Bihaqo / t3f

Tensor Train decomposition on TensorFlow

https://t3f.readthedocs.io/en/latest/index.html

MIT License

218 stars 56 forks source link

Adding support of Eager and newer versions of Tensorflow #184

Closed aiboyko closed 4 years ago

aiboyko commented 4 years ago

1) The code crashes if one uses tensorflow above 1.11 2) The eager mode is not supported.

To summarize, one needs to learn old arcane syntax to somehow embed t3f into modern DL software stack. In order to use it one needs to do some homunculus code that runs (tf 1.11 + t3f) + (modern tensorflow or pytorch) on the same GPU.

@Bihaqo Could you please add TF 2.0 compatibility and Eager mode when you have time.

oseledets commented 4 years ago

You probably need a simple example to reproduce. I think, eager is partially supported. Also, there is an "upgrade script" that you can try running and help the community.

Bihaqo commented 4 years ago

Thanks for creating the issue!

As discussed offline, the eager mode works fine on the example you have in mind. As for the later versions of tf, that's a good point, I'll add tests for that.

Bihaqo commented 4 years ago

185

Bihaqo commented 4 years ago

It seems that tf 1.13.2 and tf 1.14 are fine even without any fixes. https://travis-ci.org/Bihaqo/t3f/builds/588457426?utm_source=github_status&utm_medium=notification

Bihaqo commented 4 years ago

185 is merged now and t3f should work with latest tf.