PaddlePaddle / PaddleOCR

Awesome multilingual OCR toolkits based on PaddlePaddle (practical ultra lightweight OCR system, support 80+ languages recognition, provide data annotation and synthesis tools, support training and deployment among server, mobile, embedded and IoT devices)
https://paddlepaddle.github.io/PaddleOCR/
Apache License 2.0
44.23k stars 7.82k forks source link

训练参数与模型不一致 #11723

Closed yun-zero closed 7 months ago

yun-zero commented 8 months ago

从日志中大概可以看出是训练的参数与模型不一致,但我是跟着教程一起做的,训练模型用的是https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar。训练的配置文件是ch_PP-OCRv3_rec_distillation.yml训练参数在上面日志中有。请问是否改更换模型还是修改配置参数

yun-zero commented 8 months ago

明白了,是字典的字符数和模型的不一致,请问对训练有什么影响吗

TingquanGao commented 8 months ago

这是因为加载的预训练模型的分类头和字典分类数不一致。只要确认字典是正确的,就没问题。

TingquanGao commented 7 months ago

长时间未回复,已close,如有问题,可reopen并补充