Custom dataset trained Mobile Detection Model v3 0 accuracy

[saylan21]

Hello everyone, I am trying to train ch_PP-OCRv3_det_slim model that shared on the Model List Page as the config file i am using https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml the default .yaml file only changing the `pretrained_model as the weights i downloaded from model list v3 page and using a custom dataset which has no problem while training recognation model. Currently the model i downloaded from model list page, is working with 70 percent accuracy on my dataset, i would to increase the accuracy by training on this model weights with my own dataset but the training is ending up with 0 accuracy. Could you tell me how to train this ch_PP-OCRv3_det_slim model with my own dataset and using the pretrained model on the model list page. Thank you :)

[andyjiang1116]

if you want to finetune your own model, you should use this model ch_PP-OCRv3_det not slim model. also you should use detection datasets not recognition dataset. https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/doc/doc_en/dataset/ocr_datasets_en.md#1-text-detection

[saylan21]

Hi thank you for the answer, I am already using the detection dataset which I created using PPOCRLabel tool. The model will run on a mobile device so I think it would be better to fine tune slim model. Is it possible to fine-tune slim model ?

[andyjiang1116]

It needs 2 step first you finetune the base model (not slim) second you slim the model according this doc https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/deploy/slim/quantization/README_en.md#3-quant-aware-training

[saylan21]

Thank you for the answer @andyjpaddle . 1-) I am trying to fine-tune base model ch_PP-OCRv3_det using ch_PP-OCRv3_det_cml.yml config. Is it enough to change the dataset folder and the pretrained weight for the Teacher model, Do i need to modify anything for the student models. And in the second step should i quantize the teacher model or the student model ?

2-) i already trained student model and run it on a mobile device but the model didn't give any output for the image size 150x340 then i upscale the images to 300x680 then it started to give output for a reason that i do not know ? So Is it possible to fine tune the model in a way that it runs without needing to be upscaled ?

[andyjiang1116]

1. it depends on what your application scenarios are, we provide 3 fine-tune methods, but the english doc is not ready, we'll update later, you can refer the Chinese doc https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/doc/doc_ch/PPOCRv3_det_train.md#3-%E5%9F%BA%E4%BA%8Eppocrv3%E6%A3%80%E6%B5%8Bfinetune%E8%AE%AD%E7%BB%83 Only change the dataset folder and the pretrained weight is enough for cml method. The teacher model usually has a big model size, student model is little. you can choose that you need.
2. the input image size should be same with your training shape

[saylan21]

@andyjpaddle thank you for the answer 👍 i find it out that my custom trained inference model is working better with a size multiple of 32. I also see this in the line 264 https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.6/ppocr/data/imaug/operators.py , Could you please tell me 1-) why it is required to resize my image to a multiple of 32 ? 2-) why It works better the more i upscale my image size ? For example it works better when i upscale my image size to 320 instead of 240 both of them are multiple of 32 ?

Thank you

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions.