rmrao / msa-transformer Goto Github PK

First of all, great work!

I am experienceing the train hangs problem, I have the same conda environment as yours and use 8 V100 GPUs and set train_batch_size =1.

I tried DDP training of single machine with 8cards, but I often encountered the phenomenon that 100% of the training of the 2 or 5 epoch was stuck. I checked the checkpoint after each epoch and found that it was saved normally. I tried to update the latest PyTorch lighting and sometimes this worked, but sometimes it stucks agian during the training. Like this:

I change another dataset, but it still hangs during the first epoch. Like this:

And I notice that when training stucks, the GPU utilization is 100%, but GPU memory utilization is 0%.

Have you ever encountered this situation during training?

It will be great if you can give me some suggestion! Thank you!

Could you please provide some insights on how MSA Transformer handles input MSAs with varying depths and lengths? For instance, if I have an input MSA with 2 sequences and a length of 500, does the model automatically apply padding to make it 1024 by 1024 before inference?

Moreover, how many subsampled MSAs were prepared for each MSA during training? Is it possible for the same protein sequence to appear in different subsampled training MSAs?

Thank you in advance for your help!