FastQA model tensor IndexedSlices warning

[undwie]

when calling reader.train:

/Users/bloomsbury/virtual_envs/responder_venv/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "

Often these aren't really a problem, but they still leave the user thinking that there is a problem. Can this one be fixed?

[dirkweissenborn]

No, not really. They happen when we backprop through a gather op.

[riedelcastro]

Okay. Can you point me to this? Depending on the type of gather, these can actually be replaced by a somewhat equivalent operation that doesn't raise the warning. I have done this with some other code in the NTP at some point.

[dirkweissenborn]

search for "gather" in [1]. I don't know how easy it will be to change that. Can we not just suppress these warnings?

[1] https://github.com/uclmr/jtr/blob/master/jtr/tasks/xqa/fastqa.py

[undwie]

It's easy to fix, but very difficult to suppress these warnings (while keeping the ones we want to see).

On Fri, Sep 15, 2017 at 10:42 PM Dirk Weissenborn [email protected] wrote:

search for "gather" in [1]. I don't know how easy it will be to change that. Can we not just suppress these warnings?

[1] https://github.com/uclmr/jtr/blob/master/jtr/tasks/xqa/fastqa.py

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/uclmr/jtr/issues/241#issuecomment-329910994, or mute the thread https://github.com/notifications/unsubscribe-auth/AdMriICwlpqYgub5C7JT_Ei3MIKkO_CFks5siu9LgaJpZM4PX1lv .