astra-vision / poda Goto Github PK

Hello,

Interesting paper!

There is a size mismatch for running the feature optimization part below.

Traceback (most recent call last):
File "./PODA/f_aug_with_CLIP.py", line 283, in
main()
File "./PODA/f_aug_with_CLIP.py", line 253, in main
loss_CLIP1 = (1- torch.cosine_similarity(text_target, target_features_from_f1, dim=1)).mean()
RuntimeError: The size of tensor a (16) must match the size of tensor b (15) at non-singleton dimension 0

I'd appreciate it if you could help with this.

Thanks

Hello, can you open source the overall code of the detection? I tried to reproduce the code of the detection part, but it did not achieve the desired effect. Can you see how you achieved it?

Hi,

Appreciate your great work, very inspiring!

I'm wondering if there is any results for acdc foggy scene?

Hello and thank you for your great work.

I encountered an error when running the prediction code. It seems to be a problem with the pytorch version, which prevents the pre-trained model from being loaded. Can you provide the versions of python, pytorch, and torchvision you used for training?

Thanks a lot!

Thanks for your great work!
Have you tried using the ViT-b/16 as a backbone? I was replacing the backbone to align with other work and found that compared to the ResNet50, the ViT-b/16 outputs an extra global token, causing that the features can not be reshaped to [B, C, H, W].

Hello and thank you for your great work！

I use the CS source model you provided to predict citypaces images, but the visualization results are not very good. Is the source model provided only trained on the source domain without prompt? If a prompt is passed, what is the prompt corresponding to that situation?

I also evaluated the CS source model on cityscapes' val, and the accuracy was very low. Is it the wrong model?

Looking forward to your reply!

Could you release the code of image classification?