Skip to content

Scalarizer Tracker #667

@ValerianRey

Description

@ValerianRey

This issue tracks the candidate methods that could maybe be implemented as Scalarizer in torchjd.scalarization (see #666).

Status Name Ref Stateful Existing implementations Special Remarks
🟢 To do Sum - No (trivial)
🟢 To do Mean - No (trivial) Sometimes called Equal Weights (EW) in research papers.
🟢 To do Linear - No (trivial) Sometimes called Linear Scalarization (LS) in research papers. Name it Linear or Constant?
🟢 To do Random Reasonable Effectiveness of Random Weighting: A Litmus Test for Multi-Task Learning (TMLR 2022, 200 citations) No LibMTL (official) Sometimes called Random Loss Weighting (RLW) in research papers. Name it RLW or Random?
🔵 To investigate STCH (Smooth TCHebyshev) Smooth Tchebycheff Scalarization for Multi-Objective Optimization (ICML 2024, 87 citations) No official, LibMTL
🔵 To investigate GLS (Geometric Loss Strategy) MultiNet++: Multi-Stream Feature Aggregation and Geometric Loss Strategy for Multi-Task Learning (CVPR workshop 2019, 149 citations) No LibMTL
🔵 To investigate IMTL-L Towards Impartial Multi-Task Learning (ICLR 2021, 279 citations) ? official, LibMTL (maybe this is IMTL-G) Need more investigation
🔵 To investigate FAMO FAMO: Fast Adaptive Multitask Optimization (NeurIPS 2023, 124 citations) ? official, LibMTL Not sure this is a Scalarizer, need more investigation.
🔵 To investigate GradNorm GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks (ICML 2018, 2334 citations) Yes, trainable state unofficial, LibMTL Not sure this is a Scalarizer, need more investigation.
🔵 To investigate UW (UncertaintyWeighting) Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics (CVPR 2018, 5535 citations) ? unofficial, LibMTL Not sure this is a Scalarizer, need more investigation.
🔵 To investigate DWA (Dynamic Weight Averaging) End-to-End Multi-Task Learning with Attention (CVPR 2019, 1836 citations) ? official, LibMTL Not sure this is a Scalarizer, need more investigation.

Metadata

Metadata

Assignees

No one assigned

    Labels

    cc: featConventional commit type for new features.good first issueIssue that should be easy to solve for new contributorspackage: scalarization

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions