Merge branch 'main' into check-todos

Author: ValerianRey

.pre-commit-config.yaml

@@ -18,7 +18,7 @@ repos:
         ]
 
 -   repo: https://github.com/pycqa/isort
-    rev: 6.1.0
+    rev: 7.0.0
     hooks:
     -   id: isort  # Sort imports.
         args: [
@@ -31,7 +31,7 @@ repos:
         ]
 
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 25.9.0
+    rev: 25.12.0
     hooks:
     -   id: black  # Format code.
         args: [--line-length=100]

CHANGELOG.md

@@ -3,14 +3,47 @@
 All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). This changelog does not include internal
-changes that do not affect the user.
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). This
+changelog does not include internal changes that do not affect the user.
 
 ## [Unreleased]
 
+### Changed
+
+- **BREAKING**: Removed from `backward` and `mtl_backward` the responsibility to aggregate the
+  Jacobian. Now, these functions compute and populate the `.jac` fields of the parameters, and a new
+  function `torchjd.autojac.jac_to_grad` should then be called to aggregate those `.jac` fields into
+  `.grad` fields.
+  This means that users now have more control on what they do with the Jacobians (they can easily
+  aggregate them group by group or even param by param if they want), but it now requires an extra
+  line of code to do the Jacobian descent step. To update, please change:
+  ```python
+  backward(losses, aggregator)
+  ```
+  to
+  ```python
+  backward(losses)
+  jac_to_grad(model.parameters(), aggregator)
+  ```
+  and
+  ```python
+  mtl_backward(losses, features, aggregator)
+  ```
+  to
+  ```python
+  mtl_backward(losses, features)
+  jac_to_grad(shared_module.parameters(), aggregator)
+  ```
+
+- Removed an unnecessary internal cloning of gradient. This should slightly improve the memory
+  efficiency of `autojac`.
+
+## [0.8.1] - 2026-01-07
+
 ### Added
 
-- Added `__all__` in the `__init__.py` of packages. This should prevent PyLance from triggering warnings when importing from `torchjd`.
+- Added `__all__` in the `__init__.py` of packages. This should prevent PyLance from triggering
+  warnings when importing from `torchjd`.
 
 ## [0.8.0] - 2025-11-13
 

README.md

@@ -111,11 +111,11 @@ using [UPGrad](https://torchjd.org/stable/docs/aggregation/upgrad/).
       loss1 = loss_fn(output1, target1)
       loss2 = loss_fn(output2, target2)
 
-      optimizer.zero_grad()
 -     loss = loss1 + loss2
 -     loss.backward()
 +     mtl_backward(losses=[loss1, loss2], features=features, aggregator=aggregator)
       optimizer.step()
+      optimizer.zero_grad()
 ```
 
 > [!NOTE]
@@ -150,12 +150,12 @@ Jacobian descent using [UPGrad](https://torchjd.org/stable/docs/aggregation/upgr
 -     loss = loss_fn(output, target)  # shape [1]
 +     losses = loss_fn(output, target)  # shape [16]
 
-      optimizer.zero_grad()
 -     loss.backward()
 +     gramian = engine.compute_gramian(losses)  # shape: [16, 16]
 +     weights = weighting(gramian)  # shape: [16]
 +     losses.backward(weights)
       optimizer.step()
+      optimizer.zero_grad()
 ```
 
 Lastly, you can even combine the two approaches by considering multiple tasks and each element of
@@ -201,10 +201,10 @@ for input, target1, target2 in zip(inputs, task1_targets, task2_targets):
     # Obtain the weights that lead to no conflict between reweighted gradients
     weights = weighting(gramian)  # shape: [16, 2]
 
