reminisce commented on issue #14040: Reformat of TensorRT to use subgraph API URL: https://github.com/apache/incubator-mxnet/pull/14040#issuecomment-460792134 Great to see this is happening. I have two high-level comments: 1. If you use the subgraph API, there should be no needs to add specific Python functions (e.g. `init_tensorrt_params`) to use TensorRT as a backend inference engine. I think everything can be handled in backend to support Module and simple_bind APIs in the frontend. 2. Have you considered adopting TensorRT C++ APIs to convert MXNet subgraph IR to TensorRT IR? This would allow you to get rid of dependency of protobuf, onnx, and onnx-trt. I used to find building MXNet with those third-party libs is a painful process, especially on edge devices. In addition, simply relying on the TensorRT C++ APIs allows we to extend operator coverage promptly once new release of TensorRT is out. I did this while integrating TensorRT with TVM, which shares the same graph IR as in MXNet. https://github.com/reminisce/tvm/blob/subgraph_integration/src/contrib/subgraph/tensorrt_executor.cc
---------------------------------------------------------------- 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
