Continual learning datasets

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. Transformations will be added to make sure input data always has the same size and number of channels. If a single channel is specified, grayscaling will be applied. Otherwise, if 3 channels are specified, single channels will be triplicated. Bicubic interpolation or linear subsampling will be applied to meet the specified size.

Installation

1. Clone the repository to your machine.
2. Install the package:

pip install -e cl_datasets/

Note: Please use Python 3.8 or above.

Example

from cl_datasets import getDatasets

datasets = ['svhn','cifar10','fashion','mnist']
batchSize = 32
dataSize = (32,32)
nChannels = 3

dataloaders = getDatasets(datasets,batchSize,dataSize,nChannels)

for train_test_loaders in dataloaders:
    trainLoader,testLoader = train_test_loaders
    ...

List of possible datasets for training tasks

Full datasets

Description	Dataset string
MNIST	"mnist" or "MNIST"
not-MNIST	"notMnist" or "notMNIST"
Fashion MNIST	"fashion"
SVHN	"svhn"
Cifar-10	"cifar10"
Cifar-100	"cifar100"
German traffic signs	"traffic"

Incremental datasets

Description	Dataset string
Pairwise MNIST	"mnist_xy" (e.g. "mnist_01")
Incremental Cifar-100 (10 classes per task)	"cifar100_i" (e.g. "cifar100_4")
Incremental Tiny ImageNet (10 classes per task)	"TIN_i" (e.g. "TIN_3")