-    optimizer.zero_grad()
     # Do the standard backward pass, but weighted using the obtained weights
     losses.backward(weights)
     optimizer.step()
+    optimizer.zero_grad()
 ```
 
 > [!NOTE]

docs/source/docs/autojac/index.rst

@@ -10,3 +10,4 @@ autojac
 
     backward.rst
     mtl_backward.rst
+    jac_to_grad.rst

docs/source/docs/autojac/jac_to_grad.rst

@@ -0,0 +1,6 @@
+:hide-toc:
+
+jac_to_grad
+===========
+
+.. autofunction:: torchjd.autojac.jac_to_grad

docs/source/examples/amp.rst

@@ -12,15 +12,15 @@ case, the losses) should preferably be scaled with a `GradScaler
 following example shows the resulting code for a multi-task learning use-case.
 
 .. code-block:: python
-    :emphasize-lines: 2, 17, 27, 34, 36-38
+    :emphasize-lines: 2, 17, 27, 34-35, 37-38
 
     import torch
     from torch.amp import GradScaler
     from torch.nn import Linear, MSELoss, ReLU, Sequential
     from torch.optim import SGD
 
     from torchjd.aggregation import UPGrad
-    from torchjd.autojac import mtl_backward
+    from torchjd.autojac import mtl_backward, jac_to_grad
 
     shared_module = Sequential(Linear(10, 5), ReLU(), Linear(5, 3), ReLU())
     task1_module = Linear(3, 1)
@@ -48,10 +48,11 @@ following example shows the resulting code for a multi-task learning use-case.
             loss2 = loss_fn(output2, target2)
 
         scaled_losses = scaler.scale([loss1, loss2])
-        optimizer.zero_grad()
-        mtl_backward(losses=scaled_losses, features=features, aggregator=aggregator)
+        mtl_backward(losses=scaled_losses, features=features)
+        jac_to_grad(shared_module.parameters(), aggregator)
         scaler.step(optimizer)
         scaler.update()
+        optimizer.zero_grad()
 
 .. hint::
     Within the ``torch.autocast`` context, some operations may be done in ``float16`` type. For

docs/source/examples/basic_usage.rst

@@ -20,6 +20,7 @@ Import several classes from ``torch`` and ``torchjd``:
 
     from torchjd import autojac
     from torchjd.aggregation import UPGrad
+    from torchjd.autojac import jac_to_grad
 
 Define the model and the optimizer, as usual:
 
@@ -59,20 +60,16 @@ We can now compute the losses associated to each element of the batch.
 
 The last steps are similar to gradient descent-based optimization, but using the two losses.
 
-Reset the ``.grad`` field of each model parameter:
-
-.. code-block:: python
-
-    optimizer.zero_grad()
-
 Perform the Jacobian descent backward pass:
 
 .. code-block:: python
 
-    autojac.backward([loss1, loss2], aggregator)
+    autojac.backward([loss1, loss2])
+    jac_to_grad(model.parameters(), aggregator)
 
-This will populate the ``.grad`` field of each model parameter with the corresponding aggregated
-Jacobian matrix.
+The first function will populate the ``.jac`` field of each model parameter with the corresponding
+Jacobian, and the second one will aggregate these Jacobians and store the result in the ``.grad``
+field of the parameters. It also deletes the ``.jac`` fields save some memory.
 
 Update each parameter based on its ``.grad`` field, using the ``optimizer``:
 
@@ -81,3 +78,9 @@ Update each parameter based on its ``.grad`` field, using the ``optimizer``:
     optimizer.step()
 
 The model's parameters have been updated!
+
+As usual, you should now reset the ``.grad`` field of each model parameter:
+
+.. code-block:: python
+
+    optimizer.zero_grad()

docs/source/examples/iwmtl.rst

@@ -10,7 +10,7 @@ this Gramian to reweight the gradients and resolve conflict entirely.
 The following example shows how to do that.
 
 .. code-block:: python
-    :emphasize-lines: 5-6, 18-20, 31-32, 34-35, 37-38, 41-42
+    :emphasize-lines: 5-6, 18-20, 31-32, 34-35, 37-38, 40-41
 
     import torch
     from torch.nn import Linear, MSELoss, ReLU, Sequential
@@ -51,10 +51,10 @@ The following example shows how to do that.
         # Obtain the weights that lead to no conflict between reweighted gradients
         weights = weighting(gramian)  # shape: [16, 2]
 
-        optimizer.zero_grad()
         # Do the standard backward pass, but weighted using the obtained weights
         losses.backward(weights)
         optimizer.step()
+        optimizer.zero_grad()
 
 .. note::
