Simple solution (0.857 F1-Score / 0.906 Acc public LB) for take part in AIcrowd Seismic Facies Identification Challenge.
Pretrained weights from one of 5 folds
- UNet (efficientnet-b3 encoder) + scse
- Augs: HorizontalFlip, ShiftScaleRotate
- Loss: 0.75 * Lovasz + 0.25 * CE
- Optimizer: SGD with LR 0.01, momentum 0.9, weight_decay 1e-4
- Crop-predict: ImageSlicer from pytorch-toolbelt