Created by: lkevinzc

Thank you for open-sourcing this amazing project!!

Do you mind sharing what your training specifications for the released recognition crnn lite model are? I could not imagine using over 10+ millions of images and training for 3000 epochs as stated in this config file : https://github.com/PaddlePaddle/PaddleOCR/blob/549108fe0aa0d87c0a3b2d471f1c653e89daab80/configs/rec/rec_chinese_lite_train.yml#L4

May I know roughly how many gpus did you use, how long did it take, and when did it converge to the final released model?

When I tried to train from scratch, I used that lsvt cropped text data and also synthetic data from data_renderer, but got very poor results. I think there must be something wrong, thus would like to listen to your experience!

Many thanks!! :)