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_r236131449
########## File path: src/operator/contrib/nnvm_to_onnx.cc ########## @@ -78,8 +84,13 @@ op::TRTParam ConvertNnvmGraphToOnnx( ModelProto model_proto; // Need to determine IR versions and features to support - model_proto.set_ir_version(static_cast<int64>(2)); + auto opset_proto = model_proto.add_opset_import(); + opset_proto->set_version(static_cast<int64>(8)); Review comment: No worries. What you're asking is clear. So the underlying library (onnx-tensorrt) moved to this version and opset soon after migrating to TRT5. Almost immediately we began seeing serious correctness regressions with the old onnx version and opset, so updating them to a supported version seems like the path of least resistance to enable TRT5 support. After changing the opset and updating the library the output is exactly as we'd expect (i.e. matching MXNets output). So basically we want to update onnx-tensorrt to fix bugs and support new hardware, to do so we need to serialize onnx in exactly this format. ---------------------------------------------------------------- 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
