Losses are backpropagated separately about daformer

[lhoyer]

Dear @HKQX,

Thank you for your interest in our work. Doing separate backward passes for different inputs (source images and mixed images) allows freeing the first compute graph after the first backward, which reduces the overall GPU memory consumption compared to one backward pass of the accumulated loss. This is a similar concept to gradient accumulation:

• https://pytorch-lightning.readthedocs.io/en/stable/advanced/training_tricks.html#accumulate-gradients
• https://kozodoi.me/python/deep%20learning/pytorch/tutorial/2021/02/19/gradient-accumulation.html#:~:text=Gradient%20accumulation%20modifies%20the%20last,been%20processed%20by%20the%20model.

Best,
Lukas

[HKQX]

Dear @HKQX,

Thank you for your interest in our work. Doing separate backward passes for different inputs (source images and mixed images) allows freeing the first compute graph after the first backward, which reduces the overall GPU memory consumption compared to one backward pass of the accumulated loss. This is a similar concept to gradient accumulation:

• https://pytorch-lightning.readthedocs.io/en/stable/advanced/training_tricks.html#accumulate-gradients
• https://kozodoi.me/python/deep%20learning/pytorch/tutorial/2021/02/19/gradient-accumulation.html#:~:text=Gradient%20accumulation%20modifies%20the%20last,been%20processed%20by%20the%20model.

Best, Lukas

Thank you for your reply.When I run the code, I find that the data of src.loss_imnet_feat_dist is nan. The reason for nan is that during the calculation of Thing-Class ImageNet Feature Distance (FD), this image does not contain these classes [6, 7, 11, 12, 13, 14, 15, 16, 17, 18]. Have you noticed this problem? And the memory usage is close to 12G, 2080ti should not be able to complete the code running,do you need to adjust anything? Looking forward to your reply

[lhoyer]

Please, have a look at issue #11 regarding nan in the FD loss. I am able to run this repository on an RTX 2080 Ti. It's tight but it fits.