explainiverse 0.2.1__tar.gz → 0.2.2__tar.gz

{explainiverse-0.2.1 → explainiverse-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: explainiverse
-Version: 0.2.1
+Version: 0.2.2
 Summary: Unified, extensible explainability framework supporting LIME, SHAP, Anchors, Counterfactuals, PDP, ALE, SAGE, and more
 Home-page: https://github.com/jemsbhai/explainiverse
 License: MIT
@@ -17,11 +17,13 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Provides-Extra: torch
 Requires-Dist: lime (>=0.2.0.1,<0.3.0.0)
 Requires-Dist: numpy (>=1.24,<2.0)
 Requires-Dist: scikit-learn (>=1.1,<1.6)
 Requires-Dist: scipy (>=1.10,<2.0)
 Requires-Dist: shap (>=0.48.0,<0.49.0)
+Requires-Dist: torch (>=2.0) ; extra == "torch"
 Requires-Dist: xgboost (>=1.7,<3.0)
 Project-URL: Repository, https://github.com/jemsbhai/explainiverse
 Description-Content-Type: text/markdown
@@ -29,7 +31,7 @@ Description-Content-Type: text/markdown
 # Explainiverse
 
 **Explainiverse** is a unified, extensible Python framework for Explainable AI (XAI).  
-It provides a standardized interface for model-agnostic explainability with 8 state-of-the-art XAI methods, evaluation metrics, and a plugin registry for easy extensibility.
+It provides a standardized interface for model-agnostic explainability with 9 state-of-the-art XAI methods, evaluation metrics, and a plugin registry for easy extensibility.
 
 ---
 
@@ -40,6 +42,7 @@ It provides a standardized interface for model-agnostic explainability with 8 st
 **Local Explainers** (instance-level explanations):
 - **LIME** - Local Interpretable Model-agnostic Explanations ([Ribeiro et al., 2016](https://arxiv.org/abs/1602.04938))
 - **SHAP** - SHapley Additive exPlanations via KernelSHAP ([Lundberg & Lee, 2017](https://arxiv.org/abs/1705.07874))
+- **TreeSHAP** - Exact SHAP values for tree models, 10x+ faster ([Lundberg et al., 2018](https://arxiv.org/abs/1802.03888))
 - **Anchors** - High-precision rule-based explanations ([Ribeiro et al., 2018](https://ojs.aaai.org/index.php/AAAI/article/view/11491))
 - **Counterfactual** - DiCE-style diverse counterfactual explanations ([Mothilal et al., 2020](https://arxiv.org/abs/1905.07697))
 
@@ -62,7 +65,7 @@ It provides a standardized interface for model-agnostic explainability with 8 st
 ### 🧪 Standardized Interface
 - Consistent `BaseExplainer` API
 - Unified `Explanation` output format
-- Model adapters for sklearn and more
+- Model adapters for sklearn and PyTorch
 
 ---
 
@@ -74,6 +77,12 @@ From PyPI:
 pip install explainiverse
 ```
 
+With PyTorch support (for neural network explanations):
+
+```bash
+pip install explainiverse[torch]
+```
+
 For development:
 
 ```bash
@@ -100,7 +109,7 @@ adapter = SklearnAdapter(model, class_names=iris.target_names.tolist())
 
 # List available explainers
 print(default_registry.list_explainers())
-# ['lime', 'shap', 'anchors', 'counterfactual', 'permutation_importance', 'partial_dependence', 'ale', 'sage']
+# ['lime', 'shap', 'treeshap', 'anchors', 'counterfactual', 'permutation_importance', 'partial_dependence', 'ale', 'sage']
 
 # Create and use an explainer
 explainer = default_registry.create(
@@ -119,11 +128,11 @@ print(explanation.explanation_data["feature_attributions"])
 ```python
 # Find local explainers for tabular data
 local_tabular = default_registry.filter(scope="local", data_type="tabular")
-print(local_tabular)  # ['lime', 'shap', 'anchors', 'counterfactual']
+print(local_tabular)  # ['lime', 'shap', 'treeshap', 'anchors', 'counterfactual']
 
-# Find global explainers
-global_explainers = default_registry.filter(scope="global")
-print(global_explainers)  # ['permutation_importance', 'partial_dependence', 'ale', 'sage']
+# Find explainers optimized for tree models
+tree_explainers = default_registry.filter(model_type="tree")
+print(tree_explainers)  # ['treeshap']
 
 # Get recommendations
 recommendations = default_registry.recommend(
@@ -133,6 +142,64 @@ recommendations = default_registry.recommend(
 )
 ```
 
