同样的微调数据集，v5效果比v4差

[yuye0526]

🔎 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)

以下为PaddleOCR-release-2.7/configs/det/ch_PP-OCRv4/ch_PP-OCRv4_det_finetune.yml配置文件：

Global:
  debug: false
  use_gpu: true
  epoch_num: &epoch_num 300
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: ./models/ckpts
  save_epoch_step: 10
  eval_batch_step:
  - 0
  - 100
  cal_metric_during_train: false
  checkpoints:
  pretrained_model: ./pretrained/ch_PP-OCRv4_det_train/best_accuracy
  save_inference_dir: null
  use_visualdl: false
  infer_img: doc/imgs_en/img_10.jpg
  save_res_path: ./checkpoints/det_db/predicts_db.txt
  distributed: true

Architecture:
  model_type: det
  algorithm: DB
  Transform: null
  Backbone:
    name: PPLCNetV3
    scale: 0.75
    det: True
  Neck:
    name: RSEFPN
    out_channels: 96
    shortcut: True
  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: 0.0001 #(8*8c)
    warmup_epoch: 2
  regularizer:
    name: L2
    factor: 5.0e-05

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: ./data/imgs/
    label_file_list:
      - ./data/train.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - CopyPaste: 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: *epoch_num
    - MakeShrinkMap:
        shrink_ratio: 0.4
        min_text_size: 8
        total_epoch: *epoch_num
    - 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: 8

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./data/imgs/
    label_file_list:
      - ./data/test.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - DetResizeForTest:
    - 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
    num_workers: 2
profiler_options: null

以下为PaddleOCR-release-3.1/configs/det/PP-OCRv5/PP-OCRv5_mobile_det.yml配置文件：

Global:
  model_name: PP-OCRv5_mobile_det # To use static model for inference.
  debug: false
  use_gpu: true
  epoch_num: &epoch_num 500
  log_smooth_window: 20
  print_batch_step: 10
  save_model_dir: ./models/ckpts
  save_epoch_step: 10
  eval_batch_step:
  - 0
  - 100
  cal_metric_during_train: false
  checkpoints:
  pretrained_model: ./models/pretrained/PP-OCRv5_mobile_det_pretrained.pdparams
  save_inference_dir: null
  use_visualdl: false
  infer_img: doc/imgs_en/img_10.jpg
  save_res_path: ./checkpoints/det_db/predicts_db.txt
  d2s_train_image_shape: [3, 640, 640]
  distributed: true

Architecture:
  model_type: det
  algorithm: DB
  Transform: null
  Backbone:
    name: PPLCNetV3
    scale: 0.75
    det: True
  Neck:
    name: RSEFPN
    out_channels: 96
    shortcut: True
  Head:
    name: DBHead
    k: 50
    fix_nan: True

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 #(8*8c)
    warmup_epoch: 2
  regularizer:
    name: L2
    factor: 5.0e-05

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: ./data/imgs
    label_file_list:
      - ./data/train.txt
    ratio_list: [1.0]
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - CopyPaste: 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: *epoch_num
    - MakeShrinkMap:
        shrink_ratio: 0.4
        min_text_size: 8
        total_epoch: *epoch_num
    - 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: 8

Eval:
  dataset:
    name: SimpleDataSet
    data_dir: ./data/imgs
    label_file_list:
      - ./data/test.txt
    transforms:
    - DecodeImage:
        img_mode: BGR
        channel_first: false
    - DetLabelEncode: null
    - DetResizeForTest:
    - 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
    num_workers: 2
profiler_options: null

• 训练集数量在4k左右，基本来自印刷体PDF文件；
• 无论是使用旧的参数（500 epoch, 100 eval_batch_step），还是使用v5默认的参数（500 epoch, 1500 eval_batch_step），在v5上的效果都不如v4，也尝试了sever的微调，仍低于v4；
• d2s_train_image_shape: [3, 640, 640]有什么作用？
• v5的预训练模型有什么变化？v5的模型结构是否必须使用更大的微调数据量级？是否可以给一个参考的微调数据量？
• 暂时不考虑调整unclip_ratio，看了一下两者模型参数是差不多的，比较疑惑为什么v5的文本框比v4训练出来的明显小一圈，导致文字经常被截断，影响后续文字识别效果。

🏃‍♂️ Environment (运行环境)

官方docker环境

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

python3 tools/train.py

[dencentding]

我也有同样的问题，同样的数据集，微调的OCRV5，DB算法的det模型比 原来OCRV4 DB++算法的det模型准确率低很多，微调的V5 rec模型倒是比V4有所提升，特别是在手写领域

[1999swallowthemoon]

同样疑惑为什么v5的文本框比v4训练出来的小一圈，我还以为是我标注的时候把文本框画小了...

[leo-q8]

@yuye0526 mobile 模型结构没有发生变化，能提供一个 case 的不同模型可视化结果吗？另外数据集方便提供吗？我们排查一下问题

[1999swallowthemoon]

这是利用PPOCRLabel标注的时候:

这是微调后推理的结果,同样unclip_ratio=1.5:

[1999swallowthemoon]

PPOCRLabel标注:

微调PP-OCRv5_server_det后的推理:

[LMR2018]

v5训练效果问题解决了吗？你们是怎么在官方的模型基础上重新微调的模型？