Model that acheived 5th place overall on FathomNet 2025 kaggle competition (competition link: https://www.kaggle.com/competitions/fathomnet-2025/overview)
Associated code: https://github.com/robertahunt/FathomNet2025
Things I think worked well:
- Choosing the class which minimized the expected loss based on the distance matrix (and not the class with the highest probability) and implementing this as a matrix multiplication with the distance matrix.
- Implementing a small Graph Neural Network layer - the idea here was to help in cases where there are many specimens of the same species in a single overall image, and one is easy to classify, but the other instances may be blurry. Then adding a graph layer could help guide the model to the correct classification.
- Using EfficientNet as a simple and fast base network, made experimenting fairly fast and simple.
Things I wish I had done differently:
- Setting the seed earlier: I initially used a random seed each time, which made measuring progress and reproducing results difficult. It wasn't until near the end I changed this. This is why the results using this seed also sadly do not match perfectly with the public results.
- Making the saving and logging process cleaner overall so it would be easier to compare results
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
🙋
Ask for provider support