search

tensorflow / tensor2tensor

Library of deep learning models and datasets designed to make deep learning more accessible and accelerate ML research.
Apache License 2.0
15.56k stars 3.51k forks source link

when will t2t support tensorflow 2.0 #1771

Open caleb1010 opened 4 years ago

caleb1010 commented 4 years ago

Hi, there! I'm a new comer for t2t and I find some t2t codes are still using tf 1.x modules which are deprecated in 2.0 such tf.flags.

Just wondering is there a plan to upgrade to 2.0 in near future. 🙂

ksbsk commented 4 years ago

still not working... how about backward compatibility code?

lhk commented 4 years ago

same here. I would like to use t2t, but am held back by the lack of support for tensorflow 2.

the code to generate datasets relies on the tf.gfile API which was reworked in tf2.

martinpopel commented 4 years ago

It seems the authors of T2T don't wish to upgrade it to TF2, so the answer is "never". See Trax as an TF2-compatible alternative to T2T. If you want to use T2T, just install TF1 in your virtual environment, it is simple.

lhk commented 4 years ago

oh that's unfortunate.

It might be just my personal impression, but I think the tensorflow ecosystem is much too volatile. They constantly create and abandon new libraries.

To establish a technology, it would be good for the community to have a solid base to build upon. I use tensorflow for both low level AI research, training custom architectures, and for inference in bigger software architectures. Tensorflow 2 was a massive change in paradigm and required huge amounts of refactoring in my models (ok, arguably v2 has also brought some advantages). And on the inferencing side, there's just no proper ecosystem of pretrained models.