search

MCG-NJU / VideoMAE

[NeurIPS 2022 Spotlight] VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training
https://arxiv.org/abs/2203.12602
Other
1.32k stars 133 forks source link

Adding accumlate_iter. #54

Closed Zi-hao-Wei closed 2 years ago

Zi-hao-Wei commented 2 years ago

I add a accumlate_iter in the pretrain process. This may help we have a same effecitve batchsize as the original one VideoMAE used with less GPUs. I should say 64 GPUs are not usually affordable.

yztongzhan commented 2 years ago

Hi @Zi-hao-Wei! Thanks for your PR! I know adding accumlate_iter may help us pre-train VideoMAE with fewer GPUs, but it could lead to lower performance (less than 1%). We will consider supporting 8 GPUs pre-training until we find a more stable way.

Zi-hao-Wei commented 2 years ago

Looking forward to seeing the official 8 GPU pre-training code.