+### TreeSHAP for Tree Models (10x+ Faster)
+
+```python
+from explainiverse.explainers import TreeShapExplainer
+from sklearn.ensemble import RandomForestClassifier
+
+# Train a tree-based model
+model = RandomForestClassifier(n_estimators=100).fit(X_train, y_train)
+
+# TreeSHAP works directly with the model (no adapter needed)
+explainer = TreeShapExplainer(
+    model=model,
+    feature_names=feature_names,
+    class_names=class_names
+)
+
+# Single instance explanation
+explanation = explainer.explain(X_test[0])
+print(explanation.explanation_data["feature_attributions"])
+
+# Batch explanations (efficient)
+explanations = explainer.explain_batch(X_test[:10])
+
+# Feature interactions
+interactions = explainer.explain_interactions(X_test[0])
+print(interactions.explanation_data["interaction_matrix"])
+```
+
+### PyTorch Adapter for Neural Networks
+
+```python
+from explainiverse import PyTorchAdapter
+import torch.nn as nn
+
+# Define a PyTorch model
+model = nn.Sequential(
+    nn.Linear(10, 64),
+    nn.ReLU(),
+    nn.Linear(64, 3)
+)
+
+# Wrap with adapter
+adapter = PyTorchAdapter(
+    model,
+    task="classification",
+    class_names=["cat", "dog", "bird"]
+)
+
+# Use with any explainer
+predictions = adapter.predict(X)  # Returns numpy array
+
+# Get gradients for attribution methods
+predictions, gradients = adapter.predict_with_gradients(X)
+
+# Access intermediate layers
+activations = adapter.get_layer_output(X, layer_name="0")
+```
+
 ### Using Specific Explainers
 
 ```python
@@ -233,12 +300,14 @@ poetry run pytest tests/test_new_explainers.py -v
 ## Roadmap
 
 - [x] LIME, SHAP (KernelSHAP)
+- [x] TreeSHAP (optimized for tree models) ✅ NEW
 - [x] Anchors, Counterfactuals
 - [x] Permutation Importance, PDP, ALE, SAGE
 - [x] Explainer Registry with filtering
-- [ ] TreeSHAP (optimized for tree models)
+- [x] PyTorch Adapter ✅ NEW
 - [ ] Integrated Gradients (gradient-based for neural nets)
-- [ ] PyTorch/TensorFlow adapters
+- [ ] GradCAM for CNNs
+- [ ] TensorFlow adapter
 - [ ] Interactive visualization dashboard
 
 ---

{explainiverse-0.2.1 → explainiverse-0.2.2}/README.md

@@ -1,7 +1,7 @@
 # Explainiverse
 
 **Explainiverse** is a unified, extensible Python framework for Explainable AI (XAI).  
-It provides a standardized interface for model-agnostic explainability with 8 state-of-the-art XAI methods, evaluation metrics, and a plugin registry for easy extensibility.
+It provides a standardized interface for model-agnostic explainability with 9 state-of-the-art XAI methods, evaluation metrics, and a plugin registry for easy extensibility.
 
 ---
 
@@ -12,6 +12,7 @@ It provides a standardized interface for model-agnostic explainability with 8 st
 **Local Explainers** (instance-level explanations):
 - **LIME** - Local Interpretable Model-agnostic Explanations ([Ribeiro et al., 2016](https://arxiv.org/abs/1602.04938))
 - **SHAP** - SHapley Additive exPlanations via KernelSHAP ([Lundberg & Lee, 2017](https://arxiv.org/abs/1705.07874))
+- **TreeSHAP** - Exact SHAP values for tree models, 10x+ faster ([Lundberg et al., 2018](https://arxiv.org/abs/1802.03888))
 - **Anchors** - High-precision rule-based explanations ([Ribeiro et al., 2018](https://ojs.aaai.org/index.php/AAAI/article/view/11491))
 - **Counterfactual** - DiCE-style diverse counterfactual explanations ([Mothilal et al., 2020](https://arxiv.org/abs/1905.07697))
 
@@ -34,7 +35,7 @@ It provides a standardized interface for model-agnostic explainability with 8 st
 ### 🧪 Standardized Interface
 - Consistent `BaseExplainer` API
 - Unified `Explanation` output format
-- Model adapters for sklearn and more
+- Model adapters for sklearn and PyTorch
 
 ---
 
@@ -46,6 +47,12 @@ From PyPI:
 pip install explainiverse
 ```
 
