search

ninatu / everything_at_once

Official implementation of "Everything at Once - Multi-modal Fusion Transformer for Video Retrieval". CVPR 2022
96 stars 18 forks source link

What hardware resources do I need to run your model and data? #5

Closed GeoffreyWu8 closed 2 years ago

GeoffreyWu8 commented 2 years ago

Hi Shvedova, Congratulations on your team's paper being accepted by CVPR. I have a slight question. What hardware resources do you need to run your model and data? Because I have three 2080Ti and it still doesn't work. The message is below:

RuntimeError: CUDA out of memory. Tried to allocate 96.00 MiB (GPU 0; 10.76 GiB total capacity; 9.81 GiB already allocated; >>88.31 MiB free; 9.82 GiB reserved in total by PyTorch) The command that I used is : 'python test.py --n_gpu 3 --config configs/evaluation/msrvtt_at_once.yaml --resume pretrained_models/everything_at_once_tva/latest_model.pth' Looking forward to your reply. Best wishes, Feng.

ninatu commented 2 years ago

Hi!

We used one V100 32GB GPU for testing. But indeed model takes about ~ 13Gb gpu memory. And most of it is just loaded weights in checkpoint (model, optimizer, etc) You can change the line 

checkpoint = torch.load(config.resume)

to 

checkpoint = torch.load(config.resume, map_location='cpu')

And gpu utilization becomes 5Gb. I guess it should work for one 2080Ti.