libnd4j: where_np shape function needs proper implementation

[AlexDBlack]

https://github.com/deeplearning4j/deeplearning4j/blob/master/libnd4j/include/ops/declarable/generic/boolean/where_np.cpp#L121-L145

Currently hard-coded to [10,10] output.
I believe it should have shape [numMatches, rank(input)], where numMatches is the number of non-zero elements.

[shugeo]

Numpy behave the next way:

• if only one param used (condition matrix), then "If only condition is given, return the tuple condition.nonzero(), the indices where condition is True."
• if two params uses - error occurs (ValueError: either both or neither of x and y should be given).

[shugeo]

The TF behave as follow:

A Tensor with the same type and shape as x, y if they are non-None. A Tensor with shape (num_true, dim_size(condition)).

the Where op behave as TF.

[AlexDBlack]

Right, that was my conclusion. The op implementation seems correct, but the shape function is wrong.

[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.