Merge branch 'main' into release

Author: tmke8

.github/workflows/continuous_integration.yml

@@ -19,6 +19,7 @@ on:
     - LICENSE
     - make_release.sh
     - CITATION.cff
+  merge_group:
 
 
 jobs:
@@ -39,10 +40,10 @@ jobs:
         pip install ruff
     - name: Lint with ruff
       run: |
-        ruff ethicml
+        ruff check --format=github ethicml
     - name: Lint with ruff
       run: |
-        ruff tests
+        ruff check --format=github tests
 
   format_with_black:
 
@@ -149,8 +150,10 @@ jobs:
       #----------------------------------------------
       - uses: actions/checkout@v3
       - name: Install poetry
+        if: ${{ github.event_name == 'merge_group' }}
         run: pipx install poetry
       - uses: actions/setup-python@v4
+        if: ${{ github.event_name == 'merge_group' }}
         with:
           python-version: '3.8'
           cache: 'poetry'
@@ -159,6 +162,7 @@ jobs:
       #  ---------  install dependencies  --------
       #----------------------------------------------
       - name: Install dependencies
+        if: ${{ github.event_name == 'merge_group' }}
         run: |
           # keep the following in sync with `test_full_dependencies`!
           poetry env use 3.8
@@ -168,6 +172,7 @@ jobs:
       #  -----           Run MyPy             -----
       #----------------------------------------------
       - name: Type check with mypy
+        if: ${{ github.event_name == 'merge_group' }}
         run: |
           poetry run python run_mypy.py
           poetry run python run_mypy_tests.py
@@ -176,5 +181,6 @@ jobs:
       #  -----           Run Tests            -----
       #----------------------------------------------
       - name: Test with pytest
+        if: ${{ github.event_name == 'merge_group' }}
         run: |
           poetry run python -m pytest -vv -n 2 --dist loadgroup --cov=ethicml --cov-fail-under=80  tests/

ethicml/common.py

@@ -1,11 +1,9 @@
 """Common variables / constants that make things run smoother."""
 from importlib import util
-import os
 from pathlib import Path
 
-__all__ = ["TORCH_AVAILABLE", "ROOT_DIR", "ROOT_PATH"]
+__all__ = ["TORCH_AVAILABLE", "ROOT_PATH"]
 
 TORCH_AVAILABLE = util.find_spec("torch") is not None
 
-ROOT_DIR: str = os.path.abspath(os.path.join(os.path.abspath(__file__), os.pardir))
 ROOT_PATH: Path = Path(__file__).parent.resolve()

ethicml/data/csvs/make_adult_from_raw.py

@@ -36,7 +36,7 @@ def run_generate_adult() -> None:
     all_data = pd.concat([train, test], axis=0)
 
     for col in all_data.columns:
-        if all_data[col].dtype == np.object:  # type: ignore[attr-defined]
+        if all_data[col].dtype == object:
             all_data[col] = all_data[col].str.strip()
 
     # Replace full stop in the label of the test set

ethicml/data/csvs/make_crime_from_raw.py

@@ -160,7 +160,7 @@ def run_generate_crime() -> None:
     data.columns = pd.Index(columns)
 
     for col in data.columns:
-        if data[col].dtype == np.object:  # type: ignore[attr-defined]
+        if data[col].dtype == object:
             data[col] = data[col].str.strip()
 
     # Drop NaNs

ethicml/data/dataset.py

@@ -9,7 +9,7 @@
 from typing_extensions import override
 
 import pandas as pd
-from ranzen import StrEnum
+from ranzen.misc import StrEnum
 
 from ethicml.common import ROOT_PATH
 from ethicml.utility import DataTuple, undo_one_hot
@@ -417,6 +417,7 @@ class LegacyDataset(CSVDataset):
 
     def __init__(
         self,
+        *,
         name: str,
         filename_or_path: str | Path,
         features: Sequence[str],

ethicml/data/tabular_data/acs.py

@@ -89,7 +89,7 @@
 
 @contextlib.contextmanager
 def _download_dir(root: Path) -> Generator[None, None, None]:
-    curdir = os.getcwd()
+    curdir = Path.cwd()
     os.chdir(root.expanduser().resolve())
     try:
         yield

