HengLan / CGSTVG

[CVPR 2024] Context-Guided Spatio-Temporal Video Grounding
41 stars 2 forks source link

With each iteration, the model’s video memory gradually increases #2

Closed tongzhang111 closed 7 months ago

tongzhang111 commented 7 months ago

With each iteration, the model’s video memory gradually increases

GX77 commented 7 months ago

I didn't quite understand your question, but I think it's normal for the memory usage to gradually increase in the early stages of training.

tongzhang111 commented 7 months ago
WeChat4a73b46a7215f6446143ab556d4619d8

With each iteration, the max mem gradually increases

GX77 commented 7 months ago

I understand what you mean now, I didn't notice this situation before, I'm not quite sure about the reason, it might need careful debugging.

GX77 commented 7 months ago

One speculation is that the image size is dynamic, which would lead to a gradual increase in video memory.

tongzhang111 commented 7 months ago

As the number of iterations increases, the maximum memory also increases. Therefore, the problem of video memory explosion will occur. I determine whether there is a memory leak

GX77 commented 7 months ago

If the OOM (Out of Memory) error occurs, you may need to reduce the resolution.