feat(scalarization)!: Add UW

CHANGELOG.md

@@ -10,6 +10,12 @@ changelog does not include internal changes that do not affect the user.
 
 ### Added
 
+- Added `UW` (Uncertainty Weighting) from [Multi-Task Learning Using Uncertainty to Weigh Losses
+  for Scene Geometry and
+  Semantics](https://openaccess.thecvf.com/content_cvpr_2018/papers/Kendall_Multi-Task_Learning_Using_CVPR_2018_paper.pdf),
+  a `Scalarizer` that combines the values using learned per-task uncertainties. It is the first
+  stateful, trainable scalarizer: its log-variances are an `nn.Parameter` that must be passed to
+  the optimizer.
 - Added `STCH` from [Smooth Tchebycheff Scalarization for Multi-Objective
   Optimization](https://openreview.net/pdf?id=m4dO5L6eCp), a `Scalarizer` that combines the input
   tensor of values into a smooth approximation of their (weighted, shifted) maximum.
@@ -19,6 +25,12 @@ changelog does not include internal changes that do not affect the user.
   Learning](https://openaccess.thecvf.com/content_CVPRW_2019/papers/WAD/Chennupati_MultiNet_Multi-Stream_Feature_Aggregation_and_Geometric_Loss_Strategy_for_Multi-Task_CVPRW_2019_paper.pdf),
   a `Scalarizer` that returns the geometric mean of the input tensor of values.
 
+### Changed
+
+- **BREAKING**: Moved the `Stateful` mixin from `torchjd.aggregation` to the top-level `torchjd`
+  namespace, so it can be shared between the aggregation and scalarization packages. Import it as
+  `torchjd.Stateful` instead of `torchjd.aggregation.Stateful`.
+
 ## [0.12.0] - 2026-05-28
 
 ### Added

docs/source/docs/aggregation/index.rst

@@ -19,10 +19,6 @@ Abstract base classes
 .. autoclass:: torchjd.aggregation.Weighting
     :members: __call__
 
-.. autoclass:: torchjd.aggregation.Stateful
-    :members: reset
-
-
 .. toctree::
     :hidden:
     :maxdepth: 1

docs/source/docs/scalarization/index.rst

@@ -10,7 +10,6 @@ Abstract base class
 .. autoclass:: torchjd.scalarization.Scalarizer
     :members: __call__
 
-
 .. toctree::
     :hidden:
     :maxdepth: 1
@@ -21,3 +20,4 @@ Abstract base class
     random.rst
     stch.rst
     sum.rst
+    uw.rst

docs/source/docs/scalarization/uw.rst

@@ -0,0 +1,7 @@
+:hide-toc:
+
+UW
+==
+
+.. autoclass:: torchjd.scalarization.UW
+    :members: __call__, reset

docs/source/docs/stateful.rst

@@ -0,0 +1,4 @@
+:orphan:
+
+.. autoclass:: torchjd.Stateful
+    :members: reset

docs/source/examples/grouping.rst

@@ -19,8 +19,8 @@ the parameters:
 
 In TorchJD, grouping is achieved by calling :func:`~torchjd.autojac.jac_to_grad` once per group
 after :func:`~torchjd.autojac.backward` or :func:`~torchjd.autojac.mtl_backward`, with a dedicated
-aggregator instance per group. For :class:`~torchjd.aggregation.Stateful` aggregators, each instance
-should independently maintain its own state (e.g. the EMA :math:`\hat{\phi}` state in
+aggregator instance per group. For :class:`~torchjd.Stateful` aggregators, each instance should
+independently maintain its own state (e.g. the EMA :math:`\hat{\phi}` state in
 :class:`~torchjd.aggregation.GradVac`, matching the per-block targets from the original paper).
 
