parseq/config.yaml

model:  
  image_size : [224, 224]
  patch_size : [16, 16]
  max_len : 25
  d_model : 384
  enc_num_heads : 6
  enc_mlp_ratio : 4
  enc_depth : 12
  dec_num_heads : 12
  dec_mlp_ratio : 4
  dec_depth : 1
  perm_num : 8
  perm_forward : true
  perm_mirrored : true
  decode_ar : true
  refine_iter : 2
  num_tokens : 97
  pretrained : false
  train_charset : 0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!\"#$%&'()*+,-./:;<=>?@[\\]^_`{|}~
  weight_url : https://github.com/baudm/parseq/releases/download/v1.0.0/parseq_small_patch16_224-fcf06f5a.pt
trainer:
  lr : 3e-4
  batch_size : 4
  weight_decay : 0.0
  warm_pct : 0.075