Masked Autoencoders Enable Efficient Knowledge Distillers
share
This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, , forward propagate inputs through the first few layers, for obtaining intermediate feature maps. Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves 84.0 outperforming the baseline of directly distilling a fine-tuned ViT-L by 1.2 More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with 95 where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of 83.6 still secure 82.4 FOUR visible patches (98 available at https://github.com/UCSC-VLAA/DMAE.
READ FULL TEXT