运行的结果没有示例图中好 about recognize-anything

[xinyu1205]

你好我运行测试命令的时候报错，请问您有遇到吗：python inference_tag2text.py --image 042.jpg --pretrained tag2text_swin_14m.pth 报错： magic_number = pickle_module.load(f, **pickle_load_args) _pickle.UnpicklingError: invalid load key, 'v'.

你的这个报错我没有遇到，我倒是遇到了另一个报错

Traceback (most recent call last):
  File "inference_tag2text.py", line 94, in <module>
    res = inference(image, model, args.specified_tags)
  File "inference_tag2text.py", line 43, in inference
    caption, tag_predict = model.generate(image,
  File "/data2/home/tyu/stable_diffusion/promt_gen/Recognize_Anything-Tag2Text/models/tag2text.py", line 364, in generate
    torch.sigmoid(logits) > self.class_threshold,
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

因为我是在显卡上跑的，所以遇到这样的报错，可以通过将对应的代码

https://github.com/xinyu1205/Recognize_Anything-Tag2Text/blob/ffd1a283caea70ab8436645c0fd0f366ae7de3f8/models/tag2text.py#L364

修改为

torch.sigmoid(logits) > self.class_threshold.to(image.device),

就行了，小问题 @Coler1994 @xinyu1205

感谢你非常有价值的bug反馈，我已经修改对应的代码~

[onefish51]

你好我运行测试命令的时候报错，请问您有遇到吗：python inference_tag2text.py --image 042.jpg --pretrained tag2text_swin_14m.pth 报错： magic_number = pickle_module.load(f, **pickle_load_args) _pickle.UnpicklingError: invalid load key, 'v'.

你的这个报错我没有遇到，我倒是遇到了另一个报错

Traceback (most recent call last):
  File "inference_tag2text.py", line 94, in <module>
    res = inference(image, model, args.specified_tags)
  File "inference_tag2text.py", line 43, in inference
    caption, tag_predict = model.generate(image,
  File "/data2/home/tyu/stable_diffusion/promt_gen/Recognize_Anything-Tag2Text/models/tag2text.py", line 364, in generate
    torch.sigmoid(logits) > self.class_threshold,
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

因为我是在显卡上跑的，所以遇到这样的报错，可以通过将对应的代码
https://github.com/xinyu1205/Recognize_Anything-Tag2Text/blob/ffd1a283caea70ab8436645c0fd0f366ae7de3f8/models/tag2text.py#L364

修改为

torch.sigmoid(logits) > self.class_threshold.to(image.device),

就行了，小问题
@Coler1994 @xinyu1205

[majinyu666]

是model.threshold由0.68降到了0.64？我刚才改了但是好像没起到作用。还是其他哪个参数？谢谢

应该只是阈值问题，我这儿降到0.63能出lamp，door还要更低些

[Coler1994]

demo为了保证准确率，调高了阈值，牺牲了些召回，
grounded sam的pipeline由于有grounding dino兜底，阈值会偏低些。
我们在精细的调调每个类的阈值。

[onefish51]

是model.threshold由0.68降到了0.64？我刚才改了但是好像没起到作用。还是其他哪个参数？谢谢

[cpperrpr]

你好我运行测试命令的时候报错，请问您有遇到吗：python inference_tag2text.py --image 042.jpg --pretrained tag2text_swin_14m.pth
报错：
magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: invalid load key, 'v'.

[cpperrpr]

你好我运行测试命令的时候报错，请问您有遇到吗：python inference_tag2text.py --image 042.jpg --pretrained tag2text_swin_14m.pth 报错： magic_number = pickle_module.load(f, **pickle_load_args) _pickle.UnpicklingError: invalid load key, 'v'.

你的这个报错我没有遇到，我倒是遇到了另一个报错

Traceback (most recent call last):
  File "inference_tag2text.py", line 94, in <module>
    res = inference(image, model, args.specified_tags)
  File "inference_tag2text.py", line 43, in inference
    caption, tag_predict = model.generate(image,
  File "/data2/home/tyu/stable_diffusion/promt_gen/Recognize_Anything-Tag2Text/models/tag2text.py", line 364, in generate
    torch.sigmoid(logits) > self.class_threshold,
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

因为我是在显卡上跑的，所以遇到这样的报错，可以通过将对应的代码

https://github.com/xinyu1205/Recognize_Anything-Tag2Text/blob/ffd1a283caea70ab8436645c0fd0f366ae7de3f8/models/tag2text.py#L364

修改为

torch.sigmoid(logits) > self.class_threshold.to(image.device),

就行了，小问题 @Coler1994 @xinyu1205

谢谢，发现问题了是模型文件没clone好，谢谢你的回复