ethicml/implementations/adv_debiasing_modules/model.py

@@ -192,6 +192,7 @@ class AdvDebiasingClassLearner:
 
     def __init__(
         self,
+        *,
         lr: float,
         n_clf_epochs: int,
         n_adv_epochs: int,
@@ -231,9 +232,9 @@ def __init__(
 
         self.n_epoch_combined = n_epoch_combined
 
-    def fit(self, train: DataTuple, seed: int) -> Self:  # type: ignore[valid-type]
+    def fit(self, train: DataTuple, seed: int) -> Self:
         """Fit."""
-        train_data, train_loader = make_dataset_and_loader(
+        _, train_loader = make_dataset_and_loader(
             train, batch_size=self.batch_size, shuffle=True, seed=seed, drop_last=True
         )
 
@@ -268,9 +269,9 @@ def fit(self, train: DataTuple, seed: int) -> Self:  # type: ignore[valid-type]
     @torch.no_grad()
     def predict(self, x: pd.DataFrame) -> np.ndarray:
         """Predict."""
-        x = torch.from_numpy(x.to_numpy()).float()
+        x_ = torch.from_numpy(x.to_numpy()).float()
         self.clf.eval()
-        yhat = self.clf(x)
+        yhat = self.clf(x_)
         sm = nn.Softmax(dim=1)
         yhat = sm(yhat)
         yhat = yhat.detach().numpy()
@@ -283,6 +284,7 @@ class AdvDebiasingRegLearner:
 
     def __init__(
         self,
+        *,
         lr: float,
         n_clf_epochs: int,
         n_adv_epochs: int,
@@ -322,7 +324,7 @@ def __init__(
 
         self.n_epoch_combined = n_epoch_combined
 
-    def fit(self, train: DataTuple, seed: int) -> Self:  # type: ignore[valid-type]
+    def fit(self, train: DataTuple, seed: int) -> Self:
         """Fit."""
         # The features are X[:,1:]
 
@@ -361,9 +363,9 @@ def fit(self, train: DataTuple, seed: int) -> Self:  # type: ignore[valid-type]
     @torch.no_grad()
     def predict(self, x: pd.DataFrame) -> torch.Tensor:
         """Predict."""
-        x = torch.from_numpy(x.to_numpy()).float()
+        x_ = torch.from_numpy(x.to_numpy()).float()
         self.clf.eval()
-        yhat = self.clf(x).squeeze().detach().numpy()
+        yhat = self.clf(x_).squeeze().detach().numpy()
         if self.out_shape == 1:
             out = yhat
         else:

ethicml/implementations/agarwal.py

@@ -6,7 +6,7 @@
 from pathlib import Path
 import random
 import sys
-from typing import TYPE_CHECKING, Generator
+from typing import TYPE_CHECKING, Generator, Union
 
 from joblib import dump, load
 import numpy as np
@@ -26,16 +26,17 @@
 )
 
 if TYPE_CHECKING:
-    from fairlearn.reductions import ExponentiatedGradient
+    from fairlearn.reductions import ExponentiatedGradient  # pyright: ignore
 
     from ethicml.models.inprocess.agarwal_reductions import AgarwalArgs
     from ethicml.models.inprocess.in_subprocess import InAlgoArgs
+    from ethicml.models.inprocess.shared import LinearModel
 
 
 def fit(train: DataTuple, args: AgarwalArgs, seed: int = 888) -> ExponentiatedGradient:
     """Fit a model."""
     try:
-        from fairlearn.reductions import (
+        from fairlearn.reductions import (  # pyright: ignore
             DemographicParity,
             EqualizedOdds,
             ExponentiatedGradient,
@@ -50,13 +51,14 @@ def fit(train: DataTuple, args: AgarwalArgs, seed: int = 888) -> ExponentiatedGr
     fairness_class: UtilityParity
     fairness_type = FairnessType(args["fairness"])
     classifier_type = ClassifierType(args["classifier"])
-    kernel_type = None if args["kernel"] == "" else KernelType[args["kernel"]]
+    kernel_type = None if not args["kernel"] else KernelType[args["kernel"]]
 
     if fairness_type is FairnessType.dp:
         fairness_class = DemographicParity(difference_bound=args["eps"])
     else:
         fairness_class = EqualizedOdds(difference_bound=args["eps"])
 
