@lazarevevgeny Could you please take a look?

Just noticed that some constant_folding tests do pass negative indices to OneHot. This seems a little bit strange, since all negative indices will be statically casted to huge size_t numbers, and almost certainly ignored (at least for reasonable depth values, not smth like 2^30). I’m not sure if this negative-indices feature is needed, but if it does, I suggest to add smth like “When some elements from the indices are negative, or greater or equal to the depth it is a well-formed operation…”.
@lazarevevgeny, what do you think?

Just noticed that some constant_folding tests do pass negative indices to OneHot. This seems a little bit strange, since all negative indices will be statically casted to huge size_t numbers, and almost certainly ignored (at least for reasonable depth values, not smth like 2^30). I’m not sure if this negative-indices feature is needed, but if it does, I suggest to add smth like “When some elements from the indices are negative, or greater or equal to the depth it is a well-formed operation…”.
@lazarevevgeny, what do you think?

CC @dmitry-gorokhov, @nshchego, @andrei-kochin One more example with negative indices like with Gather operation.

@IvanNovoselov We preliminary agreed to that the behaviour with negative indices is undefined, because different plugins may work differently since we don't support negative indices.

@lazarevevgeny could you please clarify what this undefined behavior means from the implementation point of view? Specifically, I would suggest:

1. The ref. implementation should not throw on negative indices, but have to ignore them silently (since they are not supported).
2. Negative indices should be eliminated from the constant folding tests for one_hot, because undefined behavior should not be tested.
3. This behavior should be explicitly stated in the specs, smth like "Behavior for negative indices is undefined."

Please, correct me if I'm getting smth wrong.

@lazarevevgeny could you please clarify what this undefined behavior means from the implementation point of view? Specifically, I would suggest:

1. The ref. implementation should not throw on negative indices, but have to ignore them silently (since they are not supported).
2. Negative indices should be eliminated from the constant folding tests for one_hot, because undefined behavior should not be tested.
3. This behavior should be explicitly stated in the specs, smth like "Behavior for negative indices is undefined."

Please, correct me if I'm getting smth wrong.

Agree. Let's write 0 value to elements corresponding to negative elements.

@lazarevevgeny, please take a look at the corrected version, and approve the changes if everything is Ok.

Please, revert commit of the mkl-dnn.

Looks like you changed the mkl-dnn submodule by mistake.

@ilyachur, you are right. Fixed.