|
|
--- |
|
|
license: cc |
|
|
--- |
|
|
|
|
|
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: |
|
|
|
|
|
1. 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. |
|
|
2. 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. |
|
|
3. Using EfficientNet as a simple and fast base network, made experimenting fairly fast and simple. |
|
|
|
|
|
Things I wish I had done differently: |
|
|
|
|
|
1. 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. |
|
|
2. Making the saving and logging process cleaner overall so it would be easier to compare results |
