Hi all and thanks if you read this
I have a special case of transformer where the input of the decoder is a list, and one of the element of the list is a list of 2D Tensors (size of list=batch size) and the ouput is a tuple(List Tensors)
I’m able to convert my model in torchscript via script method and wrap the model
Unfortinately i don’t know howto expose this kind of input/ouput to torch.onnx.export
torch.onnx.export(
decoder,
(output_encoder,), <----- Ouput Encoder is a list of 2d Tensors (with a dynamic dimentions)
"decoder.onnx",
input_names=["input"],
output_names=["output"],
dynamic_axes={"input": {0: {0: "sequence"}}, "output": {0: "sequence"}}, <---- fail, howto define this use case ?
)
Regards