visionlan_resnet45_LF_1.yaml

system:
  mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
  distribute: True
  amp_level: 'O2'
  seed: 42
  log_interval: 200
  val_while_train: True
  drop_overflow_update: False

common:
  character_dict_path: &character_dict_path
  num_classes: &num_classes 37 # num_chars_in_dict+1,  TODO: retreive it from dict or check correctness
  max_text_len: &max_text_len 25
  infer_mode: &infer_mode False
  use_space_char: &use_space_char False
  batch_size: &batch_size 192
  blank_at_last: &blank_at_last False # pad blank at the last or the first of the dictionary
  lower: &lower True
  training_step: &training_step 'LF_1'

model:
  type: rec
  transform: null
  resume: False
  backbone:
    name: rec_resnet45
    pretrained: False
    strides: [2, 2, 2, 1, 1]

  head:
    name: VisionLANHead
    n_layers: 3
    n_position: 256
    n_dim: 512
    max_text_length:  *max_text_len
    training_step: *training_step


postprocess:
  name: VisionLANPostProcess
  character_dict_path: *character_dict_path
  use_space_char: *use_space_char
  blank_at_last: *blank_at_last
  lower: *lower
  max_text_length: *max_text_len

metric:
  name: RecMetric
  main_indicator: acc
  character_dict_path: *character_dict_path
  ignore_space: True
  print_flag: False

loss:
  name: VisionLANLoss
  mode: *training_step
  weight_res: 0.5
  weight_mas: 0.5

scheduler:
  scheduler: step_decay
  min_lr: 0.0
  lr: &init_lr 0.0001
  decay_rate: 0.1
  decay_epochs: 6
  num_epochs: 8
  warmup_epochs: 0

optimizer:
  opt: adam
  weight_decay: 0.0
  nesterov: False
  # lr: # no need to set lr here because in train.py, the lr scheduler is passed to create_optimizer
  grouping_strategy: visionlan
  training_step: *training_step

loss_scaler:
  type: dynamic
  loss_scale: 512
  scale_factor: 2.0
  scale_window: 1000

train:
  ema: True
  ema_decay: 0.9999
  clip_norm: 20.0
  ckpt_save_dir: './tmp_visionlan/LF_1'
  dataset_sink_mode: False
  pred_cast_fp32: False
  dataset:
    type: LMDBDataset
    dataset_root: ./datasets # Optional, if set, dataset_root will be used as a prefix for data_dir
    data_dir: train
    # label_file: # not required when using LMDBDataset
    sample_ratio: 1.0
    shuffle: True
    filter_max_len: True
    max_text_len: *max_text_len
    transform_pipeline:
      - DecodeImage:
          img_mode: BGR
          to_float32: False
      - VisionLANLabelEncode:
          max_text_len: *max_text_len
          character_dict_path: *character_dict_path
          use_space_char: *use_space_char
          blank_at_last: *blank_at_last
          lower: *lower
      - SVTRRecAug:
          aug_type: 0
          deterioration_factor: null
      - SVTRRecResizeImg:
          image_shape: [64, 256] # H, W
          padding: False
      - NormalizeImage:
          bgr_to_rgb: True
          is_hwc: True
          mean: [127.0, 127.0, 127.0]
          std: [127.0, 127.0, 127.0]
      - ToCHWImage:
    #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visaulize
    output_columns: ['image', 'label', 'label_res', 'label_sub', 'label_id', 'length']
    net_input_column_index: [0, 4] # input indices for network forward func in output_columns
    label_column_index: [1, 2, 3, 5] # input indices marked as label

  loader:
      shuffle: True # TODO: tbc
      batch_size: *batch_size
      drop_remainder: True
      max_rowsize: 12
      num_workers: 8

eval:
  ckpt_load_path: ./tmp_visionlan/LF_1/best.ckpt
  dataset_sink_mode: False
  dataset:
    type: LMDBDataset
    dataset_root: ./datasets
    data_dir: evaluation/Sumof6benchmarks
    # label_file: # not required when using LMDBDataset
    sample_ratio: 1.0
    shuffle: False
    transform_pipeline:
      - DecodeImage:
          img_mode: BGR
          to_float32: False
      - VisionLANLabelEncode:
          max_text_len: *max_text_len
          character_dict_path: *character_dict_path
          use_space_char: *use_space_char
          blank_at_last: *blank_at_last
          lower: *lower
      - SVTRRecResizeImg:
          image_shape: [64, 256] # H, W
          padding: False
      - NormalizeImage:
          bgr_to_rgb: True
          is_hwc: True
          mean: [127.0, 127.0, 127.0]
          std: [127.0, 127.0, 127.0]
      - ToCHWImage:
    #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visaulize
    output_columns: ['image', 'text_padded', 'length']
    net_input_column_index: [0] # input indices for network forward func in output_columns
    label_column_index: [1, 2]


  loader:
      shuffle: False # TODO: tbc
      batch_size: 64
      drop_remainder: False
      max_rowsize: 12
      num_workers: 8