Pytorch next steps

[hhbyyh]

Use this issue as a reminder to track the Pytorch-related work items

1. Support Table/multiple input and output models. [done]

2. JNI memory optimization. Avoid mem leak and unnecessary copy.

3. Scalability test for training and inference

4. Support export to Pytorch model.

5. Utility to convert data loader to RDD/DataFrame

6. RNN/LSTM training.

7. Confirm tracing batch normalization works for training.

[hhbyyh]

8. Support transfer learning with Pytorch model, i.e. lock weights etc.

[hhbyyh]

9. Investigate the performance impact of short cut the backward process from loss to model.

[hhbyyh]

10. Docs
11. Better exception handling and error messages.