     In this example, the tensor of losses is a matrix rather than a vector. The gramian is thus a

docs/source/examples/iwrm.rst

@@ -64,26 +64,26 @@ batch of data. When minimizing per-instance losses (IWRM), we use either autojac
             for x, y in zip(X, Y):
                 y_hat = model(x).squeeze(dim=1)  # shape: [16]
                 loss = loss_fn(y_hat, y)  # shape: [] (scalar)
-                optimizer.zero_grad()
                 loss.backward()
 
 
                 optimizer.step()
+                optimizer.zero_grad()
 
         In this baseline example, the update may negatively affect the loss of some elements of the
         batch.
 
     .. tab-item:: autojac
 
         .. code-block:: python
-            :emphasize-lines: 5-6, 12, 16, 21, 23
+            :emphasize-lines: 5-6, 12, 16, 21-23
 
             import torch
             from torch.nn import Linear, MSELoss, ReLU, Sequential
             from torch.optim import SGD
 
             from torchjd.aggregation import UPGrad
-            from torchjd.autojac import backward
+            from torchjd.autojac import backward, jac_to_grad
 
             X = torch.randn(8, 16, 10)
             Y = torch.randn(8, 16)
@@ -99,19 +99,19 @@ batch of data. When minimizing per-instance losses (IWRM), we use either autojac
             for x, y in zip(X, Y):
                 y_hat = model(x).squeeze(dim=1)  # shape: [16]
                 losses = loss_fn(y_hat, y)  # shape: [16]
-                optimizer.zero_grad()
-                backward(losses, aggregator)
-
+                backward(losses)
+                jac_to_grad(model.parameters(), aggregator)
 
                 optimizer.step()
+                optimizer.zero_grad()
 
         Here, we compute the Jacobian of the per-sample losses with respect to the model parameters
         and use it to update the model such that no loss from the batch is (locally) increased.
 
     .. tab-item:: autogram (recommended)
 
         .. code-block:: python
-            :emphasize-lines: 5-6, 12, 16-17, 21, 23-25
+            :emphasize-lines: 5-6, 12, 16-17, 21-24
 
             import torch
             from torch.nn import Linear, MSELoss, ReLU, Sequential
@@ -134,11 +134,11 @@ batch of data. When minimizing per-instance losses (IWRM), we use either autojac
             for x, y in zip(X, Y):
                 y_hat = model(x).squeeze(dim=1)  # shape: [16]
                 losses = loss_fn(y_hat, y)  # shape: [16]
-                optimizer.zero_grad()
                 gramian = engine.compute_gramian(losses)  # shape: [16, 16]
                 weights = weighting(gramian)  # shape: [16]
                 losses.backward(weights)
                 optimizer.step()
+                optimizer.zero_grad()
 
         Here, the per-sample gradients are never fully stored in memory, leading to large
         improvements in memory usage and speed compared to autojac, in most practical cases. The

docs/source/examples/lightning_integration.rst

@@ -11,7 +11,7 @@ The following code example demonstrates a basic multi-task learning setup using
 <../docs/autojac/mtl_backward>` at each training iteration.
 
 .. code-block:: python
-    :emphasize-lines: 9-10, 18, 32
+    :emphasize-lines: 9-10, 18, 31-32
 
     import torch
     from lightning import LightningModule, Trainer
@@ -22,7 +22,7 @@ The following code example demonstrates a basic multi-task learning setup using
     from torch.utils.data import DataLoader, TensorDataset
 
     from torchjd.aggregation import UPGrad
-    from torchjd.autojac import mtl_backward
+    from torchjd.autojac import mtl_backward, jac_to_grad
 
     class Model(LightningModule):
         def __init__(self):
@@ -43,9 +43,10 @@ The following code example demonstrates a basic multi-task learning setup using
             loss2 = mse_loss(output2, target2)
 
             opt = self.optimizers()
-            opt.zero_grad()
-            mtl_backward(losses=[loss1, loss2], features=features, aggregator=UPGrad())
+            mtl_backward(losses=[loss1, loss2], features=features)
+            jac_to_grad(self.feature_extractor.parameters(), UPGrad())
             opt.step()
+            opt.zero_grad()
 
         def configure_optimizers(self) -> OptimizerLRScheduler:
             optimizer = Adam(self.parameters(), lr=1e-3)