explainiverse 0.2.1__py3-none-any.whl → 0.2.3__py3-none-any.whl

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__ = "0.2.3"
 
 __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/adapters/__init__.py

@@ -1,9 +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
 
-__all__ = ["BaseModelAdapter", "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/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/core/registry.py

@@ -369,6 +369,7 @@ def _create_default_registry() -> ExplainerRegistry:
     from explainiverse.explainers.global_explainers.ale import ALEExplainer
     from explainiverse.explainers.global_explainers.sage import SAGEExplainer
     from explainiverse.explainers.counterfactual.dice_wrapper import CounterfactualExplainer
+    from explainiverse.explainers.gradient.integrated_gradients import IntegratedGradientsExplainer
     
     registry = ExplainerRegistry()
     
@@ -461,6 +462,23 @@ def _create_default_registry() -> ExplainerRegistry:
         )
     )
     
+    # Register Integrated Gradients (for neural networks)
+    registry.register(
+        name="integrated_gradients",
+        explainer_class=IntegratedGradientsExplainer,
+        meta=ExplainerMeta(
+            scope="local",
+            model_types=["neural"],
+            data_types=["tabular", "image"],
+            task_types=["classification", "regression"],
+            description="Integrated Gradients - axiomatic attributions for neural networks (requires PyTorch)",
+            paper_reference="Sundararajan et al., 2017 - 'Axiomatic Attribution for Deep Networks' (ICML)",
+            complexity="O(n_steps * forward_pass)",
+            requires_training_data=False,
+            supports_batching=True
+        )
+    )
+    
     # =========================================================================
     # Global Explainers (model-level)
     # =========================================================================

explainiverse/explainers/__init__.py

@@ -8,6 +8,7 @@ Local Explainers (instance-level):
 - TreeSHAP: Optimized exact SHAP for tree-based models
 - Anchors: High-precision rule-based explanations
 - Counterfactual: Diverse counterfactual explanations
+- Integrated Gradients: Gradient-based attributions for neural networks
 
 Global Explainers (model-level):
 - Permutation Importance: Feature importance via permutation
@@ -25,6 +26,7 @@ from explainiverse.explainers.global_explainers.permutation_importance import Pe
 from explainiverse.explainers.global_explainers.partial_dependence import PartialDependenceExplainer
 from explainiverse.explainers.global_explainers.ale import ALEExplainer
 from explainiverse.explainers.global_explainers.sage import SAGEExplainer
+from explainiverse.explainers.gradient.integrated_gradients import IntegratedGradientsExplainer
 
 __all__ = [
     # Local explainers
@@ -33,6 +35,7 @@ __all__ = [
     "TreeShapExplainer",
     "AnchorsExplainer",
     "CounterfactualExplainer",
+    "IntegratedGradientsExplainer",
     # Global explainers
     "PermutationImportanceExplainer",
     "PartialDependenceExplainer",

explainiverse/explainers/gradient/__init__.py

@@ -0,0 +1,11 @@
+# src/explainiverse/explainers/gradient/__init__.py
+"""
+Gradient-based explainers for neural networks.
+
+These explainers require models that support gradient computation,
+typically via the PyTorchAdapter.
+"""
+
+from explainiverse.explainers.gradient.integrated_gradients import IntegratedGradientsExplainer
+
+__all__ = ["IntegratedGradientsExplainer"]