+With PyTorch support (for neural network explanations):
+
+```bash
+pip install explainiverse[torch]
+```
+
 For development:
 
 ```bash
@@ -72,7 +79,7 @@ adapter = SklearnAdapter(model, class_names=iris.target_names.tolist())
 
 # List available explainers
 print(default_registry.list_explainers())
-# ['lime', 'shap', 'anchors', 'counterfactual', 'permutation_importance', 'partial_dependence', 'ale', 'sage']
+# ['lime', 'shap', 'treeshap', 'anchors', 'counterfactual', 'permutation_importance', 'partial_dependence', 'ale', 'sage']
 
 # Create and use an explainer
 explainer = default_registry.create(
@@ -91,11 +98,11 @@ print(explanation.explanation_data["feature_attributions"])
 ```python
 # Find local explainers for tabular data
 local_tabular = default_registry.filter(scope="local", data_type="tabular")
-print(local_tabular)  # ['lime', 'shap', 'anchors', 'counterfactual']
+print(local_tabular)  # ['lime', 'shap', 'treeshap', 'anchors', 'counterfactual']
 
-# Find global explainers
-global_explainers = default_registry.filter(scope="global")
-print(global_explainers)  # ['permutation_importance', 'partial_dependence', 'ale', 'sage']
+# Find explainers optimized for tree models
+tree_explainers = default_registry.filter(model_type="tree")
+print(tree_explainers)  # ['treeshap']
 
 # Get recommendations
 recommendations = default_registry.recommend(
@@ -105,6 +112,64 @@ recommendations = default_registry.recommend(
 )
 ```
 
+### TreeSHAP for Tree Models (10x+ Faster)
+
+```python
+from explainiverse.explainers import TreeShapExplainer
+from sklearn.ensemble import RandomForestClassifier
+
+# Train a tree-based model
+model = RandomForestClassifier(n_estimators=100).fit(X_train, y_train)
+
+# TreeSHAP works directly with the model (no adapter needed)
+explainer = TreeShapExplainer(
+    model=model,
+    feature_names=feature_names,
+    class_names=class_names
+)
+
+# Single instance explanation
+explanation = explainer.explain(X_test[0])
+print(explanation.explanation_data["feature_attributions"])
+
+# Batch explanations (efficient)
+explanations = explainer.explain_batch(X_test[:10])
+
+# Feature interactions
+interactions = explainer.explain_interactions(X_test[0])
+print(interactions.explanation_data["interaction_matrix"])
+```
+
+### PyTorch Adapter for Neural Networks
+
+```python
+from explainiverse import PyTorchAdapter
+import torch.nn as nn
+
+# Define a PyTorch model
+model = nn.Sequential(
+    nn.Linear(10, 64),
+    nn.ReLU(),
+    nn.Linear(64, 3)
+)
+
+# Wrap with adapter
+adapter = PyTorchAdapter(
+    model,
+    task="classification",
+    class_names=["cat", "dog", "bird"]
+)
+
+# Use with any explainer
+predictions = adapter.predict(X)  # Returns numpy array
+
+# Get gradients for attribution methods
+predictions, gradients = adapter.predict_with_gradients(X)
+
+# Access intermediate layers
+activations = adapter.get_layer_output(X, layer_name="0")
+```
+
 ### Using Specific Explainers
 
 ```python
@@ -205,12 +270,14 @@ poetry run pytest tests/test_new_explainers.py -v
 ## Roadmap
 
