GreatV
commented
3 days ago The issue you're facing, where the PaddleOCR inference model produces incorrect predictions (e.g., Chinese symbols instead of digits) despite correct training and evaluation results, is a common problem that arises in OCR workflows due to mismatched configurations or issues during model export and inference. Below are the likely causes and steps to resolve the issue:
1. Mismatch in rec_char_dict_path
- Ensure that the
rec_char_dict_pathused during training, evaluation, and inference is consistent. - Your
digit_dict.txtshould contain only the digits0-9, one per line. Double-check that this file does not include any non-numeric characters or extra blank spaces.
Example:
0
1
2
3
4
5
6
7
8
9If there is any discrepancy in the character dictionary, the inference model may produce unexpected characters, such as Chinese symbols, because it interprets the output indices incorrectly.
2. Incorrect Model Export
When exporting the trained model to an inference model, ensure that the correct checkpoints and configurations are used. Verify that the following steps were executed correctly:
- The
export_model.pyscript should reference the same configuration file (en_PP-OCRv3_rec.yml) and checkpoint file (iter_epoch_24.pdparams) as used during training and evaluation. - The
Global.rec_char_dict_pathparameter in the configuration file should point to yourdigit_dict.txt.
Command you ran:
python3 tools/export_model.py \
-c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml \
-o Global.checkpoints=./output/v3_en_mobile/iter_epoch_24.pdparams \
Global.save_inference_dir=./inference/rec_digitsVerify the above command and ensure the paths are correct. If there was an issue during export, the inference model may not align with the trained weights.
3. Inference Configuration Issues
The issue may lie in the parameters passed to the PaddleOCR class during inference. Based on your code:
ocr = PaddleOCR(
use_gpu=False,
rec_char_dict_path='./digit_dict.txt',
rec_model_dir="./PaddleOCR/inference/rec_digits"
)Check the following:
rec_model_dirshould point to the folder containing the exported inference model files (inference.pdiparams,inference.pdiparams.info, andinference.pdmodel).- Ensure there are no typos in the
rec_char_dict_pathparameter.
If the configuration is incorrect, the model may fail to interpret its predictions properly.
4. Model Evaluation vs. Inference
Even though the model achieves 100% accuracy during evaluation, this does not guarantee it will work perfectly during inference. The evaluation step uses the training pipeline, which differs slightly from the inference pipeline in terms of preprocessing and postprocessing. Here are some steps to debug:
- Run
infer_rec.py: Use theinfer_rec.pyscript to test the exported model directly on your test images. This script uses the same preprocessing pipeline as the training process.python3 tools/infer_rec.py \ --image_dir=path_to_test_image \ --rec_model_dir=./PaddleOCR/inference/rec_digits \ --rec_char_dict_path=./digit_dict.txt \ --use_gpu=False
Compare the results from infer_rec.py with those from the PaddleOCR API. If the results from infer_rec.py are correct, the issue may lie in how the PaddleOCR class is configured.
5. Preprocessing Differences
Differences in preprocessing between training and inference can also cause discrepancies. For example:
- Resize settings, normalization, or padding during inference might differ from the training pipeline.
- The
rec_image_shapeparameter in your configuration file (e.g.,en_PP-OCRv3_rec.yml) should match the input shape expected by the inference model.
Verify the rec_image_shape in your config file (e.g., ["3", "32", "100"] for recognition) and ensure it matches the preprocessing applied in your inference script.
6. Debugging Steps
- Check Model Export Logs: Ensure no warnings or errors occurred during model export.
- Verify Character Set: Use the same
digit_dict.txtfile across training, evaluation, export, and inference. - Test with
infer_rec.py: Run the exported model using theinfer_rec.pyscript and compare results. - Inspect Preprocessing: Verify that the preprocessing pipeline used during inference matches the training pipeline.
References to Related Issues
-
Issue #13972: 导出模型inference后，paddleocr命令行执行结果不一致
- This issue highlights inconsistencies in results due to differences in configurations or parameters during inference. Ensure that the character dictionary and model paths are consistent.
-
Discussion #14088: 用自己数据集训练的识别模型的train_model转换为inference_model后识别结果错误
- Similar issue where the exported inference model produced incorrect results. The resolution involved double-checking the character dictionary and preprocessing settings.
Conclusion
To resolve your issue:
- Verify the
digit_dict.txtfile and ensure it is correctly referenced during all steps. - Re-export the model using consistent configurations and checkpoints.
- Test the model with
infer_rec.pyto isolate whether the issue lies in the export process or thePaddleOCRAPI configuration. - Align preprocessing between training and inference.
If the problem persists, share the outputs of infer_rec.py and your configuration file for further debugging.
Response generated by 🤖 feifei-bot | chatgpt-4o-latest
Pedro69491
🔎 Search before asking
🐛 Bug (问题描述)
I have been using PaddleOCR training capabilities in a small dataset of digits, after 25 epochs the accuracy of the model reaches 100%, I then evaluate the model and I get an accuracy of 100% too. The problem is that when I try to test the model on the exact same images I used to eval, I get completely different results, of course Iexported the best weights from the trained recognition model
from PaddleOCR.paddleocr import PaddleOCR
ocr = PaddleOCR( use_gpu=False, rec_char_dict_path='./digit_dict.txt', rec_model_dir="./PaddleOCR/inference/rec_digits", # Path to the saved model )
result = ocr.ocr(image_path)
output: chinese symbol instead of a number
notes: digit_dict refers to a small text file containing the numbers from 0 to 9
I tried to use infer_rec.py but the results were no good at all again with data that we already used to validate, not sure what I should do next.
🏃♂️ Environment (运行环境)
linux: ubuntu 22.04 python3: 10.12
🌰 Minimal Reproducible Example (最小可复现问题的Demo)
python3 tools/train.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.use_gpu=False
python3 tools/eval.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.checkpoints=./output/v3_en_mobile/iter_epoch_24.pdparams Global.use_gpu=False
python3 tools/export_model.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.checkpoints=./output/v3_en_mobile/iter_epoch_24.pdparams Global.save_inference_dir=./inference/rec_digits