PaddleOCR2Pytorch icon indicating copy to clipboard operation
PaddleOCR2Pytorch copied to clipboard

PaddleOCR inference in PyTorch. Converted from [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR)

Results 81 PaddleOCR2Pytorch issues
Sort by recently updated
recently updated
newest added

1、转换的pp模型链接: https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar 2、运行脚本: python ./converter/ch_ppocr_v3_rec_converter.py 3、报错: Traceback (most recent call last): File "./converter/ch_ppocr_v3_rec_converter.py", line 60, in load_paddle_weights self.net.state_dict()[k].copy_(torch.Tensor(para_state_dict[ppname])) KeyError: 'Student.backbone.conv1._conv.weight' During handling of the above exception, another exception occurred: Traceback...

大佬什么时候开始更新3.0版本的svtr呢?期待

请问检测到的弯曲文字(如圆形徽章),怎么进行水平纠正?相关的纠正代码在哪个文件

请问这个模型能训练自己的数据集吗?

我在视频和图像数据上测试了西班牙语言识别模型,发现无法识别空格,使用Paddle原始模型转换pytorch模型后也不可以识别空格,不知道是因为什么,作者大大能帮忙看一下吗 ![image](https://user-images.githubusercontent.com/51230137/173328339-dcc637bd-cb30-45c1-bee8-119eaca8152d.png) ![image](https://user-images.githubusercontent.com/51230137/173328310-7b4432d3-ad03-4a88-911f-19575d9f0e34.png)

Can you share the dataloader (simple_dataset as well as lmdb_dataset) implemented in pytorch in data module and the loss function as well. I was able to reimplemented but can't reproduce...

##### 背景: 自己训练了个`ppocrv3 rec`模型,使用`ch_PP-OCRv3_rec.yml` 配置文件,训练完成后使用paddleOCR推理精度很高。 ##### 问题: 1. 模型在转换过程中报错:`KeyError: Student.backbone.conv1._conv.weight` 2. 将源码该处修改后,转换最后报错`RuntimeError: The size of tensor a (34) must match the size of tensor b (33) at non-singleton dimension 0`,...

调用ch_ptocr_v3_rec_infer.pth https://github.com/frotms/PaddleOCR2Pytorch/blob/main/tools/infer/predict_rec.py#L286 图片resize时高为48,宽被压缩,在识别一些英文和数字时,精度受到影响