 - [x] LIME, SHAP (KernelSHAP)
+- [x] TreeSHAP (optimized for tree models) ✅ NEW
 - [x] Anchors, Counterfactuals
 - [x] Permutation Importance, PDP, ALE, SAGE
 - [x] Explainer Registry with filtering
-- [ ] TreeSHAP (optimized for tree models)
+- [x] PyTorch Adapter ✅ NEW
 - [ ] Integrated Gradients (gradient-based for neural nets)
-- [ ] PyTorch/TensorFlow adapters
+- [ ] GradCAM for CNNs
+- [ ] TensorFlow adapter
 - [ ] Interactive visualization dashboard
 
 ---

{explainiverse-0.2.1 → explainiverse-0.2.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "explainiverse"
-version = "0.2.1"
+version = "0.2.2"
 description = "Unified, extensible explainability framework supporting LIME, SHAP, Anchors, Counterfactuals, PDP, ALE, SAGE, and more"
 authors = ["Muntaser Syed <jemsbhai@gmail.com>"]
 license = "MIT"
@@ -29,6 +29,10 @@ scikit-learn = ">=1.1,<1.6"
 shap = "^0.48.0"
 scipy = ">=1.10,<2.0"
 xgboost = ">=1.7,<3.0"
+torch = { version = ">=2.0", optional = true }
+
+[tool.poetry.extras]
+torch = ["torch"]
 
 [tool.poetry.group.dev.dependencies]
 pytest = "^8.0"

{explainiverse-0.2.1 → explainiverse-0.2.2}/src/explainiverse/__init__.py

@@ -2,8 +2,9 @@
 """
 Explainiverse - A unified, extensible explainability framework.
 
-Supports multiple XAI methods including LIME, SHAP, Anchors, Counterfactuals,
-Permutation Importance, PDP, ALE, and SAGE through a consistent interface.
+Supports multiple XAI methods including LIME, SHAP, TreeSHAP, Anchors, 
+Counterfactuals, Permutation Importance, PDP, ALE, and SAGE through a 
+consistent interface.
 
 Quick Start:
     from explainiverse import default_registry
@@ -14,6 +15,10 @@ Quick Start:
     # Create an explainer
     explainer = default_registry.create("lime", model=adapter, training_data=X, ...)
     explanation = explainer.explain(instance)
+    
+For PyTorch models:
+    from explainiverse import PyTorchAdapter  # Requires torch
+    adapter = PyTorchAdapter(model, task="classification")
 """
 
 from explainiverse.core.explainer import BaseExplainer
@@ -25,9 +30,10 @@ from explainiverse.core.registry import (
     get_default_registry,
 )
 from explainiverse.adapters.sklearn_adapter import SklearnAdapter
+from explainiverse.adapters import TORCH_AVAILABLE
 from explainiverse.engine.suite import ExplanationSuite
 
-__version__ = "0.2.1"
+__version__ = "0.2.2"
 
 __all__ = [
     # Core
@@ -40,6 +46,12 @@ __all__ = [
     "get_default_registry",
     # Adapters
     "SklearnAdapter",
+    "TORCH_AVAILABLE",
     # Engine
     "ExplanationSuite",
 ]
+
+# Conditionally export PyTorchAdapter if torch is available
+if TORCH_AVAILABLE:
+    from explainiverse.adapters import PyTorchAdapter
+    __all__.append("PyTorchAdapter")

explainiverse-0.2.2/src/explainiverse/adapters/__init__.py

@@ -0,0 +1,19 @@
+# src/explainiverse/adapters/__init__.py
+"""
+Model adapters - wrappers that provide a consistent interface for different ML frameworks.
+
+Available adapters:
+- SklearnAdapter: For scikit-learn models (always available)
+- PyTorchAdapter: For PyTorch nn.Module models (requires torch)
+"""
+
+from explainiverse.adapters.base_adapter import BaseModelAdapter
+from explainiverse.adapters.sklearn_adapter import SklearnAdapter
+
+# Conditionally import PyTorchAdapter if torch is available
+try:
+    from explainiverse.adapters.pytorch_adapter import PyTorchAdapter, TORCH_AVAILABLE
+    __all__ = ["BaseModelAdapter", "SklearnAdapter", "PyTorchAdapter", "TORCH_AVAILABLE"]
+except ImportError:
+    TORCH_AVAILABLE = False
+    __all__ = ["BaseModelAdapter", "SklearnAdapter", "TORCH_AVAILABLE"]

