search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
415 stars 77 forks source link

Can we run evaluate script on multiple GPUS? #96

Closed bhosalems closed 1 year ago

bhosalems commented 1 year ago

When I ran evaluate ion single GPU it gave me out of memory CUDA error (for torch.cuda.synchronize()). So I added one more GPU in the default config, and it gives me the below issue now.

File "actionformer_release/libs/utils/train_utils.py", line 391, in valid_one_epoch output = model(video_list) return forward_call(*input, **kwargs) raise RuntimeError("module must have its parameters and buffers " RuntimeError: module must have its parameters and buffers on device cuda:3 (device_ids[0]) but found one of them on device: cpu

tzzcl commented 1 year ago

The current evaluation batch size is one. i.e, using multiple GPUs won't have some memory reduction. Since mixed/half precision will hurt the model's performance, you can try CPU inference.