微调自定义数据集用ch_PP-OCRv3_rec_distillation.yml vs ch_PP-OCRv4_rec_distillation.yml都有些bug

[hecheng64]

🔎 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)

在bug14870：ch_PP-OCRv3_rec_distillation.yml vs ch_PP-OCRv4_rec_distillation.yml
如果使用的是 PP-OCRv3 作为基础模型，建议使用 ch_PP-OCRv3_rec_distillation.yml 进行蒸馏训练，以保持与 PP-OCRv3 训练策略一致。
回答：因为我当初按照官网训练：
14870

Architecture:
  model_type: &model_type "rec"
  name: DistillationModel
  algorithm: Distillation
  Models:
    Teacher:
      pretrained:
      freeze_params: false                                 -----官方默认配置false
      return_all_feats: true
      model_type: *model_type
      algorithm: SVTR_LCNet

可以训练，可以收敛，但是感觉对手写体测试有效果，原先印刷体失效的，后面怀疑 freeze_params: false 配置
有问题，根据bug 14866 建议GreatV大佬建议 改成 freeze_params: true，训练好久，损失率还很大，收敛不了

  Models:
    Teacher:
      pretrained:
      freeze_params: true
      return_all_feats: true
      model_type: *model_type
      algorithm: SVTR_LCNet
      Transform:

数据集是手写OCR汇总，由中科院手写数据和网上开源数据合并组合：https://aistudio.baidu.com/datasetdetail/102884/0

如果使用 PP-OCRv4 作为基础模型，则应使用 ch_PP-OCRv4_rec_distillation.yml，因为 PP-OCRv4 可能在蒸馏策略上有优化或新的调整。
用 ch_PP-OCRv4_rec_distillation.yml 训练会报错

python tools/train.py -c configs/rec/PP-OCRv4/ch_PP-OCRv4_rec_distillation.yml

[2025/03/17 15:13:06] ppocr ERROR: When parsing line train_data/138440.png      率普遍较高如方正
, error happened with msg: Traceback (most recent call last):
  File "/home/PaddleOCR/ppocr/data/simple_dataset.py", line 137, in __getitem__
    outs = transform(data, self.ops)
  File "/home/PaddleOCR/ppocr/data/imaug/__init__.py", line 73, in transform
    data = op(data)
  File "/home/PaddleOCR/ppocr/data/imaug/operators.py", line 123, in __call__
    data_list.append(data[key])
KeyError: 'valid_ratio'   然后尝试在配置里面把     - KeepKeys:
        keep_keys:
        - image
        - label_ctc
        - label_gtc
        - length
        - valid_ratio   把valid_ratio 去掉 要报其他错误
Exception in thread Thread-1 (_thread_loop):
Traceback (most recent call last):
  File "/root/anaconda3/envs/ocr/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
    self.run()
  File "/root/anaconda3/envs/ocr/lib/python3.10/threading.py", line 953, in run
    self._target(*self._args, **self._kwargs)
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/dataloader_iter.py", line 619, in _thread_loop
    batch = self._get_data()
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/dataloader_iter.py", line 766, in _get_data
    batch.reraise()
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/worker.py", line 195, in reraise
    raise self.exc_type(msg)
ValueError: DataLoader worker(1) caught ValueError with message:
Traceback (most recent call last):
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/worker.py", line 380, in _worker_loop
    batch = fetcher.fetch(indices)
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/fetcher.py", line 85, in fetch
    data = self.collate_fn(data)
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/collate.py", line 75, in default_collate_fn
    return [default_collate_fn(fields) for fields in zip(*batch)]
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/collate.py", line 75, in <listcomp>
    return [default_collate_fn(fields) for fields in zip(*batch)]
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/collate.py", line 56, in default_collate_fn
    batch = np.stack(batch, axis=0)
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/numpy/core/shape_base.py", line 449, in stack
    raise ValueError('all input arrays must have the same shape')
ValueError: all input arrays must have the same shape

Traceback (most recent call last):
  File "/home/PaddleOCR/tools/train.py", line 270, in <module>
    main(config, device, logger, vdl_writer, seed)
  File "/home/PaddleOCR/tools/train.py", line 223, in main
    program.train(
  File "/home/PaddleOCR/tools/program.py", line 312, in train
    for idx, batch in enumerate(train_dataloader):
  File "/root/anaconda3/envs/ocr/lib/python3.10/site-packages/paddle/io/dataloader/dataloader_iter.py", line 840, in __next__
    self._reader.read_next_list()[0]
SystemError: (Fatal) Blocking queue is killed because the data reader raises an exception.
  [Hint: Expected killed_ != true, but received killed_:1 == true:1.] (at ../paddle/phi/core/operators/reader/blocking_queue.h:175)

