Source of design choices in vit_l_dp005_mask_005 config?

Hi team,

I had a question regarding the vit_l_dp005_mask_005 config in the InsightFace repo. I noticed that it deviates from the standard Vision Transformer (ViT) implementation in several ways:

1. Random masking of input tokens.

2. No use of [CLS] token for classification.

3. The final output of the ViT encoder is flattened and passed directly to the classification head.

4. Learnable positional encodings for input.

5. Learnable mask tokens to join the unmasked to tokens of output.

Could you please point me to the research paper or documentation that these design decisions are based on? I’d like to understand the motivation or theoretical grounding behind these modifications.

Thanks in advance!