saurabhshri/CCAligner

Changes for TF 1.13.1 and CUDA 10.0

aolney opened this issue · 4 comments

Getting error AttributeError: module 'tensorflow.contrib.rnn' has no attribute 'core_rnn_cell'

My TF is 1.13.1 and CUDA is 10.1

Guessing this is b/c of incompatibility b/w my version of TF and the presumed 1.0.0 version. However, I do not want to downgrade my CUDA from 10 to 8 just to use TF 1.0.0.

To fix this, I:

  1. Made the minor code change suggested in this closed PR
  2. Downloaded the current seq2seq, unzipped into the dependencies directory, and installed
  3. Recompiled using build.sh as described in the README
  4. Replaced the install/g2p-seq2seq-cmudict directory with one matching the seq2seq version

This seems to be working with my configuration. Since the aforementioned PR was declined, I'm documenting this only as an issue.

Im having the same issue now.
I'm going to try with your solution. Hope It works.

I'd followed your instructions, and now it is working for me @aolney. Thanks for the Guidance.

@aolney Thank you very much for looking into this. For some reason, I can't recall, we decided to not use the recent ones. But I see no reason right now why we should not support the latest dependencies.

Unfortunately, I am not able to get enough time to look into this (hence, my superlate reply to this).

Feel free to open a PR with changes that would work with latest libs, or re-open the one I closed.

Thanks again for documenting this.

Unfortunately I blew my code away later in the summer to free up disk space. So to do this would require complete recreation.

@joseamair Would you consider a PR?