Any progress for pytorch 1.5? about deq

[jerrybai1995]

Yes, that is a design choice due to PyTorch's nn.DataParallel. If you use only 1 GPU (i.e., no nn.DataParallel), then you are able to do the actual implicit differentiation all in the backward() in deq.py and without func_copy. You can simply do it through one layer, as we hoped.

However, the weird thing we found was, once nn.DataParallel was invoked, the parameter gradients on the replica will all vanish. In other words, the gradients computed in the backward() will disappear. This happened in PyTorch 1.4, I'm not so sure about 1.5. But anyway, that was the rationale behind this design choice; we found no good choice but to leave a func_copy there for the Jacobian-vector product part computation.

Indeed, once we are able to solve this issue, we will have much better memory efficiency than the ones reported in our paper.

[jerrybai1995]

Yup, I have pushed a branch named "pytorch-1.5" for the repo. Please pull the repo, do git checkout pytorch-1.5 and train the model there. Also, see the updated README on what's been changed.

Let me know if it works!

[LuChengTHU]

Thanks! And I'm confused about the 'func_copy' model. It seems that we need to use 2x GPU memory because of this implementation. Is there a more efficient way of implementing the backward method?

[LuChengTHU]

Thanks! I'm waiting for the better implementation!