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
1.36k stars 135 forks source link

Why do not load the best ckpt for final test during downstream finetuning? #68

Closed youcaiSUN closed 1 year ago

youcaiSUN commented 1 year ago

Hi, It seems that the model in the last epoch will be used for test, why not use the best saved ckpt during validation? https://github.com/MCG-NJU/VideoMAE/blob/826dfa6d4fa54da21fe470bc0471a4de83f3acd1/run_class_finetuning.py#L542

Thanks!

congee524 commented 1 year ago

There is little difference between the last ckpt and the best ckpt unless overfit. But you may modify the code to reload the best-ckpt.

youcaiSUN commented 1 year ago

Got it, thanks!