Is this repository the final version used to produce the CVPR results?

[Becomeyourself]

Hi, thanks for releasing your code and great work.

I am trying to reproduce your CVPR paper, and I found a few issues that I hope you could help clarify:

1. Multi-weather (rain/snow/fog) joint training

The paper mentions joint training on three weather types, but the released training script mainly uses get_training_set_derain(...). I do not see a unified loader that mixes derain, dehaze, and desnow data together, nor the procedure for generating the 3-type classification label.

Question: Is there another training script or configuration that supports true multi-weather joint training as used in the paper?

---

2. Testing and evaluation scripts

There seem to be several inconsistencies:

• The model forward arguments differ between training and testing (the alpha argument is missing in test).
• num_queries is 48 during training but becomes 48 * nFrames in testing, which changes the model architecture.
• The PSNR/SSIM script matches GT and predicted frames using hard-coded indexing (i/5, i%5), and the Y-channel conversion formula appears unusual.

Question: Could you confirm the correct testing/evaluation procedure that matches the paper’s results?

---

3. Is this repo the final version used for CVPR?

Because of the mismatches above, I am unsure if this GitHub code is the exact version used for the paper.

Question: Is this repository the final version used to produce the CVPR results?
If not, is it possible to share the final scripts or the correct configuration?

---

Thanks a lot for your time and for sharing the code.