explainiverse/explainers/gradient/integrated_gradients.py

@@ -0,0 +1,348 @@
+# src/explainiverse/explainers/gradient/integrated_gradients.py
+"""
+Integrated Gradients - Axiomatic Attribution for Deep Networks.
+
+Integrated Gradients computes feature attributions by accumulating gradients
+along a straight-line path from a baseline to the input. It satisfies two
+key axioms:
+- Sensitivity: If a feature differs between input and baseline and changes
+  the prediction, it receives non-zero attribution.
+- Implementation Invariance: Attributions are identical for functionally
+  equivalent networks.
+
+Reference:
+    Sundararajan, M., Taly, A., & Yan, Q. (2017). Axiomatic Attribution for
+    Deep Networks. ICML 2017. https://arxiv.org/abs/1703.01365
+
+Example:
+    from explainiverse.explainers.gradient import IntegratedGradientsExplainer
+    from explainiverse.adapters import PyTorchAdapter
+    
+    adapter = PyTorchAdapter(model, task="classification")
+    
+    explainer = IntegratedGradientsExplainer(
+        model=adapter,
+        feature_names=feature_names,
+        n_steps=50
+    )
+    
+    explanation = explainer.explain(instance)
+"""
+
+import numpy as np
+from typing import List, Optional, Union, Callable
+
+from explainiverse.core.explainer import BaseExplainer
+from explainiverse.core.explanation import Explanation
+
+
+class IntegratedGradientsExplainer(BaseExplainer):
+    """
+    Integrated Gradients explainer for neural networks.
+    
+    Computes attributions by integrating gradients along the path from
+    a baseline (default: zero vector) to the input. The integral is
+    approximated using the Riemann sum.
+    
+    Attributes:
+        model: Model adapter with predict_with_gradients() method
+        feature_names: List of feature names
+        class_names: List of class names (for classification)
+        n_steps: Number of steps for integral approximation
+        baseline: Baseline input (default: zeros)
+        method: Integration method ("riemann_middle", "riemann_left", "riemann_right", "riemann_trapezoid")
+    """
+    
+    def __init__(
+        self,
+        model,
+        feature_names: List[str],
+        class_names: Optional[List[str]] = None,
+        n_steps: int = 50,
+        baseline: Optional[np.ndarray] = None,
+        method: str = "riemann_middle"
+    ):
+        """
+        Initialize the Integrated Gradients explainer.
+        
+        Args:
+            model: A model adapter with predict_with_gradients() method.
+                   Use PyTorchAdapter for PyTorch models.
+            feature_names: List of input feature names.
+            class_names: List of class names (for classification tasks).
+            n_steps: Number of steps for approximating the integral.
+                    More steps = more accurate but slower. Default: 50.
+            baseline: Baseline input for comparison. If None, uses zeros.
+                     Can also be "random" for random baseline or a callable.
+            method: Integration method:
+                   - "riemann_middle": Middle Riemann sum (default, most accurate)
+                   - "riemann_left": Left Riemann sum
+                   - "riemann_right": Right Riemann sum
+                   - "riemann_trapezoid": Trapezoidal rule
+        """
+        super().__init__(model)
+        
+        # Validate model has gradient capability
+        if not hasattr(model, 'predict_with_gradients'):
+            raise TypeError(
+                "Model adapter must have predict_with_gradients() method. "
+                "Use PyTorchAdapter for PyTorch models."
+            )
+        
+        self.feature_names = list(feature_names)
+        self.class_names = list(class_names) if class_names else None
+        self.n_steps = n_steps
+        self.baseline = baseline
+        self.method = method
+    
+    def _get_baseline(self, instance: np.ndarray) -> np.ndarray:
+        """Get the baseline for a given input shape."""
+        if self.baseline is None:
+            # Default: zero baseline
+            return np.zeros_like(instance)
+        elif isinstance(self.baseline, str) and self.baseline == "random":
+            # Random baseline (useful for images)
+            return np.random.uniform(
+                low=instance.min(),
+                high=instance.max(),
+                size=instance.shape
+            ).astype(instance.dtype)
+        elif callable(self.baseline):
+            return self.baseline(instance)
+        else:
+            return np.array(self.baseline).reshape(instance.shape)
