中文crnn识别以及其模型转onnx
更正了原demo图片resize参数错误的问题
Win10 + torch1.8.1+cu111+cudnn8.1.1
- Download the dataset
- Edit lib/config/360CC_config.yaml DATA:ROOT to you image path
DATASET:
ROOT: 'to/your/images/path'
-
Download the labels (password: eaqb)
-
Put char_std_5990.txt in lib/dataset/txt/
-
And put train.txt and test.txt in lib/dataset/txt/
eg. test.txt
20456343_4045240981.jpg 89 201 241 178 19 94 19 22 26 656
20457281_3395886438.jpg 120 1061 2 376 78 249 272 272 120 1061
...
- Edit lib/config/OWN_config.yaml DATA:ROOT to you image path
DATASET:
ROOT: 'to/your/images/path'
-
And put your train_own.txt and test_own.txt in lib/dataset/txt/
eg. test_own.txt
20456343_4045240981.jpg 你好啊!祖国!
20457281_3395886438.jpg 晚安啊!世界!
...
note: fixed-length training is supported. yet you can modify dataloader to support random length training.
注:将训练的字符添加至lib/config/alphabets.py后,再进行训练,否则训练将出错
[run] python train.py --cfg lib/config/360CC_config.yaml
or [run] python train.py --cfg lib/config/OWN_config.yaml
#### loss curve
```angular2html
[run] cd output/360CC/crnn/xxxx-xx-xx-xx-xx/
[run] tensorboard --logdir log
[run] python demo.py --image_path images/test.png --checkpoint output/checkpoints/mixed_second_finetune_acc_97P7.pth
将export_onnx.py的模型位置替换成你自己的位置,在onnx_test.py进行测试
https://github.com/Sierkinhane/CRNN_Chinese_Characters_Rec