add subgraph detection for convoluation with dilations > 1

[qigtang]

Tensorflow conv2d with dilations > 1 is in a graph with

SpaceToBatchND -> conv -> BatchtoSpaceND etc.

SpaceToBatchND and BatchtoSpaceND are not supported in onnx, should combine this subgraph to normal conv2d with dilation > 1 operator.

[guschmue]

not sure if I interpret this correctly:
you say that for example conv2d(dilations=[1,3,3,1]) will compose the graph as SpaceToBatchND -> conv -> BatchtoSpaceND since tf.conv2d has no native support conv2d with dilations ?

I think this has changed, just tested with 1.8 and it has only conv2d:
https://github.com/onnx/tensorflow-onnx/blob/master/tests/test_backend.py#L269

Can't find it in the tensorflow release notes, but I think I recall something after tf-1.5.

[qigtang]

Haven't tried recent latest tensorflow version. We decided to switch to pytorch for our training instead.

[sulc]

Hi,
I would also appreciate the support for SpaceToBatchND, and I believe this is still an issue with newer versions of TF:
I'm working with TF 1.12, and this opp is still used by tf.contrib.layers.conv2d with dilation (i.e. with parameter rate>1) - tf2onnx crashes with:
ValueError: tensorflow op SpaceToBatchND is not supported

Looking deeper into the layer definition, I can see _WithSpaceToBatch being used in Convolution in nn_ops.py..

[nbcsm]

SpaceToBatchND is supported via #472
please give a try with latest mater branch.

[sulc]

Thank you! My bad - indeed it's supported by the latest master - sorry, I missed this update..

Unfortunately, I'm getting another error related to SpaceToBatchND:
node model/Conv_1/SpaceToBatchND: exception get tensor value: model/Conv_1/concat must be Const
I suppose it cannot handle input with variable size..? (It's a detection network, where images may have different resolutions).
I'm using the latest onnx, but onnx.defs.has("DynamicSlice") is False (not sure if it's relevant)..

[vadimkantorov]

@guschmue This is still an issue for old TensorFlow checkpoints, like NVidia OpenSeq2Seq wav2letter model. Currently it exports a SpaceToBatchND + Conv + BatchToSpaceND subgraph (indeed with a dynamic crop issue #571, fix in #635 ?).

The corresponding TfLite transform is in https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/toco/graph_transformations/identify_dilated_conv.cc

[anmolduainter]

I want to convert tensorflow model (tf version: 1.12) where dilation rate > 1 in many convs into ONNX. I am unable to do so as there is Space To Batch error. Can anyone help me to resolve this issue. Desired ONNX version (1.3.0).