as_default():
            output_graph_def = tf.compat.v1.GraphDef()
            output_graph_path = "model.pb/saved_model.pb"

            with open(output_graph_path, 'rb') as f:
                output_graph_def.ParseFromString(f.read())
                _ = tf.import_graph_def(output_graph_def, name="")
        return output_graph_def

However I get following bug:

  File "architecture_test.py", line 396, in torchToTensorflow
    output_graph_def.ParseFromString(f.read())
google.protobuf.message.DecodeError: Error parsing message

Could you help me deal with the problem?

[chinhuang007]

The exported TF model is in the Saved Model format, so the pb file cannot be loaded as a TF computational graph.

[gunerhui]

The exported TF model is in the Saved Model format, so the pb file cannot be loaded as a TF computational graph.

how can i use it

[chenyuqi990215]

Thank you, I have solved the problem.

[farazBhatti]

@chenyuqi990215 can you share the solution?