v0.17.0

🍍 ExcessMTL

This release introduces ExcessMTL and ExcessMTLWeighting. Thanks a lot to @KhusPatel4450 for the contribution!

We're trying to grow the community and build even more features! To participate, you can join the Discord community!

Changelog

Added

• Added ExcessMTL and ExcessMTLWeighting from Robust Multi-Task Learning with Excess Risks (ICML 2024). ExcessMTLWeighting is a stateful Weighting that maintains task weights across calls via an exponentiated gradient update driven by per-task excess risk estimates. The excess risk is approximated using an AdaGrad-style diagonal Hessian. An optional n_warmup_steps parameter controls how many forward calls collect gradient statistics before weight updates begin.