user1018

Results 84 comments of user1018

无法复现ppstructure 的recovery效果

英文文档推荐使用英文模型，添加`--lang='en`, 详细内容可以参考[这里](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.6/ppstructure/docs/quickstart.md#216-%E7%89%88%E9%9D%A2%E6%81%A2%E5%A4%8D)

Difference with original implementation

@liuguilin1225 If I have a real broken image, how can I generate the input mask of the broken image by extracting the part of broken pixels ? Thanks for your...

Difference with original implementation

@liuguilin1225 Another question： In your paper, only images generated by original images and masks are used?

[Question] loss is too low, only 0.3

@haotian-liu Could you help me see how I should change it

Code doesn't work with numpy>=1.24

Thanks for your suggestion, we'll discuss it further.

Code doesn't work with numpy>=1.24

We'vd solved this problem, you can try again

[Help Wanted]: Grounding Bias After Fine-Tuning on Qwen2.5-VL-7B

> I tried again, this time using the format exactly as shown in the Qwen2.5-VL blog. However, the bbox still shifts upward 😭 > > 你好，请问你的训练grounding的prompt是什么样

Qwen2.5-VL grounding

Thank you very much. I'll try

swift微调Qwen2.5-VL，合并模型后，分别使用swift和transform推理结果不一致

还是不一样，请问还有哪里可能设置的不同

training code

Thanks for your reply, could you tell us your training environment（such as, the number and model of GPU）、the training time of geometric unwarping transformer and illumination correction transformer