Pair-wise losses fail when providing dynamically sized logits

[tensorflow]

Pair-wise losses fail when providing dynamically sized logits

[xuanhuiwang]

Can you try topn = array_ops.shape(logits)[1] to see whether it solves your problem?

[felixmaximilian]

I can confirm that using array_ops.shape does not have the same issue as the variant in the code which uses unstack(). It would be great if your code would be adapted to use the more reliable way to get the shape.
I was also able to temporarily fix the issue by giving a hint for the shape to the logits tensor:

self.logits.set_shape(target_array.shape)
for some reason, the target_array tensor shape can be inferred although being a placeholder with uninitialized shape:
self.target_array = tf.placeholder(tf.int32, [None, None], name='targets')

I consider this issue as resolved. Hopefully, it will be considered in one of the next releases.

[xuanhuiwang]

We will use array index to replace unstack. closing this now.