net_decode.pt模型转onnx

[HandsLing]

请问能提供一些net_decode.pt模型转onnx的思路吗。我转onnx的时候报错
torch.onnx.errors.SymbolicValueError: Unsupported: ONNX export of instance_norm for unknown channel size. [Caused by the value 'input.5 defined in (%skip.1 : Tensor, %skip0.1 : Tensor, %skip1.1 : Tensor, %input.5 : Tensor = prim::ListUnpack(%x.3), scope: model_p2p.conv_blocks_dbb.Hourglass_DBB_Mobile_Decode::
)' (type 'Tensor') in the TorchScript graph. The containing node has kind 'prim::ListUnpack'.]

我转onnx的代码如下：
decoder_input1 = torch.randn(1, 3, 384, 384).to(self.device)
decoder_input2 = torch.randn(1, 16, 192, 192).to(self.device)
decoder_input3 = torch.randn(1, 32, 96, 96).to(self.device)
decoder_input4 = torch.randn(1, 64, 48, 48).to(self.device)
decoder_input = (decoder_input1, decoder_input2, decoder_input3, decoder_input4)
param_val = torch.randn(1, 32).to(self.device)
input_names =["decoder_input1", "decoder_input2", "decoder_input3", "decoder_input4", "param_val"]
output_names = ['decoder_output']

    torch.onnx.export(self.generator, 
                    (decoder_input, param_val),
                    f'{data_dir}/net_decode.onnx',
                    opset_version=11,
                    input_names = input_names,
                    output_names = output_names,
                    dynamic_axes={  "decoder_input1": {0: "batch_size"}, 
                                    "decoder_input2": {0: "batch_size"},
                                    "decoder_input3": {0: "batch_size"},
                                    "decoder_input4": {0: "batch_size"},
                                    "param_val": {0: "batch_size"},
                                    'decoder_output': {0: 'batch_size'}}
                    )