Note:
Approach:
- Converted Original data to TFRecords, for using the Kaggle's TPU using this
- All images were scaled to 299x299 considering majority of the images had similar resolution.
- Model used EfficientNet-B4 (noisy student weights) freezed BatchNorm layers.
- Augmentations used:
- Horizontal Fips
- Shearing
- H/W Zoom
- H/W Shift
- Custom Cosine Annealed Learning Schedule
- 5-Fold Cross Validation pipeline.
- Class weights to combat class-imbalance.
Will be making the kaggle notebook public soon!