src/models/vit/config.yml

model:
  # deit
  deit_tiny_distilled_patch16_224:
    image_size: 224
    patch_size: 16
    d_model: 192
    n_heads: 3
    n_layers: 12
    normalization: deit
    distilled: true
  deit_small_distilled_patch16_224:
    image_size: 224
    patch_size: 16
    d_model: 384
    n_heads: 6
    n_layers: 12
    normalization: deit
    distilled: true
  deit_base_distilled_patch16_224:
    image_size: 224
    patch_size: 16
    d_model: 768
    n_heads: 12
    n_layers: 12
    normalization: deit
    distilled: true
  deit_base_distilled_patch16_384:
    image_size: 384
    patch_size: 16
    d_model: 768
    n_heads: 12
    n_layers: 12
    normalization: deit
    distilled: true
  # vit
  vit_base_patch8_384:
    image_size: 384
    patch_size: 8
    d_model: 768
    n_heads: 12
    n_layers: 12
    normalization: vit
    distilled: false
  vit_tiny_patch16_384:
    image_size: 384
    patch_size: 16
    d_model: 192
    n_heads: 3
    n_layers: 12
    normalization: vit
    distilled: false
  vit_small_patch16_384:
    image_size: 384
    patch_size: 16
    d_model: 384
    n_heads: 6
    n_layers: 12
    normalization: vit
    distilled: false
  vit_base_patch16_384:
    image_size: 384
    patch_size: 16
    d_model: 768
    n_heads: 12
    n_layers: 12
    normalization: vit
    distilled: false
  vit_large_patch16_384:
    image_size: 384
    patch_size: 16
    d_model: 1024
    n_heads: 16
    n_layers: 24
    normalization: vit
  vit_small_patch32_384:
    image_size: 384
    patch_size: 32
    d_model: 384
    n_heads: 6
    n_layers: 12
    normalization: vit
    distilled: false
  vit_base_patch32_384:
    image_size: 384
    patch_size: 32
    d_model: 768
    n_heads: 12
    n_layers: 12
    normalization: vit
  vit_large_patch32_384:
    image_size: 384
    patch_size: 32
    d_model: 1024
    n_heads: 16
    n_layers: 24
    normalization: vit
decoder:
  linear: {}
  deeplab_dec:
    encoder_layer: -1
  mask_transformer:
    drop_path_rate: 0.0
    dropout: 0.1
    n_layers: 2
dataset:
  ade20k:
    epochs: 64
    eval_freq: 2
    batch_size: 8
    learning_rate: 0.001
    im_size: 512
    crop_size: 512
    window_size: 512
    window_stride: 512
  pascal_context:
    epochs: 256
    eval_freq: 8
    batch_size: 16
    learning_rate: 0.001
    im_size: 520
    crop_size: 480
    window_size: 480
    window_stride: 320
  cityscapes:
    epochs: 216
    eval_freq: 4
    batch_size: 8
    learning_rate: 0.01
    im_size: 1024
    crop_size: 768
    window_size: 768
    window_stride: 512