 .. note::

src/torchjd/__init__.py

@@ -1,13 +1,16 @@
 from collections.abc import Callable
 from warnings import warn as _warn
 
+from ._mixins import Stateful
 from .autojac import backward as _backward, mtl_backward as _mtl_backward
 
 _deprecated_items: dict[str, tuple[str, Callable]] = {
     "backward": ("autojac", _backward),
     "mtl_backward": ("autojac", _mtl_backward),
 }
 
+__all__ = ["Stateful"]
+
 
 def __getattr__(name: str) -> Callable:
     """

src/torchjd/_mixins.py

@@ -1,7 +1,16 @@
+from abc import ABC, abstractmethod
 from importlib.util import find_spec
 from typing import Any
 
 
+class Stateful(ABC):
+    """Mixin adding a reset method."""
+
+    @abstractmethod
+    def reset(self) -> None:
+        """Resets the internal state."""
+
+
 class _WithOptionalDeps:
     """
     Mixin that raises :class:`ImportError` at instantiation time if required optional dependencies

src/torchjd/aggregation/__init__.py

@@ -52,7 +52,6 @@
 from ._krum import Krum, KrumWeighting
 from ._mean import Mean, MeanWeighting
 from ._mgda import MGDA, MGDAWeighting
-from ._mixins import Stateful
 from ._modo import MoDoWeighting
 from ._nash_mtl import NashMTL
 from ._pcgrad import PCGrad, PCGradWeighting
@@ -94,7 +93,6 @@
     "PCGradWeighting",
     "Random",
     "RandomWeighting",
-    "Stateful",
     "Sum",
     "SumWeighting",
     "TrimmedMean",

src/torchjd/aggregation/_cr_mogm.py

@@ -4,7 +4,7 @@
 
 from torch import Tensor
 
-from torchjd.aggregation._mixins import Stateful
+from torchjd._mixins import Stateful
 
 from ._weighting_bases import Weighting
 
@@ -13,7 +13,7 @@
 
 class CRMOGMWeighting(Weighting[_T], Stateful):
     r"""
-    :class:`~torchjd.aggregation._mixins.Stateful`
+    :class:`~torchjd.Stateful`
     :class:`~torchjd.aggregation._weighting_bases.Weighting` that wraps another
     :class:`~torchjd.aggregation._weighting_bases.Weighting` and stabilises the weights it
     produces with an exponential moving average (EMA) across calls. This is the weight-smoothing
@@ -61,7 +61,7 @@ class CRMOGMWeighting(Weighting[_T], Stateful):
 
     This weighting is stateful: it keeps :math:`\lambda_{k-1}` across calls. Use :meth:`reset`
     to restart the smoothing from the initial state. Note that calling :meth:`reset` will also
-    reset the wrapped weighting if it is :class:`~torchjd.aggregation.Stateful`.
+    reset the wrapped weighting if it is :class:`~torchjd.Stateful`.
 
     :param weighting: The wrapped weighting whose output is smoothed.
     :param alpha: EMA coefficient on the previous weights. ``alpha=0`` disables smoothing
@@ -120,7 +120,7 @@ def alpha(self, value: float) -> None:
     def reset(self) -> None:
         r"""
         Clears the EMA state so the next forward restarts from the initial state. Also resets the
-        wrapped weighting if it is :class:`~torchjd.aggregation._mixins.Stateful`.
+        wrapped weighting if it is :class:`~torchjd.Stateful`.
         """
 
         if isinstance(self.weighting, Stateful):

src/torchjd/aggregation/_gradvac.py

@@ -5,7 +5,8 @@
 import torch
 from torch import Tensor
 
-from torchjd.aggregation._mixins import Stateful, _NonDifferentiable
+from torchjd._mixins import Stateful
+from torchjd.aggregation._mixins import _NonDifferentiable
 from torchjd.linalg import PSDMatrix
 
 from ._aggregator_bases import GramianWeightedAggregator
@@ -15,7 +16,7 @@
 # Non-differentiable: weights are modified in-place during the gradient correction loop.
 class GradVacWeighting(_GramianWeighting, Stateful, _NonDifferentiable):
     r"""
-    :class:`~torchjd.aggregation._mixins.Stateful`
+    :class:`~torchjd.Stateful`
     :class:`~torchjd.aggregation.Weighting` [:class:`~torchjd.linalg.PSDMatrix`]
     giving the weights of :class:`~torchjd.aggregation.GradVac`.
 
