🍉 UW, IMTL-L
This release introduces UW and IMTL-L, two very similar scalarization methods. Thanks a lot to @ppraneth for the contributions!
We also thank @KhusPatel4450 for becoming code owner of the aggregation package and @ppraneth for becoming code owner of the scalarization package. This should make the project slightly more decentralized.
Changelog
Added
- Added
IMTL-L(the loss-balancing variant of Impartial Multi-Task Learning) from Towards
Impartial Multi-Task Learning (ICLR 2021), a stateful
Scalarizerthat learns a per-task scales_iand combines the values as
Σ (exp(s_i) · L_i − s_i). - Added
UW(Uncertainty Weighting) from Multi-Task Learning Using Uncertainty to Weigh Losses
for Scene Geometry and
Semantics,
aScalarizerthat combines the values using learned per-task uncertainties. It is the first
stateful, trainable scalarizer: its log-variances are annn.Parameterthat must be passed to
the optimizer.