group_vit_gcc_redcap_30e.yml

_base_: 'default.yml'

model:
  type: MultiLabelContrastive
  img_encoder:
    type: GroupViT
    embed_dim: 384
    num_heads: [6, 6, 6]
    depths: [6, 3, 3]
    num_group_tokens: [64, 8, 0]
    num_output_groups: [64, 8]
    drop_rate: 0.0
    drop_path_rate: 0.1
  text_encoder:
    type: TextTransformer
    context_length: 77
    width: 256
    layers: 12
    vocab_size: 49408
  contrast_temperature: 0.07
  proj_num_layers: 2
  output_dim: 256
  multi_label: ${data.text_aug.multi_label} # multi_label=0 is better for RedCap
data:
  dataset:
    train:
      - gcc3m
      - gcc12m
      - redcap12m