Libnd4j: embedding_lookup op needs to support multiple input arrays

[AlexDBlack]

Currently, the embedding_lookup op assumes that all data is in a single array.
i.e., we have exactly 2 arrays: input and indices.

deeplearning4j/libnd4j/include/ops/declarable/generic/parity_ops/embedding_lookup.cpp

Lines 37 to 38 in 451dd76

	NDArray<T>* input = INPUT_VARIABLE(0); // lookup param
	NDArray<T>* indeces = INPUT_VARIABLE(1); // indeces, as is

For TF import, we need this op (or a new op, if that's easier) to support multiple input arrays - i.e., a "split" input array (still with single indices array), with partitioning strategies "mod" and "div":
https://www.tensorflow.org/api_docs/python/tf/nn/embedding_lookup

[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.