Main Architecture: Unet with ResNet34 encoder.
Training:
- RandomFlip augmentation
- Batch sizes: 64, 128, 256 with cycle length = 30
Cross Validation:
- 5 folds
- Test time augmentation with horizontal flip
Final Submission:
- Model Averaging of 5 models
Getting Started? You can check fast.ai lesson 14 where Jeremy Howard an amazing mentor and teacher shows how to solve a similar problem.