@@ -130,7 +131,7 @@ def _ensure_state(self, m: int, dtype: torch.dtype) -> None:
 
 class GradVac(GramianWeightedAggregator, Stateful, _NonDifferentiable):
     r"""
-    :class:`~torchjd.aggregation._mixins.Stateful`
+    :class:`~torchjd.Stateful`
     :class:`~torchjd.aggregation.GramianWeightedAggregator` implementing the aggregation step of
     Gradient Vaccine (GradVac) from `Gradient Vaccine: Investigating and Improving Multi-task
     Optimization in Massively Multilingual Models (ICLR 2021 Spotlight)

src/torchjd/aggregation/_mixins.py

@@ -1,18 +1,9 @@
-from abc import ABC, abstractmethod
 from typing import Any
 
 import torch
 from torch import nn
 
 
-class Stateful(ABC):
-    """Mixin adding a reset method."""
-
-    @abstractmethod
-    def reset(self) -> None:
-        """Resets the internal state."""
-
-
 class _NonDifferentiable(nn.Module):
     """
     Mixin making a nn.Module non-differentiable, preventing autograd graph construction by wrapping

src/torchjd/aggregation/_modo.py

@@ -7,15 +7,16 @@
 import torch
 from torch import Tensor
 
-from torchjd.aggregation._mixins import Stateful, _NonDifferentiable
+from torchjd._mixins import Stateful
+from torchjd.aggregation._mixins import _NonDifferentiable
 from torchjd.linalg import Matrix
 
 from ._weighting_bases import _MatrixWeighting
 
 
 class MoDoWeighting(_MatrixWeighting, Stateful, _NonDifferentiable):
     r"""
-    :class:`~torchjd.aggregation._mixins.Stateful`
+    :class:`~torchjd.Stateful`
     :class:`~torchjd.aggregation.Weighting` [:class:`~torchjd.linalg.Matrix`] from `Three-Way
     Trade-Off in Multi-Objective Learning: Optimization, Generalization and Conflict-Avoidance
     <https://www.jmlr.org/papers/volume25/23-1287/23-1287.pdf>`_ (JMLR 2024).

src/torchjd/aggregation/_nash_mtl.py

@@ -8,8 +8,8 @@
 import torch
 from torch import Tensor
 
-from torchjd._mixins import _WithOptionalDeps
-from torchjd.aggregation._mixins import Stateful, _NonDifferentiable
+from torchjd._mixins import Stateful, _WithOptionalDeps
+from torchjd.aggregation._mixins import _NonDifferentiable
 
 from ._aggregator_bases import WeightedAggregator
 from ._weighting_bases import _MatrixWeighting
@@ -25,7 +25,7 @@ class _NashMTLWeighting(_WithOptionalDeps, _MatrixWeighting, Stateful, _NonDiffe
     _REQUIRED_DEPS = ["numpy", "cvxpy", "ecos"]
     _INSTALL_HINT = 'Install them with: pip install "torchjd[nash_mtl]"'
     """
-    :class:`~torchjd.aggregation._mixins.Stateful`
+    :class:`~torchjd.Stateful`
     :class:`~torchjd.aggregation.Weighting` [:class:`~torchjd.linalg.Matrix`] that
     extracts weights using the step decision of Algorithm 1 of `Multi-Task Learning as a Bargaining
     Game <https://arxiv.org/pdf/2202.01017.pdf>`_.
@@ -206,7 +206,7 @@ def reset(self) -> None:
 
 class NashMTL(WeightedAggregator, Stateful, _NonDifferentiable):
     """
-    :class:`~torchjd.aggregation._mixins.Stateful`
+    :class:`~torchjd.Stateful`
     :class:`~torchjd.aggregation.WeightedAggregator` as proposed in Algorithm 1 of
     `Multi-Task Learning as a Bargaining Game <https://arxiv.org/pdf/2202.01017.pdf>`_.
 

src/torchjd/scalarization/__init__.py

@@ -26,5 +26,15 @@
 from ._scalarizer_base import Scalarizer
 from ._stch import STCH
 from ._sum import Sum
+from ._uw import UW
 