explainiverse-0.2.2/src/explainiverse/adapters/pytorch_adapter.py

@@ -0,0 +1,396 @@
+# src/explainiverse/adapters/pytorch_adapter.py
+"""
+PyTorch Model Adapter for Explainiverse.
+
+Provides a unified interface for PyTorch neural networks, enabling
+compatibility with all explainers in the framework.
+
+Example:
+    import torch.nn as nn
+    from explainiverse.adapters import PyTorchAdapter
+    
+    model = nn.Sequential(
+        nn.Linear(10, 64),
+        nn.ReLU(),
+        nn.Linear(64, 3)
+    )
+    
+    adapter = PyTorchAdapter(
+        model,
+        task="classification",
+        class_names=["cat", "dog", "bird"]
+    )
+    
+    probs = adapter.predict(X)  # Returns numpy array
+"""
+
+import numpy as np
+from typing import List, Optional, Union, Callable
+
+from .base_adapter import BaseModelAdapter
+
+# Check if PyTorch is available
+try:
+    import torch
+    import torch.nn as nn
+    TORCH_AVAILABLE = True
+except ImportError:
+    TORCH_AVAILABLE = False
+    torch = None
+    nn = None
+
+
+def _check_torch_available():
+    """Raise ImportError if PyTorch is not installed."""
+    if not TORCH_AVAILABLE:
+        raise ImportError(
+            "PyTorch is required for PyTorchAdapter. "
+            "Install it with: pip install torch"
+        )
+
+
+class PyTorchAdapter(BaseModelAdapter):
+    """
+    Adapter for PyTorch neural network models.
+    
+    Wraps a PyTorch nn.Module to provide a consistent interface for
+    explainability methods. Handles device management, tensor/numpy
+    conversions, and supports both classification and regression tasks.
+    
+    Attributes:
+        model: The PyTorch model (nn.Module)
+        task: "classification" or "regression"
+        device: torch.device for computation
+        class_names: List of class names (for classification)
+        feature_names: List of feature names
+        output_activation: Optional activation function for outputs
+    
+    Example:
+        >>> model = MyNeuralNetwork()
+        >>> adapter = PyTorchAdapter(model, task="classification")
+        >>> probs = adapter.predict(X_numpy)  # Returns probabilities
+    """
+    
+    def __init__(
+        self,
+        model,
+        task: str = "classification",
+        feature_names: Optional[List[str]] = None,
+        class_names: Optional[List[str]] = None,
+        device: Optional[str] = None,
+        output_activation: Optional[str] = "auto",
+        batch_size: int = 32
+    ):
+        """
+        Initialize the PyTorch adapter.
+        
+        Args:
+            model: A PyTorch nn.Module model.
+            task: "classification" or "regression".
+            feature_names: List of input feature names.
+            class_names: List of output class names (classification only).
+            device: Device to run on ("cpu", "cuda", "cuda:0", etc.).
+                   If None, auto-detects based on model parameters.
+            output_activation: Activation for output layer:
+                - "auto": softmax for classification, none for regression
+                - "softmax": Apply softmax (classification)
+                - "sigmoid": Apply sigmoid (binary classification)
+                - "none" or None: No activation (raw logits/values)
+            batch_size: Batch size for large inputs (default: 32).
+        """
+        _check_torch_available()
+        
+        if not isinstance(model, nn.Module):
+            raise TypeError(
+                f"Expected nn.Module, got {type(model).__name__}. "
+                "For sklearn models, use SklearnAdapter instead."
+            )
+        
+        super().__init__(model, feature_names)
+        
+        self.task = task
+        self.class_names = list(class_names) if class_names else None
+        self.batch_size = batch_size
+        
+        # Determine device
+        if device is not None:
+            self.device = torch.device(device)
+        else:
+            # Auto-detect from model parameters
+            try:
+                param = next(model.parameters())
+                self.device = param.device
+            except StopIteration:
+                # Model has no parameters, use CPU
+                self.device = torch.device("cpu")
+        
+        # Move model to device and set to eval mode
+        self.model = model.to(self.device)
+        self.model.eval()
+        
+        # Configure output activation
+        if output_activation == "auto":
+            if task == "classification":
+                self.output_activation = "softmax"
+            else:
+                self.output_activation = None
+        else:
+            self.output_activation = output_activation if output_activation != "none" else None