+    
+    def _get_interpolation_alphas(self) -> np.ndarray:
+        """Get interpolation points based on method."""
+        if self.method == "riemann_left":
+            return np.linspace(0, 1 - 1/self.n_steps, self.n_steps)
+        elif self.method == "riemann_right":
+            return np.linspace(1/self.n_steps, 1, self.n_steps)
+        elif self.method == "riemann_middle":
+            return np.linspace(0.5/self.n_steps, 1 - 0.5/self.n_steps, self.n_steps)
+        elif self.method == "riemann_trapezoid":
+            return np.linspace(0, 1, self.n_steps + 1)
+        else:
+            raise ValueError(f"Unknown method: {self.method}")
+    
+    def _compute_integrated_gradients(
+        self,
+        instance: np.ndarray,
+        baseline: np.ndarray,
+        target_class: Optional[int] = None
+    ) -> np.ndarray:
+        """
+        Compute integrated gradients for a single instance.
+        
+        The integral is approximated as:
+        IG_i = (x_i - x'_i) * sum_{k=1}^{m} grad_i(x' + k/m * (x - x')) / m
+        
+        where x is the input, x' is the baseline, and m is n_steps.
+        """
+        # Get interpolation points
+        alphas = self._get_interpolation_alphas()
+        
+        # Compute path from baseline to input
+        # Shape: (n_steps, n_features)
+        delta = instance - baseline
+        interpolated_inputs = baseline + alphas[:, np.newaxis] * delta
+        
+        # Compute gradients at each interpolation point
+        all_gradients = []
+        for interp_input in interpolated_inputs:
+            _, gradients = self.model.predict_with_gradients(
+                interp_input.reshape(1, -1),
+                target_class=target_class
+            )
+            all_gradients.append(gradients.flatten())
+        
+        all_gradients = np.array(all_gradients)  # Shape: (n_steps, n_features)
+        
+        # Approximate the integral
+        if self.method == "riemann_trapezoid":
+            # Trapezoidal rule: (f(0) + 2*f(1) + ... + 2*f(n-1) + f(n)) / (2n)
+            weights = np.ones(self.n_steps + 1)
+            weights[0] = 0.5
+            weights[-1] = 0.5
+            avg_gradients = np.average(all_gradients, axis=0, weights=weights)
+        else:
+            # Standard Riemann sum: average of gradients
+            avg_gradients = np.mean(all_gradients, axis=0)
+        
+        # Scale by input - baseline difference
+        integrated_gradients = delta * avg_gradients
+        
+        return integrated_gradients
+    
+    def explain(
+        self,
+        instance: np.ndarray,
+        target_class: Optional[int] = None,
+        baseline: Optional[np.ndarray] = None,
+        return_convergence_delta: bool = False
+    ) -> Explanation:
+        """
+        Generate Integrated Gradients explanation for an instance.
+        
+        Args:
+            instance: 1D numpy array of input features.
+            target_class: For classification, which class to explain.
+                         If None, uses the predicted class.
+            baseline: Override the default baseline for this explanation.
+            return_convergence_delta: If True, include the convergence delta
+                                     (difference between sum of attributions
+                                     and prediction difference).
+        
+        Returns:
+            Explanation object with feature attributions.
+        """
+        instance = np.array(instance).flatten().astype(np.float32)
+        
+        # Get baseline
+        if baseline is not None:
+            bl = np.array(baseline).flatten().astype(np.float32)
+        else:
+            bl = self._get_baseline(instance)
+        
+        # Determine target class if not specified
+        if target_class is None and self.class_names:
+            predictions = self.model.predict(instance.reshape(1, -1))
+            target_class = int(np.argmax(predictions))
+        
+        # Compute integrated gradients
+        ig_attributions = self._compute_integrated_gradients(
+            instance, bl, target_class
+        )
+        
+        # Build attributions dict
+        attributions = {
+            fname: float(ig_attributions[i])
+            for i, fname in enumerate(self.feature_names)
+        }
+        
+        # Determine class name
+        if self.class_names and target_class is not None:
+            label_name = self.class_names[target_class]
+        else:
+            label_name = f"class_{target_class}" if target_class is not None else "output"
