BatchToSpaceND: crops parameter need not be const?

[vadimkantorov]

I'm trying to convert the OpenSeq2Seq's wav2letter model checkpoints from TensorFlow to ONNX.

It contains BatchToSpaceND ops. In this model the crops argument is Identity, not Const.

However, tfonnx requires crops to be Const.

Failed to convert node import/ForwardPass/w2l_encoder/conv71/SpaceToBatchND
OP=SpaceToDepth
Name=import/ForwardPass/w2l_encoder/conv71/SpaceToBatchND
Inputs:
        Dropout__40:0=Dropout, [64, -1, 768], 10
        import/ForwardPass/w2l_encoder/conv71/SpaceToBatchND/block_shape:0=Const, [1], 6
        import/ForwardPass/w2l_encoder/conv71/concat:0=Identity, [1, 2], 6
Outpus:
        import/ForwardPass/w2l_encoder/conv71/SpaceToBatchND:0=[128, -1, 768], 10
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/tfonnx.py", line 574, in tensorflow_onnx_mapping
    func(g, node, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/onnx_opset/tensor.py", line 915, in version_4
    paddings = node.inputs[2].get_tensor_value()
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/graph.py", line 239, in get_tensor_value
    raise ValueError("get tensor value: {} must be Const".format(self.name))
ValueError: get tensor value: import/ForwardPass/w2l_encoder/conv71/concat must be Const
Failed to convert node import/ForwardPass/w2l_encoder/conv71/BatchToSpaceND
OP=DepthToSpace
Name=import/ForwardPass/w2l_encoder/conv71/BatchToSpaceND
Inputs:
        import/ForwardPass/w2l_encoder/conv71/conv1d/Squeeze:0=Squeeze, [128, -1, 896], 10
        import/ForwardPass/w2l_encoder/conv71/BatchToSpaceND/block_shape:0=Const, [1], 6
        import/ForwardPass/w2l_encoder/conv71/concat_1:0=Identity, [1, 2], 6
Outpus:
        import/ForwardPass/w2l_encoder/conv71/BatchToSpaceND:0=[64, -1, 896], 10
Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/tfonnx.py", line 574, in tensorflow_onnx_mapping
    func(g, node, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/onnx_opset/tensor.py", line 859, in version_4
    cls._convert_since_4(ctx, node, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/onnx_opset/tensor.py", line 871, in _convert_since_4
    crops = node.inputs[2].get_tensor_value()
  File "/usr/local/lib/python3.6/dist-packages/tf2onnx/graph.py", line 239, in get_tensor_value
    raise ValueError("get tensor value: {} must be Const".format(self.name))
ValueError: get tensor value: import/ForwardPass/w2l_encoder/conv71/concat_1 must be Const

` `

[zhijxu-MS]

onnx padhttps://github.com/onnx/onnx/blob/master/docs/Operators.md#Pad need padding to an attribute, namely a const.

@vadimkantorov the input of identity is a const or not? if yes, you could hack the code to specify its value, also you could wait for our tensorflow_const_folding feature.

[vadimkantorov]

I think it is not const. TensorFlow exports it as applying a shape op and then doing a much of slices and manipulation to it.

[zhijxu-MS]

sorry, I located the wrong code line, there is nothing about pad.

just looked at the conversion logic of BatchToSpaceND, it looks like we need enhace it, will go back to you later.

[vadimkantorov]

Then there is another question of folding the BatchToSpace + Conv + SpaceToBatch TensorFlow's pattern to a dilated conv.

[vadimkantorov]

Fix incoming in #635 ?

[fatcat-z]

The related PR was closed because of the lack of CLA, close it for now.