Scored 7th Place in the On Cloud N: Cloud Cover Detection Challenge

My main ideas that got me a high score were

• Significant cleaning of the dataset based on model disagreement and manual visual inspection
• Ensembling many models trained on different subsets of the data and different hyperparameter choices
• Exporting the trained models to JIT, making them more efficient and allowing me to use more models for the final predictions in the drivendata testing environment