This repository contains code to train the image classification model using the semi-supervised approach known as Unsupervised Data Augmentation
I ported the initial implementation for images in Tensorflow to PyTorch, so this trainer uses:
- RandAugment to augment image data (using albumentations to improve the performance)
- Confidence-based masking
- Training Signal Annealing
- Predictions Sharpening using Softmax temperature
The code was primarily developed for the Wildlife Image Classification Competition, but may be pretty easily adapted to any other image classification task.