CUDA memory efficiency

[Doan-IT]

• Thanks for sharing a great project. I tested offline and had a problem with Cuda out of memory.
• When I looked at the demo.ipynb notebooks tutorial, I saw that you put the whole video on GPU: video=video.cuda() and CoTrackerThreeOffline() class forward processing with fmaps_chunk_size=200. This process is expensive for GPU and inefficient.
• I want to ask if it is possible to split video into frames in CPU and process each frame in turn on GPU?
• I am not sure if the logic of model but such processing will reduce GPU memory a lot and be more accessible.
  Thanks.

[Haozong-Zeng]

Same issue here. I was trying to run the demo on a 640X480 video with 3000 frames, and it took 57 VRAM (20G + shared memory) and thus ran very slow. My project involves processing usually longer videos.

[gjpblabla]

Hello, have you solved this problem? I was wondering if I could change the batchsize, but I don't know how to do it