Skip to content

ViT

ViT backbone based on An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.

ViT (Vision Transformer) does not have a stage configuration and therefore does not support compatibility with neck modules. Currently, it only supports the FC head. When using the ViT model for classification tasks, users can decide whether to use a classification token. Additionally, users can flexibly configure the settings of the transformer encoder.

Field list

Field Description
name (str) Name must be "vit" to use ViT backbone.
params.patch_size (int) Size of the image patch to be treated as a single embedding.
params.attention_channels (int) Dimension for the encoder.
params.num_blocks (int) The number of self-attention blocks in the encoder.
params.num_attention_heads (int) The number of heads in the multi-head attention.
params.attention_dropout_prob (float) Dropout probability in the attention block.
params.ffn_intermediate_channels (int) Intermediate dimension of the feed-forward network inside the attention block.
params.ffn_dropout_prob (float) Dropout probability in the feed-forward network inside the attention block.
params.use_cls_token (bool) Whether to use the classification token.
params.vocab_size (int) Maximum token length for positional encoding.

Model configuration examples

ViT-tiny 
model:
  architecture:
    backbone:
      name: vit
      params:
        patch_size: 16
        attention_channels: 192
        num_blocks: 12
        num_attention_heads: 3
        attention_dropout_prob: 0.0
        ffn_intermediate_channels: 768  # hidden_size * 4
        ffn_dropout_prob: 0.1
        use_cls_token: True
        vocab_size: 1000
      stage_params: ~
ViT-small 
model:
  architecture:
    backbone:
      name: vit
      params:
        patch_size: 16
        attention_channels: 384
        num_blocks: 12
        num_attention_heads: 6
        attention_dropout_prob: 0.0
        ffn_intermediate_channels: 1536  # hidden_size * 4
        ffn_dropout_prob: 0.0
        use_cls_token: True
        vocab_size: 1000
      stage_params: ~