This repository contains the codebase used to generate the main results of Mitigating Gradient Bias in Multi-objective Learning: A Provably Convergent Stochastic Approach, which has been accepted to ICLR 2023. This codebase is an extension to/ modification of Conflict-Averse Gradient Descent for Multitask Learning (CAGrad) and LibMTL codebases.
For running multi-task-supervised learning tasks (NYUv2 and Cityscapes datasets), follow the instructions in CAGrad/REAME.md.
For running multi-task-supervised learning tasks (Office-31 and Office-home datasets), follow the instructions in LibMTL/examples/office/REAME.md.