KellenSunderland commented on a change in pull request #13310: [MXNET-703] Update to TensorRT 5, ONNX IR 3. Fix inference bugs. URL: https://github.com/apache/incubator-mxnet/pull/13310#discussion_r236058512
########## File path: src/operator/contrib/nnvm_to_onnx.cc ########## @@ -248,8 +259,12 @@ void ConvertPooling(NodeProto* node_proto, const NodeAttrs& attrs, AttributeProto* const pads = node_proto->add_attribute(); pads->set_name("pads"); pads->set_type(AttributeProto::INTS); - for (int kval : pad) { - pads->add_ints(static_cast<int64>(kval)); + + // Convert from MXNet symetric pads to ONNX non-symetric by running through padding twice. + for (int i =0; i < 2; i++) { Review comment: Yeah I want to have a conversation with them offline, I've already brought this up a few times. The trick here is we need to export to ONNX from native code. This is more or less 100% duplication (minus tests). I'd be tempted to politely suggest we migrate all ONNX stuff to native code and core functionality so that for example I can export to ONNX, and java users running inference can import from ONNX. I think this feature is important enough it shouldn't be python only. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services