+        
+        explanation_data = {
+            "feature_attributions": attributions,
+            "attributions_raw": ig_attributions.tolist(),
+            "baseline": bl.tolist(),
+            "n_steps": self.n_steps,
+            "method": self.method
+        }
+        
+        # Optionally compute convergence delta
+        if return_convergence_delta:
+            # The sum of attributions should equal F(x) - F(baseline)
+            pred_input = self.model.predict(instance.reshape(1, -1))
+            pred_baseline = self.model.predict(bl.reshape(1, -1))
+            
+            if target_class is not None:
+                pred_diff = pred_input[0, target_class] - pred_baseline[0, target_class]
+            else:
+                pred_diff = pred_input[0, 0] - pred_baseline[0, 0]
+            
+            attribution_sum = np.sum(ig_attributions)
+            convergence_delta = abs(pred_diff - attribution_sum)
+            
+            explanation_data["convergence_delta"] = float(convergence_delta)
+            explanation_data["prediction_difference"] = float(pred_diff)
+            explanation_data["attribution_sum"] = float(attribution_sum)
+        
+        return Explanation(
+            explainer_name="IntegratedGradients",
+            target_class=label_name,
+            explanation_data=explanation_data
+        )
+    
+    def explain_batch(
+        self,
+        X: np.ndarray,
+        target_class: Optional[int] = None
+    ) -> List[Explanation]:
+        """
+        Generate explanations for multiple instances.
+        
+        Note: This is not optimized for batching - it processes
+        instances sequentially. For large batches, consider using
+        the batched gradient computation in a custom implementation.
+        
+        Args:
+            X: 2D numpy array of instances (n_samples, n_features).
+            target_class: Target class for all instances.
+        
+        Returns:
+            List of Explanation objects.
+        """
+        X = np.array(X)
+        if X.ndim == 1:
+            X = X.reshape(1, -1)
+        
+        return [
+            self.explain(X[i], target_class=target_class)
+            for i in range(X.shape[0])
+        ]
+    
+    def compute_attributions_with_noise(
+        self,
+        instance: np.ndarray,
+        target_class: Optional[int] = None,
+        n_samples: int = 5,
+        noise_scale: float = 0.1
+    ) -> Explanation:
+        """
+        Compute attributions averaged over noisy baselines (SmoothGrad-style).
+        
+        This can help reduce noise in the attributions by averaging over
+        multiple baselines sampled around the zero baseline.
+        
+        Args:
+            instance: Input instance.
+            target_class: Target class for attribution.
+            n_samples: Number of noisy baselines to average.
+            noise_scale: Standard deviation of Gaussian noise.
+        
+        Returns:
+            Explanation with averaged attributions.
+        """
+        instance = np.array(instance).flatten().astype(np.float32)
+        
+        all_attributions = []
+        for _ in range(n_samples):
+            # Create noisy baseline
+            noise = np.random.normal(0, noise_scale, instance.shape).astype(np.float32)
+            noisy_baseline = noise  # Noise around zero
+            
+            ig = self._compute_integrated_gradients(
+                instance, noisy_baseline, target_class
+            )
+            all_attributions.append(ig)
+        
+        # Average attributions
+        avg_attributions = np.mean(all_attributions, axis=0)
+        std_attributions = np.std(all_attributions, axis=0)
+        
+        attributions = {
+            fname: float(avg_attributions[i])
+            for i, fname in enumerate(self.feature_names)
+        }
+        
+        if self.class_names and target_class is not None:
+            label_name = self.class_names[target_class]
+        else:
+            label_name = f"class_{target_class}" if target_class is not None else "output"
+        
+        return Explanation(
+            explainer_name="IntegratedGradients_Smooth",
+            target_class=label_name,
+            explanation_data={
+                "feature_attributions": attributions,
+                "attributions_raw": avg_attributions.tolist(),
+                "attributions_std": std_attributions.tolist(),
+                "n_samples": n_samples,
+                "noise_scale": noise_scale
+            }
+        )

