configs/rec/svtrv2/svtrv2_smtr_gtc_rctc_ch.yml

Global:
  device: gpu
  epoch_num: 100
  log_smooth_window: 20
  print_batch_step: 10
  output_dir: ./output/rec/ch/svtrv2_smtr_gtc_rctc_ch
  save_epoch_step: 1
  # evaluation is run every 2000 iterations
  eval_batch_step: [0, 2000]
  eval_epoch_step: [0, 1]
  cal_metric_during_train: False
  pretrained_model:
  # ./output/rec/u14m_filter/svtrv2_rctc/best.pth
  checkpoints:
  use_tensorboard: false
  infer_img:
  # for data or label process
  character_dict_path: &character_dict_path ./tools/utils/ppocr_keys_v1.txt # ch
  max_text_length: &max_text_length 25
  use_space_char: &use_space_char False
  save_res_path: ./output/rec/ch/predicts_svtrv2_smtr_gtc_rctc_ch.txt
  use_amp: True

Optimizer:
  name: AdamW
  lr: 0.00065 # for 4gpus bs256/gpu
  weight_decay: 0.05
  filter_bias_and_bn: True

LRScheduler:
  name: OneCycleLR
  warmup_epoch: 5
  cycle_momentum: False

Architecture:
  model_type: rec
  algorithm: SVTRv2
  in_channels: 3
  Transform:
  Encoder:
    name: SVTRv2LNConvTwo33
    use_pos_embed: False
    dims: [128, 256, 384]
    depths: [6, 6, 6]
    num_heads: [4, 8, 12]
    mixer: [['Conv','Conv','Conv','Conv','Conv','Conv'],['Conv','Conv','FGlobal','Global','Global','Global'],['Global','Global','Global','Global','Global','Global']]
    local_k: [[5, 5], [5, 5], [-1, -1]]
    sub_k: [[1, 1], [2, 1], [-1, -1]]
    last_stage: false
    feat2d: True
  Decoder:
    name: GTCDecoder
    infer_gtc: True
    detach: False
    gtc_decoder:
      name: SMTRDecoder
      num_layer: 1
      ds: True
      max_len: *max_text_length
      next_mode: &next True
      sub_str_len: &subsl 5
    ctc_decoder:
      name: RCTCDecoder

Loss:
  name: GTCLoss
  ctc_weight: 0.1
  gtc_loss:
    name: SMTRLoss

PostProcess:
  name: GTCLabelDecode
  gtc_label_decode:
    name: SMTRLabelDecode
    next_mode: *next
  character_dict_path: *character_dict_path
  use_space_char: *use_space_char

Metric:
  name: RecGTCMetric
  main_indicator: acc
  # is_filter: True

Train:
  dataset:
    name: RatioDataSetTVResize
    ds_width: True
    padding: false
    data_dir_list: ['../benchmark_bctr/benchmark_bctr_train/document_train',
    '../benchmark_bctr/benchmark_bctr_train/handwriting_train',
    '../benchmark_bctr/benchmark_bctr_train/scene_train',
    '../benchmark_bctr/benchmark_bctr_train/web_train'
    ]
    transforms:
      - DecodeImagePIL: # load image
          img_mode: RGB
      - PARSeqAugPIL:
      - GTCLabelEncode: # Class handling label
          gtc_label_encode:
            name: SMTRLabelEncode
            sub_str_len: *subsl
          character_dict_path: *character_dict_path
          use_space_char: *use_space_char
          max_text_length: *max_text_length
      - KeepKeys:
          keep_keys: ['image', 'label', 'label_subs', 'label_next', 'length_subs',
          'label_subs_pre', 'label_next_pre', 'length_subs_pre', 'length', 'ctc_label', 'ctc_length'] # dataloader will return list in this order
  sampler:
    name: RatioSampler
    scales: [[128, 32]] # w, h
    # divide_factor: to ensure the width and height dimensions can be devided by downsampling multiple
    first_bs: &bs 256
    fix_bs: false
    divided_factor: [4, 16] # w, h
    is_training: True
  loader:
    shuffle: True
    batch_size_per_card: *bs
    drop_last: True
    max_ratio: &max_ratio 8
    num_workers: 4

Eval:
  dataset:
    name: RatioDataSetTVResize
    ds_width: True
    padding: False
    data_dir_list: ['../benchmark_bctr/benchmark_bctr_test/scene_test']
    transforms:
      - DecodeImagePIL: # load image
          img_mode: RGB
      - GTCLabelEncode: # Class handling label
          gtc_label_encode:
            name: ARLabelEncode
          character_dict_path: *character_dict_path
          use_space_char: *use_space_char
          max_text_length: *max_text_length
      - KeepKeys:
          keep_keys: ['image', 'label', 'length', 'ctc_label', 'ctc_length'] # dataloader will return list in this order
  sampler:
    name: RatioSampler
    scales: [[128, 32]] # w, h
    # divide_factor: to ensure the width and height dimensions can be devided by downsampling multiple
    first_bs: *bs
    fix_bs: false
    divided_factor: [4, 16] # w, h
    is_training: False
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: *bs
    max_ratio: *max_ratio
    num_workers: 4