-__all__ = ["Constant", "GeometricMean", "Mean", "Random", "Scalarizer", "STCH", "Sum"]
+__all__ = [
+    "Constant",
+    "GeometricMean",
+    "Mean",
+    "Random",
+    "Scalarizer",
+    "STCH",
+    "Sum",
+    "UW",
+]

src/torchjd/scalarization/_uw.py

@@ -0,0 +1,81 @@
+from collections.abc import Sequence
+
+import torch
+from torch import Tensor, nn
+
+from torchjd._mixins import Stateful
+
+from ._scalarizer_base import Scalarizer
+
+
+class UW(Scalarizer, Stateful):
+    r"""
+    :class:`~torchjd.Stateful`
+    :class:`~torchjd.scalarization.Scalarizer` that combines the input tensor of values using
+    learned per-task uncertainties. ``UW`` is short for Uncertainty Weighting, the method proposed
+    in `Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics
+    <https://openaccess.thecvf.com/content_cvpr_2018/papers/Kendall_Multi-Task_Learning_Using_CVPR_2018_paper.pdf>`_.
+
+    Each value :math:`L_i` is assigned a learnable log-variance :math:`s_i`, and the values are
+    combined as
+
+    .. math::
+        \sum_i \left( \frac{1}{2} e^{-s_i} L_i + \frac{1}{2} s_i \right)
+
+    where:
+
+    - :math:`L_i` is the :math:`i`-th value (typically the loss of task :math:`i`);
+    - :math:`s_i = \log \sigma_i^2` is the learnable log-variance of task :math:`i`.
+
+    Following the paper, the log-variance :math:`s_i` is learned rather than the variance
+    :math:`\sigma_i^2` directly: this is numerically more stable (the combination never divides by
+    zero) and keeps :math:`s_i` unconstrained, since :math:`e^{-s_i}` is always positive. The
+    :math:`s_i` are stored as an ``nn.Parameter``, so the parameters of this scalarizer must be
+    passed to the optimizer to be learned jointly with the model.
+
+    :param shape: The shape of the values to scalarize, used to create one log-variance per value.
+        An ``int`` ``n`` is interpreted as the shape ``(n,)``.
+
+    The following example shows train a model with Uncertainty Weighting, as described in the paper.
+
+        >>> import torch
+        >>> from torch.nn import Linear
+        >>>
+        >>> from torchjd.scalarization import UW
+        >>>
+        >>> model = Linear(3, 2)
+        >>> scalarizer = UW(2)  # Move to the right device with e.g. UW(2).to(device="cuda")
+        >>> optimizer = torch.optim.SGD([*model.parameters(), *scalarizer.parameters()], lr=0.1)
+        >>>
+        >>> features = torch.randn(8, 3)
+        >>> # Compute some dummy losses just for the sake of the example
+        >>> losses = model(features).pow(2).mean(dim=0)  # One loss per output dimension.
+        >>> loss = scalarizer(losses)
+        >>> loss.backward()
+        >>> optimizer.step()
+
+    .. note::
+        The log-variances are initialized to ``0`` (i.e. :math:`\sigma_i^2 = 1`), which gives
+        uniform weights at the start of training. The paper reports that the result is robust to
+        this initialization. (`LibMTL <https://github.com/median-research-group/LibMTL>`_
+        initializes them to ``-0.5`` instead.)
+    """
+
+    def __init__(self, shape: int | Sequence[int]) -> None:
+        super().__init__()
+        self.log_var = nn.Parameter(torch.zeros(shape))
+
+    def forward(self, values: Tensor, /) -> Tensor:
+        if values.shape != self.log_var.shape:
+            raise ValueError(
+                f"Parameter `values` should have shape {tuple(self.log_var.shape)} (matching the "
+                f"shape of the log-variances). Found `values.shape = {tuple(values.shape)}`.",
+            )
+        return (0.5 * torch.exp(-self.log_var) * values + 0.5 * self.log_var).sum()
+
+    def reset(self) -> None:
+        with torch.no_grad():
+            self.log_var.zero_()
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(shape={tuple(self.log_var.shape)})"

tests/trajectories/optimize.py

@@ -19,7 +19,7 @@
 import torch
 
 from tests.paths import TRAJECTORIES_RESULTS_DIR
-from torchjd.aggregation import Stateful
+from torchjd._mixins import Stateful
 from trajectories._constants import (
     AGGREGATORS,
     BASE_LEARNING_RATES,