+    
+    def _to_tensor(self, data: np.ndarray) -> "torch.Tensor":
+        """Convert numpy array to tensor on the correct device."""
+        if isinstance(data, torch.Tensor):
+            return data.to(self.device).float()
+        return torch.tensor(data, dtype=torch.float32, device=self.device)
+    
+    def _to_numpy(self, tensor: "torch.Tensor") -> np.ndarray:
+        """Convert tensor to numpy array."""
+        return tensor.detach().cpu().numpy()
+    
+    def _apply_activation(self, output: "torch.Tensor") -> "torch.Tensor":
+        """Apply output activation function."""
+        if self.output_activation == "softmax":
+            return torch.softmax(output, dim=-1)
+        elif self.output_activation == "sigmoid":
+            return torch.sigmoid(output)
+        return output
+    
+    def predict(self, data: np.ndarray) -> np.ndarray:
+        """
+        Generate predictions for input data.
+        
+        Args:
+            data: Input data as numpy array. Shape: (n_samples, n_features)
+                  or (n_samples, channels, height, width) for images.
+        
+        Returns:
+            Predictions as numpy array:
+            - Classification: probabilities of shape (n_samples, n_classes)
+            - Regression: values of shape (n_samples, n_outputs)
+        """
+        data = np.array(data)
+        
+        # Handle single instance
+        if data.ndim == 1:
+            data = data.reshape(1, -1)
+        
+        n_samples = data.shape[0]
+        outputs = []
+        
+        with torch.no_grad():
+            for i in range(0, n_samples, self.batch_size):
+                batch = data[i:i + self.batch_size]
+                tensor_batch = self._to_tensor(batch)
+                
+                output = self.model(tensor_batch)
+                output = self._apply_activation(output)
+                outputs.append(self._to_numpy(output))
+        
+        return np.vstack(outputs)
+    
+    def predict_with_gradients(
+        self,
+        data: np.ndarray,
+        target_class: Optional[int] = None
+    ) -> tuple:
+        """
+        Generate predictions and compute gradients w.r.t. inputs.
+        
+        This is essential for gradient-based attribution methods like
+        Integrated Gradients, GradCAM, and Saliency Maps.
+        
+        Args:
+            data: Input data as numpy array.
+            target_class: Class index for gradient computation.
+                         If None, uses the predicted class.
+        
+        Returns:
+            Tuple of (predictions, gradients) as numpy arrays.
+        """
+        data = np.array(data)
+        if data.ndim == 1:
+            data = data.reshape(1, -1)
+        
+        # Convert to tensor with gradient tracking
+        tensor_data = self._to_tensor(data)
+        tensor_data.requires_grad_(True)
+        
+        # Forward pass
+        output = self.model(tensor_data)
+        activated_output = self._apply_activation(output)
+        
+        # Determine target for gradient
+        if self.task == "classification":
+            if target_class is None:
+                target_class = output.argmax(dim=-1)
+            elif isinstance(target_class, int):
+                target_class = torch.tensor([target_class] * data.shape[0], device=self.device)
+            
+            # Select target class scores for gradient
+            target_scores = output.gather(1, target_class.view(-1, 1)).squeeze()
+        else:
+            # Regression: gradient w.r.t. output
+            target_scores = output.squeeze()
+        
+        # Backward pass
+        if target_scores.dim() == 0:
+            target_scores.backward()
+        else:
+            target_scores.sum().backward()
+        
+        gradients = tensor_data.grad
+        
+        return (
+            self._to_numpy(activated_output),
+            self._to_numpy(gradients)
+        )
+    
+    def get_layer_output(
+        self,
+        data: np.ndarray,
+        layer_name: str
+    ) -> np.ndarray:
+        """
+        Get intermediate layer activations.
+        
+        Useful for methods like GradCAM that need feature map activations.
+        
+        Args:
+            data: Input data as numpy array.
+            layer_name: Name of the layer to extract (as registered in model).
+        
+        Returns:
+            Layer activations as numpy array.
+        """
+        data = np.array(data)
+        if data.ndim == 1:
+            data = data.reshape(1, -1)
+        
+        activations = {}
+        
+        def hook_fn(module, input, output):
+            activations['output'] = output
+        
