TensorFlow r2.7 branch cut, September 23

Hi - we would like to announce that a new TF release 2.7.0 is coming up and we will be cutting the r2.7 branch on 9/23/2021 end of day. You can refer to TF 2.7.0 GitHub RELEASE.md document for the current plans for what’s upcoming, and will be updated with changes as we prepare the final release.

A key focus for this release is Performance and Usability, and we look forward to delivering these improvements in the final release in late October.

6 Likes

Hi!

I work at Rasa, and we’ve been eager to update our tensorflow dependency. We’re currently blocked by a significant increases in training time, which we’ve narrowed down to the tf.scan function. See related issue here.

It’d be great if you could prioritise this as part of your Performance and Usability focus!

Please let us know if we can help at all.

1 Like

Hi - thank you for the follow up and the detailed issue, that really helps. It seems this is an issue with training on MacOS Mojave Version 10.14.5? Have you reached out to Apple on GitHub as well? They can probably advise you more quickly on any macOS dependencies related issues.

Hi,

Thanks for your quick response!

Actually I don’t believe this is a macOS issue: we did originally notice this issue on macOS but we’ve since also reproduced this on debian-10 Linux, and someone from your team has reproduced this on Google colab (see here)

1 Like

I see that now, thanks. Let me try to escalate this as a performance regression issue. I do agree with the comments about needing to simplify the repro.

2 Likes

Thank you!

needing to simplify the repro

My colleague just posted a simpler example, hopefully that helps.

1 Like

Hi, just a quick update to say that we are working to land updates that address this performance issue in the upcoming TF2.7 release.

1 Like

@yarri-oss thank you for keeping us posted! Very glad to hear it :slight_smile:

1 Like

@yarri-oss I have a question about

“Improve the informativeness of error messages raised by Keras Layer.__call__(), by adding the full list of argument values passed to the layer in every exception.”.

Does this mean that we’ll have access to additional arguments in the call()? If so, is there the list of these arguments somewhere?

TF 2.7-RC0 has been uploaded! You can try it with pip install tensorflow==2.7.0rc0.

Currently we have not yet uploaded the CPU-only wheels so tensorflow-cpu==2.7.0rc0 is still not working. We’re going to get this fixed by next week.

1 Like

pip install tensorflow-cpu==2.7.0rc0 is now also possible

@SmacznaKawusia sorry for the delay in responding, could you repost this in the General Discussion forum? I would recommend adding the ‘keras’ and ‘help_request’ tags as well. Thanks!

@yarri-oss Can you share what is the current target of final release date - do you still aim for late October?

1 Like

We should have RC1 released by early next week and then we’ll analyze depending on the number of reported regressions

TF 2.7 RC1 has just been released. You can test it with one of the next commands

pip install tensorflow==2.7.0rc1
pip install tensorflow-cpu==2.7.0rc1
pip install tensorflow-gpu==2.7.0rc1

Just as in all other releases after 2.1, tensorflow-gpu is the same as tensorflow on Windows and Linux whereas tensorflow-cpu is the same as tensorfow on MacOS (no GPU build available on MacOS)