RFC: SameDiff op class creator class heirarchy

[AlexDBlack]

Something I've been considering doing recently is reorganizing the SameDiff class more along the lines of TF and PyTorch.
Currently all of the "op creator" methods are defined on the SameDiff class itself:
https://github.com/deeplearning4j/deeplearning4j/blob/master/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/SameDiff.java

Instead, we would have something like:

SameDiff sd = SameDiff.create();
...
SDVariable relu = sd.nn.relu(...);                //Currently: sd.relu(...)
SDVariable count = sd.math.countNonZero(...);     //Currently: sd.countNonZero(...)

The motivation is three-fold:
(a) In theory it should be easier for users to find relevant ops
(b) Ease of documentation (autogen docs by section)
(c) Maintenance - a lot of the op methods could be moved to other classes, making SameDiff.java more managable (it's 11,500 lines so far)

As for categories, here's a start:

• nn (neural network)
• math
• loss
• random
• image? (We do have stuff like resizeBilinear)
• linalg?

Some open questions:

• What should be on the "main" SameDiff namespace vs. the sub-categories?
• Should we have more than 1 level? Or allow things like sd.nn.rnn.LSTM(...)?
• Should there be similar organization for SDVariable methods?
• What categories should we use?

If this change is to be implemented, it should happen before the next release (and sooner if possible).

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.