+    model: Union[LinearModel, GradientBoostingClassifier]
     if classifier_type is ClassifierType.svm:
         assert kernel_type is not None
         model = select_svm(C=args["C"], kernel=kernel_type, seed=seed)
@@ -79,7 +81,7 @@ def fit(train: DataTuple, args: AgarwalArgs, seed: int = 888) -> ExponentiatedGr
     exponentiated_gradient.fit(data_x, data_y, sensitive_features=data_a)
 
     min_class_label = train.y.min()
-    exponentiated_gradient.min_class_label = min_class_label
+    exponentiated_gradient.min_class_label = min_class_label  # pyright: ignore
 
     return exponentiated_gradient
 
@@ -90,7 +92,7 @@ def predict(exponentiated_gradient: ExponentiatedGradient, test: TestTuple) -> p
     preds = pd.DataFrame(randomized_predictions, columns=["preds"])
 
     if (min_val := preds["preds"].min()) != preds["preds"].max():
-        preds = preds.replace(min_val, exponentiated_gradient.min_class_label)
+        preds = preds.replace(min_val, exponentiated_gradient.min_class_label)  # pyright: ignore
     return preds
 
 
@@ -105,7 +107,7 @@ def train_and_predict(
 @contextlib.contextmanager
 def working_dir(root: Path) -> Generator[None, None, None]:
     """Change the working directory to the given path."""
-    curdir = os.getcwd()
+    curdir = Path.cwd()
     os.chdir(root.expanduser().resolve().parent)
     try:
         yield
@@ -118,7 +120,7 @@ def main() -> None:
     in_algo_args: InAlgoArgs = json.loads(sys.argv[1])
     flags: AgarwalArgs = json.loads(sys.argv[2])
     try:
-        import cloudpickle
+        import cloudpickle  # pyright: ignore
 
         # Need to install cloudpickle for now. See https://github.com/fairlearn/fairlearn/issues/569
     except ImportError as e:

ethicml/implementations/beutel.py

@@ -80,14 +80,13 @@ def fit(train: DataTuple, flags: BeutelArgs, seed: int = 888) -> tuple[DataTuple
     set_seed(seed)
     fairness = FairnessType[flags["fairness"]]
 
-    post_process = False
+    processor: LabelBinarizer | None = None
     if flags["y_loss"] == "BCELoss()":
         try:
             assert_binary_labels(train)
         except AssertionError:
             processor = LabelBinarizer()
             train = processor.adjust(train)
-            post_process = True
 
     # By default we use 10% of the training data for validation
     train_, validation = train_test_split(train, train_percentage=1 - flags["validation_pcnt"])
@@ -136,6 +135,8 @@ def fit(train: DataTuple, flags: BeutelArgs, seed: int = 888) -> tuple[DataTuple
                 raise NotImplementedError("Not implemented Eq. Odds yet")
             elif fairness is FairnessType.dp:
                 mask = torch.ones(s_pred.shape, dtype=torch.uint8)
+            else:
+                raise NotImplementedError(f"Unknown value: {fairness}")
             loss += s_loss_fn(
                 s_pred, torch.masked_select(sens_label, mask).view(-1, int(train_data.sdim))
             )
@@ -169,8 +170,8 @@ def fit(train: DataTuple, flags: BeutelArgs, seed: int = 888) -> tuple[DataTuple
     enc.load_state_dict(best_enc)
 
     transformed_train = encode_dataset(enc, all_train_data_loader, train)
-    if post_process:
-        transformed_train = processor.post(encode_dataset(enc, all_train_data_loader, train))
+    if processor is not None:
+        transformed_train = processor.post(transformed_train)
     return transformed_train, enc
 
 
@@ -207,6 +208,8 @@ def get_mask(flags: BeutelArgs, s_pred: Tensor, class_label: Tensor) -> Tensor:
         raise NotImplementedError("Not implemented Eq. Odds yet")
     elif fairness is FairnessType.dp:
         mask = torch.ones(s_pred.shape, dtype=torch.uint8)
+    else:
+        raise NotImplementedError("Shouldn't be hit.")
     return mask