configs/rec/mgpstr/vit_mgpstr.yml

Global:
  device: gpu
  epoch_num: 20
  log_smooth_window: 20
  print_batch_step: 10
  output_dir: ./output/rec/u14m_filter/vit_mgpstr/
  eval_epoch_step: [0, 1]
  eval_batch_step: [100000, 2000]
  cal_metric_during_train: False
  pretrained_model:
  checkpoints:
  use_tensorboard: false
  infer_img:
  # for data or label process
  character_dict_path: &character_dict_path ./tools/utils/EN_symbol_dict.txt
  max_text_length: &max_text_length 25
  use_space_char: &use_space_char False
  use_amp: True
  save_res_path: ./output/rec/u14m_filter/predicts_vit_mgpstr.txt
  grad_clip_val: 5

Optimizer:
  name: Adam
  lr: 0.000325 # 4gpus 128bs/gpu
  weight_decay: 0.
  filter_bias_and_bn: False

LRScheduler:
  name: OneCycleLR
  warmup_epoch: 1.5 # pct_start 0.075*20 = 1.5ep
  cycle_momentum: False

Architecture:
  model_type: rec
  algorithm: MGPSTR
  Transform:
  Encoder:
    name: ViT
    img_size: [32,128]
    patch_size: [4, 4]
    embed_dim: 384
    depth: 12
    num_heads: 6
    mlp_ratio: 4
    qkv_bias: True
  Decoder:
    name: MGPDecoder
    only_char: &only_char False

Loss:
  name: MGPLoss
  only_char: *only_char

PostProcess:
  name: MPGLabelDecode
  character_dict_path: *character_dict_path
  use_space_char: *use_space_char
  only_char: *only_char

Metric:
  name: RecMPGMetric
  main_indicator: acc
  is_filter: True

Train:
  dataset:
    name: LMDBDataSet
    data_dir: ../Union14M-L-LMDB-Filtered
    transforms:
      - DecodeImagePIL: # load image
          img_mode: RGB
      - PARSeqAugPIL:
      - MGPLabelEncode: # Class handling label
          character_dict_path: *character_dict_path
          use_space_char: *use_space_char
          max_text_length: *max_text_length
          only_char: *only_char
      - RecTVResize:
          image_shape: [32, 128]
          padding: False
      - KeepKeys:
          keep_keys: ['image', 'char_label', 'bpe_label', 'wp_label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: True
    batch_size_per_card: 128
    drop_last: True
    num_workers: 4

Eval:
  dataset:
    name: LMDBDataSet
    data_dir: ../evaluation/
    transforms:
      - DecodeImagePIL: # load image
          img_mode: RGB
      - MGPLabelEncode: # Class handling label
          character_dict_path: *character_dict_path
          use_space_char: *use_space_char
          max_text_length: *max_text_length
          only_char: *only_char
      - RecTVResize:
          image_shape: [32, 128]
          padding: False
      - KeepKeys:
          keep_keys: ['image', 'char_label', 'bpe_label', 'wp_label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 256
    num_workers: 2