From f2327a15026ac0713226382a5a0b21cf7690996f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Val=C3=A9rian=20Rey?= Date: Mon, 15 Jun 2026 16:29:52 +0200 Subject: [PATCH] Move DWA to [Unreleased] in changelog --- CHANGELOG.md | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index eaf30c8b..44283b53 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -10,6 +10,11 @@ changelog does not include internal changes that do not affect the user. ### Added +- Added `DWA` (Dynamic Weight Average) from [End-to-End Multi-Task Learning with + Attention](https://openaccess.thecvf.com/content_CVPR_2019/papers/Liu_End-To-End_Multi-Task_Learning_With_Attention_CVPR_2019_paper.pdf) + (CVPR 2019), a stateful `Scalarizer` that weights each value by the relative rate at which its + loss decreased over the two previous epochs. It has no learnable parameters; call its `step()` + method once per epoch to roll the loss history. - Added `FAMO` (Fast Adaptive Multitask Optimization) from [FAMO: Fast Adaptive Multitask Optimization](https://proceedings.neurips.cc/paper_files/paper/2023/file/b2fe1ee8d936ac08dd26f2ff58986c8f-Paper-Conference.pdf) (NeurIPS 2023), a stateful `Scalarizer` that decreases all task losses at an approximately equal @@ -20,11 +25,6 @@ changelog does not include internal changes that do not affect the user. ### Added -- Added `DWA` (Dynamic Weight Average) from [End-to-End Multi-Task Learning with - Attention](https://openaccess.thecvf.com/content_CVPR_2019/papers/Liu_End-To-End_Multi-Task_Learning_With_Attention_CVPR_2019_paper.pdf) - (CVPR 2019), a stateful `Scalarizer` that weights each value by the relative rate at which its - loss decreased over the two previous epochs. It has no learnable parameters; call its `step()` - method once per epoch to roll the loss history. - Added `SDMGradWeighting` from [Direction-oriented Multi-objective Learning: Simple and Provable Stochastic Algorithms](https://arxiv.org/pdf/2305.18409) (NeurIPS 2023). It is a stateful `Weighting` that solves for task weights via a simplex-projected inner loop on a cross-batch matrix `A = J_1 @ J_2.T` (computed from two independent mini-batches using `autojac.jac`), with a direction-oriented regularizer pulling the descent direction toward a preference direction. - Added `IMTL-L` (the loss-balancing variant of Impartial Multi-Task Learning) from [Towards Impartial Multi-Task Learning](https://openreview.net/pdf?id=IMPnRXEWpvr) (ICLR 2021), a stateful