C. Antonio Sánchez

@tilakrayal what are you attempting to accomplish here? You have an input tensor `input0`, and a mask tensor `input1`, which is a different shape from `input0` and does not consist...

@eeDigitalSeal you're combining keras and raw tf here. Keras is a separate project, over at https://github.com/keras-team/keras. It's the keras Softmax API that is being used incorrectly. But my question stands:...

@eeDigitalSeal The issue here is that `NaN`s _are_ actually encountered in the forward pass due to your choice of mean and min/max vals. You'll see `NaN`s generated in backward gradients...

> Is there an RPC API for graph building? Go could use that. Not at the moment, but we may consider it - although it will be some time out.

I believe the index dimensions should be ``` indices = [[1,1]] ``` which allows this to work.

@VictoriaGriffith yes, we're trying to decide what to do about that. Thanks for bringing this up. We have the same "issue" across all back-ends: that `gather_nd` supports a single index...

Can you add a test for split too?

@olipinski which test is failing in your local install? For your PR, gradients are only supported for floating-point types, so you'll need to modify the `testGradientsAll` test to something like:...

@olipinski sorry for the delay here. Looks like you're missing a registration of `SplitVOpGPULaunch`. Register it in `split_lib_gpu.cu.cc`, and add a corresponding `extern template struct` declaration in `split_lib_gpu.h`.

@olipinski why do you need `int8` support? We can easily add it, but doing so increases TF's already massive binary size. For this reason, we try to keep registrations minimal....