+        # Find and hook the layer
+        layer = dict(self.model.named_modules()).get(layer_name)
+        if layer is None:
+            available = list(dict(self.model.named_modules()).keys())
+            raise ValueError(
+                f"Layer '{layer_name}' not found. Available layers: {available}"
+            )
+        
+        handle = layer.register_forward_hook(hook_fn)
+        
+        try:
+            with torch.no_grad():
+                tensor_data = self._to_tensor(data)
+                _ = self.model(tensor_data)
+        finally:
+            handle.remove()
+        
+        return self._to_numpy(activations['output'])
+    
+    def get_layer_gradients(
+        self,
+        data: np.ndarray,
+        layer_name: str,
+        target_class: Optional[int] = None
+    ) -> tuple:
+        """
+        Get gradients of output w.r.t. a specific layer's activations.
+        
+        Essential for GradCAM and similar visualization methods.
+        
+        Args:
+            data: Input data as numpy array.
+            layer_name: Name of the layer for gradient computation.
+            target_class: Target class for gradient (classification).
+        
+        Returns:
+            Tuple of (layer_activations, layer_gradients) as numpy arrays.
+        """
+        data = np.array(data)
+        if data.ndim == 1:
+            data = data.reshape(1, -1)
+        
+        activations = {}
+        gradients = {}
+        
+        def forward_hook(module, input, output):
+            activations['output'] = output
+        
+        def backward_hook(module, grad_input, grad_output):
+            gradients['output'] = grad_output[0]
+        
+        # Find and hook the layer
+        layer = dict(self.model.named_modules()).get(layer_name)
+        if layer is None:
+            available = list(dict(self.model.named_modules()).keys())
+            raise ValueError(
+                f"Layer '{layer_name}' not found. Available layers: {available}"
+            )
+        
+        forward_handle = layer.register_forward_hook(forward_hook)
+        backward_handle = layer.register_full_backward_hook(backward_hook)
+        
+        try:
+            tensor_data = self._to_tensor(data)
+            tensor_data.requires_grad_(True)
+            
+            output = self.model(tensor_data)
+            
+            if self.task == "classification":
+                if target_class is None:
+                    target_class = output.argmax(dim=-1)
+                elif isinstance(target_class, int):
+                    target_class = torch.tensor([target_class] * data.shape[0], device=self.device)
+                
+                target_scores = output.gather(1, target_class.view(-1, 1)).squeeze()
+            else:
+                target_scores = output.squeeze()
+            
+            if target_scores.dim() == 0:
+                target_scores.backward()
+            else:
+                target_scores.sum().backward()
+        finally:
+            forward_handle.remove()
+            backward_handle.remove()
+        
+        return (
+            self._to_numpy(activations['output']),
+            self._to_numpy(gradients['output'])
+        )
+    
+    def list_layers(self) -> List[str]:
+        """
+        List all named layers/modules in the model.
+        
+        Returns:
+            List of layer names that can be used with get_layer_output/gradients.
+        """
+        return [name for name, _ in self.model.named_modules() if name]
+    
+    def to(self, device: str) -> "PyTorchAdapter":
+        """
+        Move the model to a different device.
+        
+        Args:
+            device: Target device ("cpu", "cuda", "cuda:0", etc.)
+        
+        Returns:
+            Self for chaining.
+        """
+        self.device = torch.device(device)
+        self.model = self.model.to(self.device)
+        return self
+    
+    def train_mode(self) -> "PyTorchAdapter":
+        """Set model to training mode (enables dropout, batchnorm updates)."""
+        self.model.train()
+        return self
+    
+    def eval_mode(self) -> "PyTorchAdapter":
+        """Set model to evaluation mode (disables dropout, freezes batchnorm)."""
+        self.model.eval()
+        return self

explainiverse-0.2.1/src/explainiverse/adapters/__init__.py

@@ -1,9 +0,0 @@
-# src/explainiverse/adapters/__init__.py
-"""
-Model adapters - wrappers that provide a consistent interface for different ML frameworks.
-"""
-
-from explainiverse.adapters.base_adapter import BaseModelAdapter
-from explainiverse.adapters.sklearn_adapter import SklearnAdapter
-
-__all__ = ["BaseModelAdapter", "SklearnAdapter"]