Training Detection model on custom data

[gibotsgithub]

I want to finetune pretrained model on my data but just after downloading the pretrained model when checking the detection prediction for any image it always show empty So not sure weather I am using pretrained weights or training the whole model from scratch can you please suggest something I just want to finetune pretrained model on my data ?

[LDOUBLEV]

Hi, which model do you wish to finetune？
There are three models for you to choose from, namely PPOCR, PPOCRv2, PPOCRV3. model_list

Take the training of the PPOCRv3 model as an example.
The PP-OCRv3 model trained on chinese data are follows.

The model for english.
https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.5/doc/doc_en/models_list_en.md#12-english-detection-model

And corresponding config files.

First, verify that the prediction can get the result after loading the pretrained weight.

wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar
tar xf ch_PP-OCRv3_det_distill_train.tar 
# inference 
python3 tools/infer_det.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.pretrained_model=./ch_PP-OCRv3_det_distill_train/best_accuracy  Global.infer_img=./doc/imgs/11.jpg

The detected Image saved in ./checkpoints/det_db/det_results_Student/11.jpg

So, -o Global.pretrained_model=./ch_PP-OCRv3_det_distill_train/best_accuracy is OK to load pretrained model.

For finetune training of distillation models, please refer to doc.