Baseline convolutional approach to get familiar with the data. Uses green channel only.
Date | Commit | Epochs Trained | Augmentation | Artifacts | Val Macro F1 | Test Macro F1 |
---|---|---|---|---|---|---|
2018-10-06 | 1a99729 | 17 | No | link | 0.363 | 0.227 |
2018-10-07 | 172ebf0 | 23 | Yes | link | 0.315 | 0.184 |
Due to the extreme class imbalance, it could make sense to treat this as a few-shot learning task. Train an embedding using Triplet loss, and query the training set using something like nearest neighbors to classify new samples.