Research on NNVM

[zakki]

For AMD GPU and OpenCL via ONNX. https://github.com/dmlc/nnvm

[trainewbie]

Perhaps, I think it is probably possible. You can use the method to extract the parameters into a separate file and load it. The reason I said "Perhaps".... In the case of RNN, the implementation method differs for each vendor, it may not be possible to extract the parameters. Note that cuDNN_LSTM and Tensorflow BasicLSTM are different implementation methods and should be studied again. (Of course, native cuDNN_LSTM is supported as a separate API in tensor flow.) - I write with some programmer's advice -