[WIP] DomainAwareNet supports regression #338

vloison · 2025-06-25T15:06:10Z

DomainAwareNet implements a classifier. The goal is to implement regression too.

To do so, I replaced DomainAwareNet with a mother class _DomainAwareNet and two classes DomainAwareNetRegressor and DomainNetClassifier that inherit from _DomainAwareNet.
This is complementary to Arthur's work on Issue 249 . He will add a DomainNetBinaryClassifier, which will also inherit from _DomainAwareNet.

Done:

Separate three classes.
Mark DomainAwareNet as deprecated.
Replace calls to DomainAwareNet by calls to DomainAwareNetClassifier in the module.

To do:

a script to try DomainAwareNetRegressor.

Bonus: fixed tests in test_utils.py, they were too restrictive on the error message and caused the tests to fail, although the code had the expected behavior.

codecov · 2025-06-25T15:31:33Z

Codecov Report

❌ Patch coverage is 84.64730% with 37 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.80%. Comparing base (8ccfd75) to head (4d2b5af).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #338      +/-   ##
==========================================
- Coverage   96.19%   94.80%   -1.39%     
==========================================
  Files          63       51      -12     
  Lines        7020     6223     -797     
==========================================
- Hits         6753     5900     -853     
- Misses        267      323      +56

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

…guity

…regnet

antoinecollas

Thx for the PR! It's on the right track!

The nexts steps are:

I noticed many diffs that look like autoformatting. Do you have such a tool in your editor? You should revert these modifications.
Add tests for the regressor

antoinecollas · 2025-06-26T09:00:25Z

skada/deep/_adversarial.py

-    """Modified CrossEntropyLoss as described in (29) from [35]_ with label smoothing.
+    """Modified CrossEntropyLoss as described in (29) from [35]_ with label
+    smoothing.


Why is this docstring now on two lines?

Because the line seemed too long, do you want me to revert it?

antoinecollas · 2025-06-26T09:00:34Z

skada/deep/_adversarial.py

-            The modified CrossEntropyLoss with label smoothing applied to predictions.
+            The modified CrossEntropyLoss with label smoothing applied to
+            predictions.


antoinecollas · 2025-06-26T09:11:10Z

skada/deep/base.py

-_EMPTY_INT_ = torch.tensor([],dtype=torch.int64)
+_EMPTY_INT_ = torch.tensor([], dtype=torch.int64)


seems like a formatting typo

antoinecollas · 2025-06-26T09:11:19Z

skada/deep/base.py

 _DEFAULT_SAMPLE_DOMAIN_ = 0
 _NO_LABEL_ = -1

+


antoinecollas · 2025-06-26T09:11:42Z

skada/deep/base.py

        pass


-


antoinecollas · 2025-06-26T09:21:44Z

skada/deep/base.py

-            A dictionary containing 'X', 'sample_domain', and optionally 'y' and 'sample_weight' as numpy arrays.
+            A dictionary containing 'X', 'sample_domain', and optionally 'y'
+            and 'sample_weight' as numpy arrays.


antoinecollas · 2025-06-26T09:23:57Z

skada/deep/base.py

+    """
+    A domain-aware neural network classifier with sample weight support.
+
+    This class extends NeuralNetClassifier to handle domain-specific input data


skorch.classifier.NeuralNetClassifier

skada/deep/base.py

antoinecollas · 2025-06-26T09:25:34Z

skada/deep/base.py

+    def predict_proba(
+        self,
+        X: Union[Dict, torch.Tensor, np.ndarray, Dataset],
+        sample_domain: Union[torch.Tensor, np.ndarray] = None,
+        sample_weight: Union[torch.Tensor, np.ndarray] = None,
+        allow_source: bool = False,
+        **predict_params
+    ):
+        return super().predict_DA_proba(
+            X,
+            sample_domain,
+            sample_weight,
+            allow_source,
+            **predict_params
+        )


I think you do not want to predict probabilities with a regressor right?

Indeed, conceptually, we don't. In the case of a regressor, predict_proba returns the same as predict. I kept it this way to be consistent with the skorch.regressor API.

antoinecollas · 2025-06-26T09:27:18Z

skada/deep/base.py

-        X = check_array(X, 
-                ensure_2d=False,
-                allow_nd=True,
-                ensure_min_samples=0,
-                ensure_min_features=0,
-                )
+        X = check_array(


In the remaining of the file, you have many modifications just because of autoformatting.

Not sure I understand why autoformatting would need to be reverted?

antoinecollas · 2025-06-26T09:37:01Z

Also, please add the DomainAwareNet class, wrapping DomainAwareNetClassifier and marking it as deprecated, for consistency with previous releases.

