Improve OPE functionality, logging, and multiprocessing

.github/workflows/continuous_delivery.yml

@@ -43,7 +43,7 @@ jobs:
             ${{ runner.os }}-poetry-${{ matrix.python-version }}-${{ matrix.pydantic-version }}-
       - name: Install project dependencies with Poetry
         run: |
-          poetry install
+          poetry install --all-extras
       - name: Restore pyproject.toml
         run: |
           mv pyproject.toml.bak pyproject.toml

.github/workflows/continuous_documentation.yml

@@ -29,7 +29,7 @@ jobs:
           export PATH="$HOME/.poetry/bin:$PATH"
       - name: Install project dependencies with Poetry
         run: |
-          poetry install
+          poetry install --all-extras
 
       - name: Install Pandoc
         run: |

.github/workflows/continuous_integration.yml

@@ -38,7 +38,7 @@ jobs:
       - name: Install project dependencies with Poetry
         run: |
           poetry add pydantic@${{ matrix.pydantic-version }}
-          poetry install
+          poetry install --all-extras
       - name: Style check
         run: |
           # run pre-commit hooks

docs/src/authors.md

@@ -3,6 +3,7 @@
 - Dario d'Andrea, <dariod@playtika.com>
 - Shahar Bar, <shaharbar1@gmail.com>
 - Jerome Carayol, <jeromec@playtika.com>
+- Anastasiia Kabeshova, <anastasiia.kabeshova@gmail.com>
 - Stefano Piazza, <stefanop@playtika.com>
 - Ron Shiff, <ron.shiff1@gmail.com>
 - Raphael Steinmann, <raphaels@playtika.com>

pybandits/model.py

@@ -79,7 +79,7 @@ def _numpy_gelu(x: np.ndarray) -> np.ndarray:
     return 0.5 * x * (1 + erf(x / np.sqrt(2.0)))
 
 
-def _stable_sigmoid(x):
+def _numpy_sigmoid(x):
     """Stable sigmoid activation function for NumPy."""
     return np.where(x >= 0, 1 / (1 + np.exp(-x)), np.exp(x) / (1 + np.exp(x)))
 
@@ -850,7 +850,7 @@ class BaseBayesianNeuralNetwork(Model, ABC):
     _numpy_activations: ClassVar[dict] = {
         "tanh": np.tanh,
         "relu": _numpy_relu,
-        "sigmoid": _stable_sigmoid,
+        "sigmoid": _numpy_sigmoid,
         "gelu": _numpy_gelu,
     }
 
@@ -1208,7 +1208,7 @@ def sample_proba(self, context: np.ndarray) -> List[ProbabilityWeight]:
             else:
                 # Output layer - apply sigmoid
                 weighted_sum = linear_transform.squeeze(-1)
-                prob = _stable_sigmoid(weighted_sum)
+                prob = _numpy_sigmoid(weighted_sum)
 
         return list(zip(prob, weighted_sum))
 

pybandits/offline_policy_estimator.py

@@ -78,11 +78,18 @@ def _check_array(
                 raise ValueError(f"{name} must be a {ndim}D array.")
             if array.shape[0] != n_samples:
                 raise ValueError(f"action and {name} must have the same length.")
-            if array.dtype != dtype:
+            # Check dtype compatibility: use issubdtype for numpy dtypes
+            if dtype is float:
+                if not np.issubdtype(array.dtype, np.floating):
+                    raise ValueError(f"{name} must be a {dtype} array")
+            elif dtype is int:
+                if not np.issubdtype(array.dtype, np.integer):
+                    raise ValueError(f"{name} must be a {dtype} array")
+            elif array.dtype is not dtype:
                 raise ValueError(f"{name} must be a {dtype} array")
             if ndim > 1:
-                if array.shape[1] != n_actions:
-                    raise ValueError(f"{name} must have the same number of actions as the action array.")
+                if array.shape[1] < n_actions:
+                    raise ValueError(f"{name} must have at least number of actions as the action array.")
 
     @classmethod
     def _check_sum(cls, name: str, data: Dict[str, np.ndarray]):