configs/rec/cam/convnextv2_tiny_cam_tps_on.yml

Global:
  device: gpu
  epoch_num: 20
  log_smooth_window: 20
  print_batch_step: 10
  output_dir: ./output/rec/u14m_filter/convnextv2_tiny_cam_tps_on
  eval_epoch_step: [0, 1]
  eval_batch_step: [0, 500]
  cal_metric_during_train: False
  pretrained_model:
  checkpoints:
  use_tensorboard: false
  infer_img:
  # for data or label process
  character_dict_path: ./tools/utils/EN_symbol_dict.txt
  max_text_length: &max_text_length 25
  use_space_char: False
  save_res_path: ./output/rec/u14m_filter/predicts_convnextv2_cam_tps_on.txt
  use_amp: True

Optimizer:
  name: AdamW
  lr: 0.0008 # for 4gpus bs256/gpu
  weight_decay: 0.05
  filter_bias_and_bn: True
  eps: 1.e-8

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

Architecture:
  model_type: rec
  algorithm: CAM
  Transform:
    name: Aster_TPS
    tps_inputsize: [32, 64]
    tps_outputsize: &img_shape [32, 128]
  Encoder:
    name: CAMEncoder
    encoder_config:
      name: ConvNeXtV2
      depths: [3, 3, 9, 3]
      dims: [96, 192, 384, 768]
      strides: [[4,4], [2,1], [2,1], [1,1]]
      drop_path_rate: 0.2
      feat2d: True
    nb_classes: 97
    strides: [[4,4], [2,1], [2,1], [1,1]]
    deform_stride: 2
    stage_idx: 2
    use_depthwise_unet: True
    use_more_unet: False
    binary_loss_type: BanlanceMultiClassCrossEntropyLoss
    mid_size: False
    d_embedding: 512
  Decoder:
    name: CAMDecoder
    num_encoder_layers: -1
    beam_size: 0
    num_decoder_layers: 2
    nhead: 8
    max_len: *max_text_length

Loss:
  name: CAMLoss
  loss_weight_binary: 1.5
  label_smoothing: 0.

Metric:
  name: RecMetric
  main_indicator: acc
  is_filter: True

PostProcess:
  name: ARLabelDecode

Train:
  dataset:
    name: LMDBDataSet
    data_dir: ../Union14M-L-LMDB-Filtered
    transforms:
      - DecodeImagePIL: # load image
          img_mode: RGB
      - PARSeqAugPIL:
      - CAMLabelEncode: # Class handling label
          font_path: ./arial.ttf
          image_shape: *img_shape
      - RecTVResize:
          image_shape: [64, 256]
          padding: False
      - KeepKeys:
          keep_keys: ['image', 'label', 'length', 'binary_mask'] # dataloader will return list in this order
  loader:
    shuffle: True
    batch_size_per_card: 256
    drop_last: True
    num_workers: 4

Eval:
  dataset:
    name: LMDBDataSet
    data_dir: ../evaluation
    transforms:
      - DecodeImagePIL: # load image
          img_mode: RGB
      - ARLabelEncode: # Class handling label
      - RecTVResize:
          image_shape: [64, 256]
          padding: False
      - KeepKeys:
          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 256
    num_workers: 2