model:
  arch:
    type: SwinForImageClassification

  model_config:
    type: SwinConfig
    num_labels: 1000 # num classes
    image_size: 224 # input image size
    patch_size: 4 # patch size
    num_channels: 3 # channels of input images
    embed_dim: 128 # embedding dimension
    depths: [2, 2, 18, 2] # number of transformer blocks for each swin layer
    num_heads: [4, 8, 16, 32] # number of attention heads for each swin layer
    window_size: 7 # window size for swin
    mlp_ratio: 4 # ffn_hidden_size = mlp_ratio * embed_dim
    qkv_bias: True # has transformer qkv bias or not
    layer_norm_eps: 0.00001  # eps of layer_norm
    hidden_dropout_prob: 0. # drop rate of MLP
    attention_probs_dropout_prob: 0. # drop rate of Attention
    drop_path_rate: 0.1 # drop path rate of transformer blocks
    use_absolute_embeddings: False # if using absolute position embedding
    patch_norm: True # use norm in SwinPatchEmbeddings
    hidden_act: gelu # activation of MLP
    weight_init: normal # weight initialize type
    loss_type: SoftTargetCrossEntropy # loss type
    checkpoint_name_or_path: swin_base_p4w7

processor:
  type: SwinProcessor

  image_processor:
    type: SwinImageProcessor
    size: 224  # input image size