I dont understand how paddleocr is choosing rec+det models based on the different package version of paddlepaddle and paddleocr

[Greatz08]

When i was using paddleocr==2.6.1 and paddlepaddle==2.6.1 model that were auto downloaded were en_PP-OCRv3_det_infer and en_PP-OCRv3_det_infer or en_PP-OCRv2_det_infer (i dont remember correctly but their was version difference).

Right now i have upgraded to paddleocr==2.9 and paddlepaddle==3.0 , so now it is downloading and using v3_det_infer and v4_rec_infer model. Why is it happening ? Why cant it use v4 versions of both ? ( I experimented some hrs today so i dont have too much information on project working but would love to understand this ) and can i make it use latest v5 models with same version ( paddleocr==2.9 and paddlepaddle==3.0 ).

[liuhongen1234567]

Hello, prior to PaddleOCR 2.9, the models and downloaded weights for the same language and the same PP-OCR series in the PaddleOCR wheel package were fixed and could not be changed. Therefore, PaddleOCR 2.9 cannot use v5 models because PP-OCRv5 had not been developed yet when the PaddleOCR wheel was packaged. After PaddleOCR 3.0, the inference framework has been substantially upgraded, allowing users to replace models through model_name and model_dir. You can refer to the documentation here: https://paddlepaddle.github.io/PaddleOCR/latest/en/version3.x/pipeline_usage/OCR.html. Therefore, if you want to support both v3_det_infer and v4_det_infer, it is recommended to upgrade to PaddleOCR 3.0 or later versions.

[liuhongen1234567]

The reason why PaddleOCR 2.9 uses en_PP-OCRv3_det_infer.tar instead of en_PP-OCRv4_det_infer.tar might be that v3 performed better than v4 in more extensive testing. Therefore, v3 was chosen over v4 when packaging PaddleOCR 2.9.

[Greatz08]

@liuhongen1234567 thanks for the reply. Now i get it , and will soon update to 3.0