PaddleOCR自定义训练集初次记录

使用MobieNetV3预训练检测模型

开始训练检测模型

python tools/train.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model=G:/PaddleOCRv2/pretrain_models/MobieNetV3/MobileNetV3_large_x0_5_pretrained.pdparams

检测效果评估指标评估

python tools/eval.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml  -o Global.checkpoints="output/ch_db_mv3/best_accuracy"

训练参数设置

//ch_det_mv3_db_v2.0.yml
Global:
  use_gpu: true
  epoch_num: 400
  log_smooth_window: 20
  print_batch_step: 1
  save_model_dir: ./output/ch_db_mv3/
  save_epoch_step: 400
  # evaluation is run every 2000 iterations after the 3000th iteration
  eval_batch_step: [2000, 3000]
  cal_metric_during_train: False
  pretrained_model: ./pretrain_models/MobileNetV3
  checkpoints:
  save_inference_dir:
  use_visualdl: False
  infer_img:
  save_res_path: ./output/det_db/predicts_db.txt

Architecture:
  model_type: det
  algorithm: DB
  Transform:
  Backbone:
    name: MobileNetV3
    scale: 0.5
    model_name: large
    disable_se: True
  Neck:
    name: DBFPN
    out_channels: 96
  Head:
    name: DBHead
    k: 50

Loss:
  name: DBLoss
  balance_loss: true
  main_loss_type: DiceLoss
  alpha: 5
  beta: 10
  ohem_ratio: 3

Optimizer:
  name: Adam
  beta1: 0.9
  beta2: 0.999
  lr:
    name: Cosine
    learning_rate: 0.001
    warmup_epoch: 2
  regularizer:
    name: 'L2'
    factor: 0

PostProcess:
  name: DBPostProcess
  thresh: 0.3
  box_thresh: 0.6
  max_candidates: 1000
  unclip_ratio: 1.5

Metric:
  name: DetMetric
  main_indicator: hmean

Train:
  dataset:
    name: SimpleDataSet
    data_dir: ./train_data/PS/
    label_file_list:
      - ./train_data/PS/Label35.txt
    ratio_list: [1.0]
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - DetLabelEncode: # Class handling label
      - IaaAugment:
          augmenter_args:
            - { 'type': Fliplr, 'args': { 'p': 0.5 } }
            - { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
            - { 'type': Resize, 'args': { 'size': [0.5, 3] } }
      - EastRandomCropData:
          size: [960, 960]
          max_tries: 50
          keep_ratio: true
      - MakeBorderMap:
          shrink_ratio: 0.4
          thresh_min: 0.3
          thresh_max: 0.7
      - MakeShrinkMap:
          shrink_ratio: 0.4
          min_text_size: 8
      - NormalizeImage:
          scale: 1./255.
          mean: [0.485, 0.456, 0.406]
          std: [0.229, 0.224, 0.225]
          order: 'hwc'
      - ToCHWImage:
      - KeepKeys:
          keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
  loader:
    shuffle: True
    drop_last: False
    batch_size_per_card: 1 //电脑性能较差设置为1
    num_workers: 1 //同上

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./train_data/PS/
    label_file_list:
      - ./train_data/PS/Label35.txt
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - DetLabelEncode: # Class handling label
      - DetResizeForTest:
#           image_shape: [736, 1280]
      - NormalizeImage:
          scale: 1./255.
          mean: [0.485, 0.456, 0.406]
          std: [0.229, 0.224, 0.225]
          order: 'hwc'
      - ToCHWImage:
      - KeepKeys:
          keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 1 # must be 1
    num_workers: 1

训练了5个epoch后评估效果（发现训练集较少，可能存在过拟合）

[2021/12/01 09:06:02] root INFO: metric eval ***************
[2021/12/01 09:06:02] root INFO: precision:0.9127682026028843
[2021/12/01 09:06:02] root INFO: recall:0.8644237175216523
[2021/12/01 09:06:02] root INFO: hmean:0.8879384088964928
[2021/12/01 09:06:02] root INFO: fps:5.15016150411071

使用ch_ppocr_server_v2.0_rec_pre预训练识别模型

开始训练识别模型

python tools/train.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml

检测效果评估指标评估

# GPU 评估， Global.checkpoints 为待测权重
python -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_common_train_v2.0.yml -o Global.checkpoints=output/rec_chinese_common_v2.0/best_accuracy

训练参数设置

//rec_chinese_common_train_v2.0.yml
Global:
  use_gpu: true
  epoch_num: 500
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: ./output/rec_chinese_common_v2.0
  save_epoch_step: 3
  # evaluation is run every 2000 iterations after the 0th iteration
  eval_batch_step: [0, 2000]
  cal_metric_during_train: True
  pretrained_model:
  checkpoints:
  save_inference_dir:
  use_visualdl: False
  infer_img: doc/imgs_words/ch/word_1.jpg
  # for data or label process
  character_dict_path: ppocr/utils/ppocr_keys_v1.txt
  character_type: ch
  max_text_length: 25
  infer_mode: False
  use_space_char: True
  save_res_path: ./output/rec/predicts_chinese_common_v2.0.txt


Optimizer:
  name: Adam
  beta1: 0.9
  beta2: 0.999
  lr:
    name: Cosine
    learning_rate: 0.001
  regularizer:
    name: 'L2'
    factor: 0.00004

Architecture:
  model_type: rec
  algorithm: CRNN
  Transform:
  Backbone:
    name: ResNet
    layers: 34
  Neck:
    name: SequenceEncoder
    encoder_type: rnn
    hidden_size: 256
  Head:
    name: CTCHead
    fc_decay: 0.00004

Loss:
  name: CTCLoss

PostProcess:
  name: CTCLabelDecode

Metric:
  name: RecMetric
  main_indicator: acc

Train:
  dataset:
    name: SimpleDataSet
    data_dir: ./train_data/PS/
    label_file_list: [ "./train_data/PS/rec_gt35.txt" ]
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - RecAug: 
      - CTCLabelEncode: # Class handling label
      - RecResizeImg:
          image_shape: [3, 32, 320]
      - KeepKeys:
          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: True
    batch_size_per_card: 4
    drop_last: True
    num_workers: 8

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./train_data/PS/
    label_file_list: [ "./train_data/PS/rec_gt35.txt" ]
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - CTCLabelEncode: # Class handling label
      - RecResizeImg:
          image_shape: [3, 32, 320]
      - KeepKeys:
          keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
  loader:
    shuffle: False
    drop_last: False
    batch_size_per_card: 4
    num_workers: 8

训练了200个epoch后评估效果(达到了100%识别率)

[2021/12/01 13:59:15] root INFO: metric eval ***************
[2021/12/01 13:59:15] root INFO: acc:1.0
[2021/12/01 13:59:15] root INFO: norm_edit_dis:1.0
[2021/12/01 13:59:15] root INFO: fps:41.23514527782272
INFO 2021-12-01 13:59:17,096 launch.py:268] Local processes completed.

总结

PaddleOCR训练效果很好。