Transfer Learning approach using fast.ai library which makes implementing it easier. Based on 3 different approaches each with architectures- resnet34, resnet50 and resnet101... got top 5% on Kaggle leaderboard, Accuracy 99.3% and and 0.05605 binary log loss error(evaluation criteria).
Used Diffrential Learning Rates to tune arch , Test Time Augmentation and Learning Rate Anneling to improve model loss.