Reproducing S2M-Net: One Verified Claim and Lessons Learned
#5
by K4ru4k4i - opened
This was my first time participating in a reproducibility challenge and contributing to the Hugging Face open-science community.
I worked on the reproducibility analysis of S2M-Net, a lightweight architecture for medical image segmentation:
- One architectural claim was independently verified.
- A second claim concerning the FFT-based token mixer revealed discrepancies between the paper and the available implementation.
- I also prepared a matched experimental pipeline for a third claim, although personal commitments prevented me from completing the full training campaign before the deadline.
Even though I could not finish everything I had planned, this experience taught me a great deal about model auditing, reproducible experiments, scientific documentation, and the value of reporting uncertain or incomplete evidence transparently.
Thank you to the organizers and the community for this opportunity. This certificate feels like a small but meaningful first step into open and collaborative AI research. 🤗
Full reproducibility logbook:
https://huggingface.co/spaces/K4ru4k4i/eh48NIgu9z