paddleOCR训练
1
| python tools/infer/predict_system.py --image_dir="D:\User\yaoye\Pictures\QQ20241226-005135.png" --det_model_dir="./model/ch_PP-OCRv3_det_infer/" --rec_model_dir="./model/ch_PP-OCRv3_rec_infer/"
|
1
| pip install numpy==1.23.5
|
1
| python gen_ocr_train_val_test.py --trainValTestRatio 9:1:0 --datasetRootPath ../train_data/drivingData
|
1
| python tools/train.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml -o Global.checkpoints=./output/ch_PP-OCR_V3_det/latest
|
1
| python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml
|
1
| python tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec.yml
|
cuda
cudnn
1
| python tools/infer_rec.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=output/rec/best_accuracy.pdparams Global.infer_img=“C:\Users\User\Desktop\PaddleOCR-release-2.6\train_data\rec\test\0001_1 (110)_crop_3.jpg”
|
训练完成后没有best_accuracy.pdparams，需要在训练时调整配置文件eval_batch_step，源码中为 eval_batch_step: [3000, 2000]
意思是在训练时每训练 3000 和 2000 个 batch 后，进行一次评估（validation），需要调整为eval_batch_step: [ 0, 19 ]，第0与第19轮时做一组评估，否则不会出现best的相关文件，同时需要修改batch_size_per_card和epoch_num，保证batch_size_per_card<epoch_num
1
| python tools/infer_rec.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec.yml -o Global.pretrained_model=output/rec_ppocr_v3_distillation/latest.pdparams Global.infer_img="D:\Application\Develop\Project\python\PaddleOCR\train_data\drivingData\crop_img\002_crop_0.jpg"
|
93.8 98
