采用PP-OCRv3 det文字检测 教师模型dml 训练的问题

[tairen99]

Hi all and @WenmuZhou

首先感谢你们提供了这个良好的OCR平台。

• 系统环境/System Environment：x86_64 Ubuntu 20.04.6 LTS
• 版本号/Version：Paddle：2.6.0 PaddleOCR：2.7
• 问题相关组件/Related components：ch_PP-OCRv3_det_cml.yml

根据 文本检测模型训练，我想尝试使用 ch_PP-OCRv3_det_dml.yml 进行教师检测模型的训练。我首先想用 ICIDAR2015 的数据做训练测试，看看教师模型效果咋样；我测试了三种不同的YML配置文件：

第一种：将 "Global” 中的pretrained_model: ./pretrained_models/db/MobileNetV3_large_x0_5_pretrained 屏蔽了，因为可以避免导入阶段的各种警告，具体情况如下：

Global:
  use_gpu: true
  epoch_num: 1200
  log_smooth_window: 20
  print_batch_step: 2
  save_model_dir: ./output/db_icidar2015/
  save_epoch_step: 40
  # evaluation is run every 5000 iterations after the 4000th iteration
  eval_batch_step: [0, 500]
  cal_metric_during_train: False
  # pretrained_model: ./pretrained_models/db/MobileNetV3_large_x0_5_pretrained
  checkpoints:
  save_inference_dir:
  use_visualdl: False
  infer_img: doc/imgs_en/img_10.jpg
  save_res_path: ./output/det_db/predicts_db.txt
Architecture:
  name: DistillationModel
  algorithm: Distillation
  model_type: det
  Models:
    Student:
      pretrained: ./pretrained_models/db/ResNet50_vd_ssld_pretrained
      freeze_params: false 
      return_all_feats: false
      model_type: det
      algorithm: DB
      Backbone:
        name: ResNet_vd
        in_channels: 3
        layers: 50
      Neck:
        name: LKPAN
        out_channels: 256
      Head:
        name: DBHead
        kernel_list: [7,2,2]
        k: 50
    Student2:
      pretrained: ./pretrained_models/db/ResNet50_vd_ssld_pretrained
      freeze_params: false 
      return_all_feats: false
      model_type: det
      algorithm: DB
      Backbone:
        name: ResNet_vd
        in_channels: 3
        layers: 50
      Neck:
        name: LKPAN
        out_channels: 256
      Head:
        name: DBHead
        kernel_list: [7,2,2]
        k: 50
Loss:
  name: CombinedLoss
  loss_config_list:
  - DistillationDMLLoss:
      model_name_pairs:
      - ["Student", "Student2"]
      maps_name: "thrink_maps"
      weight: 1.0
      # act: None
      model_name_pairs: ["Student", "Student2"]
      key: maps
  - DistillationDBLoss:
      weight: 1.0
      model_name_list: ["Student", "Student2"]
      # key: maps
      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: DistillationDBPostProcess
  model_name: ["Student", "Student2"]
  key: head_out
  thresh: 0.3
  box_thresh: 0.6
  max_candidates: 2000
  unclip_ratio: 1.5

Metric:
  name: DistillationMetric
  base_metric_name: DetMetric
  main_indicator: hmean
  key: "Student"

Train:
  dataset:
    name: SimpleDataSet
    data_dir: ./train_data/icdar2015/
    label_file_list:
      - ./train_data/icdar2015/train_icdar2015_label.txt
    ratio_list: [1.0]
    transforms:
      - DecodeImage: # load image
          img_mode: BGR
          channel_first: False
      - DetLabelEncode: # Class handling label
      - CopyPaste:
      - 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: 2
    num_workers: 1

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./train_data/icdar2015/
    label_file_list:
      - ./train_data/icdar2015/test_icdar2015_label.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: 2

第二种情况，将 "Global” 中的pretrained_model: ./pretrained_models/db/MobileNetV3_large_x0_5_pretrained 还原
第三种情况，将所有的 student中的 pretrained: ./pretrained_models/db/ResNet50_vd_ssld_pretrained，
更换成 pretrained: ./pretrained_models/db/MobileNetV3_large_x0_5_pretrained；因为参考了 帖子，怀疑训练数据不够

可是所有这三种情况，我都训练了600 epoch，best metric hmean 都小于 0.72，无法再提高。不知道什么原因，也不知道哪里出现了问题。
恳请各位指点，支招，谢谢大家！