Our team [ods.ai] Here we go again took 46th place in this competition. This is our training pipeline for this task.
We used the following tricks:
- Four-headed models
- Augmentations
- Cutmix
- Mixup
There is nothing innovative here, we just did it carefully 😉
Our final submission consists of four se_resnext101_32x4d models, trained on four different folds for 200 epochs on 128x128 image size and then trained for 200 epochs on 224x224 image size with a smaller learning rate.