🏃‍♂️ Environment (运行环境)

release/2.10.0

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

见bug描述

[GreatV]

能提供一个最小复现的demo吗，我试试看。

[hecheng64]

你说用 ch_PP-OCRv4_rec_distillation.yml 训练会报错？ 这个必现的，我看issue里面差不多5.6个提过，好像在安排修复中 @GreatV 如果用 ch_PP-OCRv3_rec_distillation.yml 进行蒸馏训练 ，不收敛的话，我把配置给你

ch_PP-OCRv3_rec_distillation.yml.txt

，数据集链接就在https://aistudio.baidu.com/datasetdetail/102884/0

[GreatV]

感觉对手写体测试有效果，原先印刷体失效的

PaddleOCR/configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml

Lines 44 to 46 in 75526f0

	Teacher:
	pretrained:
	freeze_params: false

可能的原因

1. 预训练模型未加载

如果 pretrained 字段为空，教师模型可能从随机权重开始训练，而不是使用 PP-OCRv3 的预训练权重。这可能导致模型在印刷体上的性能下降，而更适应训练数据中的手写体。

2. 训练数据分布

训练数据中手写体样本可能较多，导致模型在手写体上表现良好，但在印刷体上表现较差。建议确保训练数据包含足够的印刷体样本。

3. 教师模型参数未冻结

配置文件中 freeze_params: false 表示教师模型参数在训练中会更新，这可能导致教师模型偏向于训练数据中的手写体特征，影响印刷体性能。

[hecheng64]

@GreatV https://aistudio.baidu.com/projectdetail/4330587 你能不能提供下你们官方例子OCR手写文字识别例子里面修改好的ch_PP-OCRv3_rec_distillation.yml 配置，我看你们代码里面doc描述https://paddlepaddle.github.io/PaddleOCR/latest/applications/%E6%89%8B%E5%86%99%E6%96%87%E5%AD%97%E8%AF%86%E5%88%AB.html ：configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml

epoch_num: 100 # 训练epoch数
save_model_dir: ./output/ch_PP-OCR_v3_rec
save_epoch_step: 10
eval_batch_step: [0, 100] # 评估间隔，每隔100step评估一次
pretrained_model: ./pretrained_models/ch_PP-OCRv3_rec_train/best_accuracy # 预训练模型路径

lr:
name: Cosine # 修改学习率衰减策略为Cosine
learning_rate: 0.0001 # 修改fine-tune的学习率
warmup_epoch: 2 # 修改warmup轮数

Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data # 训练集图片路径
ext_op_transform_idx: 1
label_file_list:
- ./train_data/chineseocr-data/rec_hand_line_all_label_train.txt # 训练集标签
- ./train_data/handwrite/HWDB2.0Train_label.txt
- ./train_data/handwrite/HWDB2.1Train_label.txt
- ./train_data/handwrite/HWDB2.2Train_label.txt
- ./train_data/handwrite/hwdb_ic13/handwriting_hwdb_train_labels.txt
- ./train_data/handwrite/HW_Chinese/train_hw.txt
ratio_list:
- 0.1
- 1.0
- 1.0
- 1.0
- 0.02
- 1.0
loader:
shuffle: true
batch_size_per_card: 64
drop_last: true
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data # 测试集图片路径
label_file_list:
- ./train_data/chineseocr-data/rec_hand_line_all_label_val.txt # 测试集标签
- ./train_data/handwrite/HWDB2.0Test_label.txt
- ./train_data/handwrite/HWDB2.1Test_label.txt
- ./train_data/handwrite/HWDB2.2Test_label.txt
- ./train_data/handwrite/hwdb_ic13/handwriting_hwdb_val_labels.txt
- ./train_data/handwrite/HW_Chinese/test_hw.txt
loader:
shuffle: false
drop_last: false
batch_size_per_card: 64
num_workers: 4

感觉很写有点含糊：上面例子感觉没有teacher，student，但是他名字ch_PP-OCRv3_rec_distillation.yml 不应该没有teacher，student，你能否提供完整的针对这个例子修改好ch_PP-OCRv3_rec_distillation.yml 。已经走好几天弯路，网上好多例子都拷贝这份描述，这个例子到底有没有配置teacher，student，如果配置teacher，student里面到底怎么配置，一会说 freeze_params: false，一会true，我就是想看看你们跑成功手写体例子时候ch_PP-OCRv3_rec_distillation.yml

[liu-jiaxuan]

@zhangyubo0722 帮忙看下这个吧~

[github-actions[bot]]

This issue is stale because it has been open for 90 days with no activity.

[NANCHENG6]

@hecheng64 解决了嘛. 好几个月前我看都有人提了, 现在貌似还是不行 一模一样的bug , 有啥替代的方法吗