使用ch_PP-OCRv4_det_cml.yml知识蒸馏训练，数据集和配置文件都和官方的一样，但是训练准确率和召回率很低，最高只有30%。

[jjkgsd]

🔎 Search before asking

• I have searched the PaddleOCR Docs and found no similar bug report.
• I have searched the PaddleOCR Issues and found no similar bug report.
• I have searched the PaddleOCR Discussions and found no similar bug report.

🐛 Bug (问题描述)

使用ch_PP-OCRv4_det_cml.yml知识蒸馏训练，数据集和配置文件都和官方的一样，但是训练准确率和召回率很低，最高只有30%，是什么原因呢？

配置文件：

Global:
  debug: false
  use_gpu: true
  epoch_num: 500
  log_smooth_window: 20
  print_batch_step: 20
  save_model_dir: ./output/ch_ppocr_v4_det_icdar_cml
  save_epoch_step: 10
  eval_batch_step:
  - 0
  - 500
  cal_metric_during_train: false
  checkpoints: null
  pretrained_model: 
  save_inference_dir: ./export_models/ch_ppocr_v4_det_cml
  use_visualdl: false
  infer_img: doc/imgs_en/img_10.jpg
  save_res_path: ./checkpoints/det_db/predicts_db.txt
  distributed: true
Architecture:
  name: DistillationModel
  algorithm: Distillation
  model_type: det
  Models:
    Student:
      pretrained: 
      model_type: det
      algorithm: DB
      Transform: null
      Backbone:
        name: PPLCNetV3
        scale: 0.75
        pretrained: false
        det: true
      Neck:
        name: RSEFPN
        out_channels: 96
        shortcut: true
      Head:
        name: DBHead
        k: 50
    Student2: 
      pretrained: 
      model_type: det
      algorithm: DB
      Transform: null
      Backbone:
        name: PPLCNetV3
        scale: 0.75
        pretrained: true
        det: true
      Neck:
        name: RSEFPN
        out_channels: 96
        shortcut: true
      Head:
        name: DBHead
        k: 50
    Teacher:
      pretrained: ./pretrained_models/teacher.pdparams
      freeze_params: true
      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:
  - DistillationDilaDBLoss:
      weight: 1.0
      model_name_pairs:
      - - Student
        - Teacher
      - - Student2
        - Teacher
      key: maps
      balance_loss: true
      main_loss_type: DiceLoss
      alpha: 5
      beta: 10
      ohem_ratio: 3
  - DistillationDMLLoss:
      model_name_pairs:
      - Student
      - Student2
      maps_name: thrink_maps
      weight: 1.0
      key: maps
  - DistillationDBLoss:
      weight: 1.0
      model_name_list:
      - Student
      - Student2
      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: 5.0e-05
PostProcess:
  name: DistillationDBPostProcess
  model_name:
  - Student
  key: head_out
  thresh: 0.3
  box_thresh: 0.6
  max_candidates: 1000
  unclip_ratio: 1.5
Metric:
  name: DistillationMetric
  base_metric_name: DetMetric
  main_indicator: hmean
  key: Student
Train:
  dataset:
    name: SimpleDataSet
    data_dir: ./datasets/icdar2015
    label_file_list:
      - ./datasets/icdar2015/train_label.txt
    ratio_list: [1.0]
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - IaaAugment:
        augmenter_args:
        - type: Fliplr
          args:
            p: 0.5
        - type: Affine
          args:
            rotate:
            - -10
            - 10
        - type: Resize
          args:
            size:
            - 0.5
            - 3
    - EastRandomCropData:
        size:
        - 640
        - 640
        max_tries: 50
        keep_ratio: true
    - MakeBorderMap:
        shrink_ratio: 0.4
        thresh_min: 0.3
        thresh_max: 0.7
        total_epoch: 500
    - MakeShrinkMap:
        shrink_ratio: 0.4
        min_text_size: 8
        total_epoch: 500
    - NormalizeImage:
        scale: 1./255.
        mean:
        - 0.485
        - 0.456
        - 0.406
        std:
        - 0.229
        - 0.224
        - 0.225
        order: hwc
    - ToCHWImage: null
    - KeepKeys:
        keep_keys:
        - image
        - threshold_map
        - threshold_mask
        - shrink_map
        - shrink_mask
  loader:
    shuffle: true
    drop_last: false
    batch_size_per_card: 8
    num_workers: 4
Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./datasets/icdar2015
    label_file_list:
      - ./datasets/icdar2015/valid_label.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - DetResizeForTest: 
        limit_side_len: 960
        limit_type: max
    - NormalizeImage:
        scale: 1./255.
        mean:
        - 0.485
        - 0.456
        - 0.406
        std:
        - 0.229
        - 0.224
        - 0.225
        order: hwc
    - ToCHWImage: null
    - KeepKeys:
        keep_keys:
        - image
        - shape
        - polys
        - ignore_tags
  loader:
    shuffle: false
    drop_last: false
    batch_size_per_card: 1

🏃‍♂️ Environment (运行环境)

python 3.8
paddlepaddle-gpu 2.5.2

🌰 Minimal Reproducible Example (最小可复现问题的Demo)

python tools/train.py -c configs/det/ch_PP-OCRv4_det_cml.yml

[changdazhou]

数据量有多少呢

[jjkgsd]

数据量有多少呢

就是官方的数据集,ICDAR2015,1000训练500验证。

[changdazhou]

几卡训练的呢，官方配置是4卡训练的，如果本地单卡训练需要减少学习率为原来的1/4

[jjkgsd]

几卡训练的呢，官方配置是4卡训练的，如果本地单卡训练需要减少学习率为原来的1/4
用的是本地单卡A100训练的

[that-learner]

我是微调训练v4 server版，指标也差不多只有30多，比微调前指标还低，可能v4不好训练

[lengmouling]

loss是多少啊？我这边训练会发现loss在10以下一点，但是普通训练server或者普通的loss是正常的0.几左右，用蒸馏loss高的离谱

[changdazhou]

V4预训练数据量大，确实不太容易微调哈，另外建议微调是时候学习率尽量小一些

[Sunting78]

您好，请问用的哪个分支呢？