Changing the names and cleaning hparams_sets of the Universal Transformer based on the NIPS submission. #837
Conversation
…er and Adaptive Universal Transformer)
Hi @MostafaDehghani, thanks for the PR. Please update such that the Travis tests pass. |
Hi @afrozenator , Thanks for your reply. I believe this is because of the problem with tf.foldl in public current version of the Tensorflow which doesn't support nested/multiple tensors for input and initialization. When I was at Google, I checked in a CL for changing the foldl and foldr to support nested tensors and it should be available internally but seems it's not publicly available yet (in Tensorflow 1.8). |
You need to update the name in the |
Also you can see the Travis test logs by clicking "Details" next to the check so that you're not flying blind. |
@MostafaDehghani Is there a kind of preprint available of that paper? I'm really interested in (I made some experiments with the |
…ansformer_test.py until tf.foldl is updated in tensorflow 1.9
Thanks @rsepassi. Solved the problem. |
@stefan-it, glad that you tried it :) We submitted the paper to NIPS and we're going to submit a preprint to arXiv soon (next week or so). In the meantime, let me know if you have any question. |
Thanks for fixing that @MostafaDehghani and @rsepassi for figuring it out. Ryan was showing me around the GitHub workflow :) @MostafaDehghani - I was trying to automatically sync this internally, and if I'm not mistaken you haven't signed the CLA yet? -- you just have to do this once. Head over to https://cla.developers.google.com/ to see your current agreements on file or to sign a new one. If you have already signed a CLA then maybe my automatic import is broken, let me know -- it would be useful to me -- I'll then sync in your changes the painful way :) |
Hi @afrozenator, Great! I have signed the Google Individual CLA on Nov 16, 2017 and I can see it in https://cla.developers.google.com/. So this might be a problem on your side. |
.travis.yml
Outdated
@@ -54,14 +54,13 @@ script: | |||
# * visualization_test | |||
# * model_rl_experiment_test | |||
# algorithmic_math_test: flaky | |||
# r_transformer_test: requires new feature in tf.foldl (rm with TF 1.9) | |||
- pytest |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hi @MostafaDehghani -- Can you put this "- pytest" line back ... I've worked out how to import this automatically internally.
Thanks again for your patience as I continue to work on the tooling!
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Done! :)
Hi @MostafaDehghani - One last thing -- Travis seems to be unhappy because of the pylint errors in the newly added files, can you fix those? Copy/pasting for convenience:
|
…ensor2tensor" This reverts commit d016ab3, reversing changes made to 41434d4.
Hi @afrozenator, it should be ready to go! let me know if there's any other problem. |
Hi @MostafaDehghani -- Thanks for fixing the linting errors. Travis still seems to be complaining, I've restarted it, I'm not sure if it is related to your PR. Let me investigate. Lets get this in today. Thanks! |
Hi @MostafaDehghani -- I guess your commits are on top of an older version of T2T -- Could you 'sync to head'? (I don't even know what it is called in git parlance). Meanwhile I'm running presubmit on my internal patch of this PR. Thanks! |
to update, should be able to run on your branch:
|
…er and Adaptive Universal Transformer)
…ansformer_test.py until tf.foldl is updated in tensorflow 1.9
…ensor2tensor" This reverts commit d016ab3, reversing changes made to 41434d4.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This seems to be 'reverting' a bunch of changes instead of applying those in this PR -- @MostafaDehghani can you check?
For example: Head ISSUE_TEMPLATE - https://raw.githubusercontent.com/tensorflow/tensor2tensor/master/ISSUE_TEMPLATE.md vs yours https://github.com/tensorflow/tensor2tensor/pull/837/files#diff-0b6c1416b1293274c4e2d11fb49e1795
A good sanity check is that the number of files changed should be the 5 that you changed |
@afrozenator yep.. sorry for that! |
No problem @MostafaDehghani -- just let me know when you are satisfied |
…hghani/tensor2tensor"" This reverts commit 06eb36d.
This reverts commit a787228.
@afrozenator fixed the "reverts" I've done mistakenly. |
@MostafaDehghani : @rsepassi is the expert on git here :) I'll let him comment. Travis seems to still have some issues, because the parameters on encode and decode were changed. This should be simple to fix, right? IIUC you are just deleting the arguments, it maybe possible to leave them as is and then add Thanks
|
Awesome @MostafaDehghani -- merging now. |
PiperOrigin-RevId: 198719318
Great...Thanks a lot @afrozenator :) |
Thank you @MostafaDehghani for the PR and the persistence :) Many regards |
…rmer based on the NIPS submission. (tensorflow#837) * Changing the names and cleaning hparams_sets based on the NIPS submission * Adding some ranges for haparams_set (for both the Universal Transformer and Adaptive Universal Transformer) * Updating name of r_transofmer in .travis.yml --to ignore universal_transformer_test.py until tf.foldl is updated in tensorflow 1.9 * Resolving the .travis.yml conflict. * Update .travis.yml * fixing lint errors * Revert "Merge branch 'master' of https://github.com/MostafaDehghani/tensor2tensor" This reverts commit d016ab3, reversing changes made to 41434d4. * Revert "Merge branch 'master' into master" This reverts commit d885757, reversing changes made to 42981f3. * Changing the names and cleaning hparams_sets based on the NIPS submission * Adding some ranges for haparams_set (for both the Universal Transformer and Adaptive Universal Transformer) * Updating name of r_transofmer in .travis.yml --to ignore universal_transformer_test.py until tf.foldl is updated in tensorflow 1.9 * fixing lint errors * Revert "Merge branch 'master' of https://github.com/MostafaDehghani/tensor2tensor" This reverts commit d016ab3, reversing changes made to 41434d4. * Revert "Merge branch 'master' into master" This reverts commit d885757, reversing changes made to 42981f3. * Revert "merge text_encoder" This reverts commit 008ff4c, reversing changes made to a787228. * Revert "Revert "Merge branch 'master' of https://github.com/MostafaDehghani/tensor2tensor"" This reverts commit 06eb36d. * Revert "Revert "Merge branch 'master' into master"" This reverts commit a787228. * fixing arguments of encode and decode to match the overridden methods
PiperOrigin-RevId: 198719318
Hey Lukasz and Ryan,
Based on the submitted paper at NIPS I changed the names used in the code to Universal Transformer (instead of the Recurrent Transformer) and also updated the names of hparams_set based on the submission and removed those that are not useful.
Cheers,