Adam Pocock

I don't think that operation is meant to be used directly, it's added to graphs when you apply gradients to them via an `Optimizer`.

Probably not, the LSTM in Keras is quite complicated and only calls the op we expose at the very end. I've not traced through the computation properly to see what...

What do you expect `tf.create` to do? As far as I can tell there isn't an equivalent method in TF Python. Depending on what your input looks like you can...

I think we'll need both. Keras will want the former as then you can freeze a layer so it's not trainable (which is very useful for transfer learning), and non-Keras...

You should compare the C API gradient code (https://github.com/tensorflow/tensorflow/blob/5dcfc51118817f27fad5246812d83e5dccdc5f72/tensorflow/cc/framework/gradients.cc#L438) to the Python TF v1 gradient code (https://github.com/tensorflow/tensorflow/blob/1bb4e4fce0c992e2d995e1663138ad98e957e050/tensorflow/python/ops/gradients_impl.py#L44). The latter is full of locks and control dependencies, and the former has...

I agree the C API version is completely different, but the fact that it's lacking the locking and dependencies is the reason it doesn't produce the same output as the...

When you ran it after upgrading cuDNN it ran in Python? TF-Java doesn't have a python VM running, so what error message do you get out of TF-Java after cuDNN...

Looks like you should be able to go directly to a TF Saved Model with the converter, but I don't think the TFJS file format is documented so we couldn't...

Unfortunately not. At the moment it looks like tfjs uses something Keras-like as it's JSON format, but we don't have all the Python bits necessary to load in Keras models...

The quick build failing in javadoc generation is a known issue we're working on in other PRs.