Work in Progress PyTorch version of the Dstl feature detection kaggle challenge on Kaggle.
The goal was to find ten potentially overlapping features (buildings, other structures, roads, tracks, trees, crops, rivers, lakes, trucks, cars) in satellite images. This solution uses the U-Net neural network architecture to segment the images for ten binary classes.
- Exploration_and_Setup.ipynb looks at the images and their labels and shows how cropping and augmentations look like
- Training.ipynb prepares and trains a U-Net model for one to ten classes in PyTorch
- While training loss curves and validation mask predictions are exported with visdom. Start the visdom server with
python -m visdom.server- Predict.ipynb predicts test image masks and saves them for later submission
- Evaluate_and_Submit.ipynb shows our results on the validation images and creates the final submission, which we also check.
