training is slow about gcvit HOT 5 CLOSED

Hi @ggsonic

That's an interesting observation. We primarily use the timm library to facilitate training, but did not notice this issue with dataloader.

Would you please provide more information regarding your setup which is used for training ?

Best,

from gcvit.

CPU:AuthenticAMD AMD EPYC 7763 64-Core Processor . GPU: A100. OS: Ubuntu 20.04 CUDA: cuda_11.3 . pytorch: 1.10.2+cu113 . timm 0.5.4.
we are using SLURM . 8 A100 GPU , each task with num_workers =8 . you can see in the snapshot only 8 cpu processes 100% working(we are using 8 A100 GPU not 4 as in your paper ).

from gcvit.

Hi @ggsonic

Thank you for sharing this information. We trained the model using 4 computational nodes, as specified in the paper. However, each node comprises of 8 GPUs, hence a total of 32 GPUs for this task. Using batch size of 128, training finished in around 22 hours on NVIDIA's NGC cluster.

According to your snapshot, I believe the issue CPU is clearly the bottleneck-- the A100 GPUs are consuming data in a higher rate and the CPU cores have a hard time to catch up. The timm library, which is primarily used by our work, has done a great job to address these bottlenecks in the dataloader, but it is still challenging for blazing fast hardware such as A100.

Best,

from gcvit.

it was my fault. i changed the code for using slurm,and set wrong create_loader params. Then i set distributed=True and everything is ok now. Thanks!

from gcvit.

Hi @ggsonic ,

Thanks for letting us know. In addition, we have updated our model and provide checkpoint with improved performance.

Best

from gcvit.