Added batch_gather to backend.

[phipleg]

Adds the method batched_gather(reference, indices which is a batched variant of gather where indices is a 1d tensor.

The numpy equivalent is reference[np.arange(batch_size), indices]. For example, if reference is [[3, 5, 7], [11, 13, 17]] and indices is [2, 1] then the result is [7, 13].

While the numpy syntax can be be directly translated to Theano, due to slicing restrictions this approach does not work in Tensorflow. Instead one can use tf.gather_nd but there is no direct support in Keras.

This extension would allow to implement the backward forward algorithm for the inference step of a Conditional Random Field with Keras (see PR #4621). It is the last necessary modification of the backend for implementing a CRF in Keras.

[fchollet]

This function is not present in either Theano or TensorFlow. This makes it a harder sell.

This extension would allow to implement the backward forward algorithm for the inference step of a Conditional Random Field with Keras

If this is the rationale, then we should package such an addition with a CRF layer (at a later time), rather than including the function in core Keras right now. We would rather not have backend functions that:

• are not used by anything in the Keras codebase
• are not a direct mirroring of the TF API

[phipleg]

Thank you very much for your answer, @fchollet. I completely agree and will try to find another solution for the batch_gather implementation without touching the backend.