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
1.38k stars 136 forks source link

About the encoder layer output #104

Open Shar-01 opened 1 year ago

Shar-01 commented 1 year ago

As understood, the MCG-NJU/videomae-base encoder outputs tensor of size (batch, 1568,768). Can we interpret this as 1567 patch tokens with emb size of 768 and the first token as the cls token? If so, for further downstream tasks, can we only take the first (cls token)and discard the rest 1567 patch tokens? I.e., Can we just take (batch, 1, 768) for further downstream tasks? Can someone say if they tried this for further downstream tasks and how much information is lost/retained compared to taking the whole 1568 tokens? (if someone compared it).