Wasserstein Task Embedding for Measuring Task Similarities

This repository provides the implementation of WTE framework on CPU device. WTE is a model-agnostic task embedding framework for measuring task (dis)similarities in supervised classification problems, with linear computational cost with respect to number of tasks.

Datasets:

• torchvision datasets: MNIST, EMNIST, FashionMNIST, KMNIST, USPS and CIFAR100
• Tiny ImageNet

Code:

• demo_NIST.ipynb: notebook to reproduce WTE distances for *NIST task group.