This solution was based on Heng CherKeng's code for PyTorch. I kindly thank him for sharing his work. 128x128, 256x256 and 512x512 U-nets are implemented. 128x128 U-net gets an LB score of 0.988.
- Keras 2.0 w/ TF backend
- sklearn
- cv2
- tqdm
Place 'train', 'train_masks' and 'test' data folders in the 'input' folder.
Convert training masks to .png format. You can do this with:
mogrify -format png *.gif
in the 'train_masks' data folder.
Run python train.py
to train from scratch. Alternatively, download pre-trained weights (for 128x128 U-net) into 'weights' folder.
Run python test_submit.py
to make predictions on test data and generate submission.