ICCV-2021 An Empirical Study of Training Self-Supervised Vision Transformers
Paper: An Empirical Study of Training Self-Supervised Vision Transformers
MoCo v3: Freezing the patch projection makes training smoother
- Self-supervised ViT training is unstable; larger batches ought to help, but in practice training gets worse. They inspect gradient backpropagation layer by layer and find that when training yields a large loss, gradients spike; the spike occurs in the first layer—the patch projection.
- Without modifying the transformer itself, only the beginning or the end can be changed. This highlights the importance of the first step—the tokenization stage—
<HR align=left color=#987cb9 SIZE=1>