Closed Park-ing-lot closed 1 year ago
Did you use mixed precision training? If it still does not work, you can try decreasing batch size which should not affect the performance too much.
When I try to change 'draw_false_video' config to 1, it works. I'll try the mixed precision training and batch size decreasing. Thanks!
Hey! When I try to fine-tune TVLT on audio-video retrieval tasks such as MSR-VTT, I encountered the cuda memory error with A6000 (49G) even though 4 batch sizes per GPU. (I didn't change anything from your code.)
The paper says all experiments are conducted with A6000, so can you tell me the exact settings?