You can use the decorator @deprecated to do this and put a warning in the docstring.

tgnassou · 2025-07-02T20:25:10Z

skada/deep/_adversarial.py

    if base_criterion is None:
        base_criterion = torch.nn.CrossEntropyLoss()

-    net = DomainAwareNet(


For the next week should have a parameter to chose between classifier, regression on binary classification

The reasoning was to have one class per task, and users would call one class depending on their use case. This parameter option seems nice though :)

tgnassou · 2025-07-02T20:28:23Z

skada/deep/base.py



-class DomainAwareNet(NeuralNetClassifier, _DAMetadataRequesterMixin):
+class _DomainAwareNet(NeuralNet, _DAMetadataRequesterMixin):


I think we dont need the underscore here. It could be use by user

Not sure when the user would need to call _DomainAwareNet. The reasoning is that the user would call one of DomainAwareNetClassifier, DomainAwareNetBinaryClassifier, or DomainAwareNetRegressor, depending on their intended use. These three classes inherit from _DomainAwareNet to avoid redundancy.

tgnassou · 2025-07-02T20:31:32Z

skada/deep/base.py

        return nonlin

-    def predict_proba(
+    def predict_DA_proba(


Yes, I think you remove it and add it only to DomainAwareNetClassiffier

vloison · 2025-09-02T08:56:23Z

Also, please add the DomainAwareNet class, wrapping DomainAwareNetClassifier and marking it as deprecated, for consistency with previous releases.

You can use the decorator @deprecated to do this and put a warning in the docstring.

Done. I didn't know this decorator, thanks for the discovery!

vloison · 2025-09-02T09:47:12Z

Based on the comments, it seems like a couple of choices need to be made.

Do we want to follow the skorch API (implemented now) or remove predict_proba from the regressor?
Do we want the user to call a different class based on their use case, like in skorch (implemented now), or to call a single class with a parameter for the use case? I think the second option will probably be more code-heavy, but maybe more user-friendly.

All options seem good, let me know what you want :)

vloison and others added 7 commits June 25, 2025 15:06

move fit method back to _DomainAwareNet

44eb8ef

format

9c92ce5

add DomainAwareNetRegressor methods

d40f116

Merge branch 'scikit-adaptation:main' into deepregnet

ac728ff

fix check_X_domain test

31a35e1

Replace calls to DomainAwareNet by DomainAwareNetClassifier

91f1000

change _DomainAwareNet.get_proba name to avoid inheritance conflict

890584d

vloison and others added 4 commits June 25, 2025 17:44

rewrite classifier and regressor predict_proba to avoid heritage ambi…

1b23d8f

…guity

Merge branch 'main' into deepregnet

ec88d55

add init

a583ace

Merge remote-tracking branch 'refs/remotes/vivi/deepregnet' into deep…

ed059e0

…regnet

antoinecollas self-requested a review June 26, 2025 08:11

Merge branch 'main' into deepregnet

04bf47b

antoinecollas reviewed Jun 26, 2025

View reviewed changes

antoinecollas changed the title ~~[ENH] DomainAwareNet supports regression~~ [WIP] DomainAwareNet supports regression Jun 26, 2025

antoinecollas and others added 2 commits July 1, 2025 08:54

Merge branch 'main' into deepregnet

e490a6c

add binaryclassifier

60e6e16

arthurdrk mentioned this pull request Jul 1, 2025

[WIP] add binaryclassifier vloison/skada#1

Closed

fix

6d17ace

tgnassou reviewed Jul 2, 2025

View reviewed changes

Merge branch 'main' into deepregnet

930f1f8

github-actions bot added deep shallow tests-shallow tests-deep labels Jul 15, 2025

correct skorch references

9d2d420

add deprecated domainawarenet

27d5fc4

add old comment

5045c95

vloison and others added 5 commits September 2, 2025 16:10

fix DomainAwareNet deprecator

b647ad6

add deprecated to pyproject.toml

ee161cc

Merge branch 'main' into deepregnet

6ad67a4

Merge branch 'main' into deepregnet

d4f8855

fix inheritance issue with predict_proba

4d2b5af

		_EMPTY_INT_ = torch.tensor([],dtype=torch.int64)
		_EMPTY_INT_ = torch.tensor([], dtype=torch.int64)



		class DomainAwareNet(NeuralNetClassifier, _DAMetadataRequesterMixin):
		class _DomainAwareNet(NeuralNet, _DAMetadataRequesterMixin):

[WIP] DomainAwareNet supports regression #338

Are you sure you want to change the base?

[WIP] DomainAwareNet supports regression #338

Uh oh!

Conversation

vloison commented Jun 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Jun 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

antoinecollas left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

antoinecollas commented Jun 26, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

vloison commented Sep 2, 2025

Uh oh!

vloison commented Sep 2, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

vloison commented Jun 25, 2025 •

edited

Loading

codecov bot commented Jun 25, 2025 •

edited

Loading