PyTorch image dataloaders and utility functions to load datasets for supervised continual learning
Introduction
This repository contains PyTorch image dataloaders and utility functions to load datasets for supervised continual learning. Currently supported datasets:
- MNIST
- Pairwise-MNIST
- Fashion-MNIST
- not-MNIST (letters version of MNIST, see EMNIST for more detail)
- CIFAR-10
- CIFAR-100
- German Traffic Signs
- Street View House Numbers (SVHN)
- Incremental CIFAR-100
- Incremental TinyImageNet
Features
The provided interface simplifies typical data loading for supervised continual learning scenarios.
-
Dataset order, additional training data (for replay buffers) and test data (for global metrics computation) can all be specified.
-
A batch balancing feature is also available to make sure data from all available classes are available in a training batch.
-
Training data size and channels can be specified.