svtr_ppocrv3_ch.yaml

system:
  mode: 0 # 0 for graph mode, 1 for pynative mode in MindSpore
  distribute: True
  amp_level: O2
  amp_level_infer: O2 # running inference in O2 mode
  seed: 42
  log_interval: 10
  val_while_train: True
  drop_overflow_update: True
  ckpt_max_keep: 5

common:
  character_dict_path: &character_dict_path mindocr/utils/dict/ch_dict.txt
  num_classes: &num_classes 6624 # num_chars_in_dict + 1
  max_text_len: &max_text_len 25
  use_space_char: &use_space_char True
  batch_size: &batch_size 128

model:
  type: rec
  transform: null
  backbone:
    name: mobilenet_v1_enhance
    scale: 0.5
    last_conv_stride: [ 1, 2 ]
    last_pool_type: avg
    last_pool_kernel_size: [ 2, 2 ]
    pretrained: False
  head:
    name: MultiHead
    out_channels_list:
      - CTCLabelDecode: 6625
      - SARLabelDecode: 6627
    head_list:
      - CTCHead:
          Neck:
            name: svtr
          out_channels: *num_classes
      - SARHead:
          enc_dim: 512
          max_text_length: *max_text_len
  pretrained: https://download-mindspore.osinfra.cn/toolkits/mindocr/svtr/svtr_lcnet_ppocrv3-6c1d0085.ckpt

postprocess:
  name: CTCLabelDecode
  character_dict_path: *character_dict_path
  use_space_char: *use_space_char

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

loss:
  name: MultiLoss
  loss_config_list:
    - CTCLossForSVTR:
    - SARLoss:

scheduler:
  scheduler: warmup_cosine_decay
  min_lr: 0.000001
  lr: 0.0001
  num_epochs: 500
  warmup_epochs: 0
  decay_epochs: 500

optimizer:
  opt: Adam
  beta1: 0.9
  beta2: 0.999
  weight_decay: 3.0e-05

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

train:
  ema: True
  ckpt_save_dir: ./tmp_rec
  dataset_sink_mode: False
  clip_grad: True
  clip_norm: 0.1
  dataset:
    type: RecDataset
    dataset_root: dir/to/data/
    data_dir: training/
    label_file: gt_training.txt
    sample_ratio: 1.0
    shuffle: True
    extra_data_num: 2
    filter_max_len: True
    max_text_len: *max_text_len
    transform_pipeline:
      - DecodeImage:
          img_mode: BGR
          to_float32: False
      - RecAug:
      - MultiLabelEncode:
          character_dict_path: *character_dict_path
          max_text_length: *max_text_len
          use_space_char: *use_space_char
      - RecResizeImgForSVTR:
          image_shape: [3, 48, 320]
    output_columns: ["image", "label_ctc", "label_sar", "text_length", "valid_ratio", "valid_width_mask"]
    net_input_column_index: [0, 1, 2, 3, 4, 5]
    label_column_index: [1, 2, 3, 4]

  loader:
    shuffle: True
    batch_size: *batch_size
    drop_remainder: False
    max_rowsize: 12
    num_workers: 1

eval:
  ckpt_load_path: ./tmp_rec/best.ckpt
  dataset_sink_mode: False
  dataset:
    type: RecDataset
    dataset_root: dir/to/data/
    data_dir: validation/
    label_file: gt_validation.txt
    sample_ratio: 1.0
    shuffle: False
    transform_pipeline:
      - DecodeImage:
          img_mode: BGR
          to_float32: False
      - MultiLabelEncode:
          character_dict_path: *character_dict_path
          max_text_length: *max_text_len
          use_space_char: *use_space_char
      - RecResizeImgForSVTR:
          image_shape: [ 3, 48, 320 ]
          width_downsample_ratio: 0.125
    output_columns: ["image", "text_padded", "text_length"]
    net_input_column_index: [ 0 ]
    label_column_index: [ 1, 2 ]

  loader:
    shuffle: False
    batch_size: 128
    drop_remainder: False
    max_rowsize: 12
    num_workers: 1