Missing CSV data splits for M&Ms dataset

[Ljh-11]

Hi authors,

Thank you for your impressive work on SicTTA!

I am trying to reproduce your results using the M&Ms dataset, but I noticed the required CSV files (e.g., all.csv, train.csv, valid.csv, test.csv) referenced in dataset.py are missing from the repo.

Could you please upload the CSV files used in your experiments?

Thanks!

[JianghaoWu]

Hi,

Thank you very much for your interest in SicTTA and for pointing this out!

We have released our processed M&MS dataset, which already includes all the CSV split files used in our experiments.
After you have obtained the official permission to access the M&MS dataset, you may use our processed version here:

🔗 https://drive.google.com/file/d/1jaT2nsbF1-rPoWs6fnF9DsFxTuaYXqW2/view?usp=sharing

If you encounter any other issues or have suggestions, feel free to let us know — we are happy to improve the repository together.

Thanks again!

[Ljh-11]

Hi Jianghao,

Thank you so much for your prompt response and for sharing the processed dataset! I will download the data and proceed with the experiments. I really appreciate your support and contribution to the community.

Best regards!

[Ljh-11]

Hi Jianghao,

I have a question regarding the consistency between the code and the paper (Figure 2).

Code Behavior: In configure_model, you set m.track_running_stats = False globally. This forces the model to use current batch statistics even during the CCD calculation inside forward_and_adapt.

In forward_and_adapt:
fine = self.get_fine_ccd(x, model.eval(), ...) # model uses batch stats here
Paper Logic (Figure 2): The logic implies a two-stage BN strategy:

Screening Stage (CCD): Single input images should use Source BN Statistics (Global Mean/Var) to accurately detect domain shift. Using batch statistics here (Instance Norm) would calibrate the features early and mask the shift, making CCD ineffective.

Adaptation Stage (SABE): Only the enhanced batch (input + SFTs) should use Current Batch Statistics for correction.

Question: Shouldn't we use self.model_anchor (which retains source running_stats) inside get_fine_ccd instead of the adapted model? The current implementation seems to apply batch statistics to both stages.

Thanks!

[JianghaoWu]

Hi, thanks a lot for the detailed observation and helpful suggestion!

You're right that CCD is supposed to be computed with the source model statistics.

During development we experimented with several BN strategies, and the uploaded version unfortunately used the adapted model (with batch-stat BN) in get_fine_ccd, which was not the final intended design.

Following your suggestion, we updated the code to:

fine = self.get_fine_ccd(
    x,
    self.model_anchor.eval(),   # source model with running stats
    self.entropy_list,
    threshold=threshold
)

We tested this modification and found the results to be very similar:

[Epoch 0] [B] dice Mean: [0.86786473 0.78108845 0.72440442 0.7911192 ], Std: [0.17189358 0.11984684 0.32954487 0.15526257]
[Epoch 0] [C] dice Mean: [0.85348185 0.74219664 0.69228899 0.76265582], Std: [0.18430156 0.16067238 0.34796776 0.18008253]
[Epoch 0] [D] dice Mean: [0.86240096 0.74731272 0.70124337 0.77031901], Std: [0.18619274 0.16281994 0.3457582  0.16898657]

, indicating that CCD is relatively stable under our setting.

The repository has been updated accordingly.

Thanks again for helping improve the code quality!