sxjscience edited a comment on issue #18256: URL: https://github.com/apache/incubator-mxnet/issues/18256#issuecomment-625050992
@leezu Previously, we have one benchmark by @Jerryzcn : https://github.com/apache/incubator-mxnet/issues/17335 . According to my recent coding experience with the new DeepNumpy interface of MXNet, I do recognize that the usability of Gluon HybridBlock is worse than the PyTorch Module. When implementing some slightly more complicated structures like TransformerXL or Transformer for NMT, it's easy to write a PyTorch-based implementation that can fully utilize the GPUs. On the other hand, it is not so trivial to do so in MXNet. One observation is that the MXNet program will occupy lots of CPU cores. I'm still trying to figure out the cause. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to 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