{explainiverse-0.2.1.dist-info → explainiverse-0.2.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: explainiverse
-Version: 0.2.1
+Version: 0.2.3
 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 10 state-of-the-art XAI methods, evaluation metrics, and a plugin registry for easy extensibility.
 
 ---
 
@@ -40,6 +42,8 @@ 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))
+- **Integrated Gradients** - Axiomatic attributions for neural networks ([Sundararajan et al., 2017](https://arxiv.org/abs/1703.01365))
 - **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 +66,7 @@ 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 +78,12 @@ From PyPI:
 pip install explainiverse
 ```
 
+With PyTorch support (for neural network explanations):
+
+```bash
+pip install explainiverse[torch]
+```
+
 For development:
 
 ```bash
@@ -100,7 +110,7 @@ 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', 'integrated_gradients', 'anchors', 'counterfactual', 'permutation_importance', 'partial_dependence', 'ale', 'sage']
 
 # Create and use an explainer
 explainer = default_registry.create(
@@ -119,11 +129,11 @@ 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', 'integrated_gradients', '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 +143,90 @@ 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")
+```
+
+### Integrated Gradients for Neural Networks
+
+```python
+from explainiverse.explainers import IntegratedGradientsExplainer
+from explainiverse import PyTorchAdapter
+
+# Wrap your PyTorch model
+adapter = PyTorchAdapter(model, task="classification", class_names=class_names)
+
+# Create IG explainer
+explainer = IntegratedGradientsExplainer(
+    model=adapter,
+    feature_names=feature_names,
+    class_names=class_names,
+    n_steps=50  # More steps = more accurate
+)
+
+# Explain a prediction
+explanation = explainer.explain(X_test[0])
+print(explanation.explanation_data["feature_attributions"])
+
+# Check convergence (sum of attributions ≈ F(x) - F(baseline))
+explanation = explainer.explain(X_test[0], return_convergence_delta=True)
+print(f"Convergence delta: {explanation.explanation_data['convergence_delta']}")
+```
+
 ### Using Specific Explainers
 
 ```python
@@ -233,12 +327,14 @@ poetry run pytest tests/test_new_explainers.py -v
 ## Roadmap
 
 - [x] LIME, SHAP (KernelSHAP)
+- [x] TreeSHAP (optimized for tree models) ✅
 - [x] Anchors, Counterfactuals
 - [x] Permutation Importance, PDP, ALE, SAGE
 - [x] Explainer Registry with filtering
-- [ ] TreeSHAP (optimized for tree models)
-- [ ] Integrated Gradients (gradient-based for neural nets)
-- [ ] PyTorch/TensorFlow adapters
+- [x] PyTorch Adapter ✅
+- [x] Integrated Gradients ✅ NEW
+- [ ] GradCAM for CNNs
+- [ ] TensorFlow adapter
 - [ ] Interactive visualization dashboard
 
 ---

{explainiverse-0.2.1.dist-info → explainiverse-0.2.3.dist-info}/RECORD

@@ -1,16 +1,17 @@
-explainiverse/__init__.py,sha256=XXf936B0hTTgCgBHnBAGYcNvYsU31yQrYDDaEABk8kQ,1207
-explainiverse/adapters/__init__.py,sha256=fNlWQ0VDjNqi4G4lwaJRTtL0wGVgvEE-4pZt6vOOjYU,322
+explainiverse/__init__.py,sha256=NmrLPOGZZPZTq1vY0G4gid5ZJWxsVGd3CfTXVIDvjaQ,1612
+explainiverse/adapters/__init__.py,sha256=HcQGISyp-YQ4jEj2IYveX_c9X5otLcTNWRnVRRhzRik,781
 explainiverse/adapters/base_adapter.py,sha256=Nqt0GeDn_-PjTyJcZsE8dRTulavqFQsv8sMYWS_ps-M,603
+explainiverse/adapters/pytorch_adapter.py,sha256=GTilJAR1VF_OgWG88qZoqlqefHaSXB3i9iOwCJkyHTg,13318
 explainiverse/adapters/sklearn_adapter.py,sha256=pzIBtMuqrG-6ZbUqUCMt7rSk3Ow0FgrY268FSweFvw4,958
 explainiverse/core/__init__.py,sha256=P3jHMnH5coFqTTO1w-gT-rurkCM1-9r3pF-055pbXMg,474
 explainiverse/core/explainer.py,sha256=Z9on-9VblYDlQx9oBm1BHpmAf_NsQajZ3qr-u48Aejo,784
 explainiverse/core/explanation.py,sha256=6zxFh_TH8tFHc-r_H5-WHQ05Sp1Kp2TxLz3gyFek5jo,881
-explainiverse/core/registry.py,sha256=_BXWi1fJY3cGjYA1Xn1DwvY91jbpJrpX6_8EVzrRT20,19876
+explainiverse/core/registry.py,sha256=AC8XDIdX2IGyx0KkmDajAjdo5YsrM3dcKvYoQu1vNCk,20711
 explainiverse/engine/__init__.py,sha256=1sZO8nH1mmwK2e-KUavBQm7zYDWUe27nyWoFy9tgsiA,197
 explainiverse/engine/suite.py,sha256=sq8SK_6Pf0qRckTmVJ7Mdosu9bhkjAGPGN8ymLGFP9E,4914
 explainiverse/evaluation/__init__.py,sha256=Y50L_b4HKthg4epwcayPHXh0l4i4MUuzvaNlqPmUNZY,212
 explainiverse/evaluation/metrics.py,sha256=tSBXtyA_-0zOGCGjlPZU6LdGKRH_QpWfgKa78sdlovs,7453
-explainiverse/explainers/__init__.py,sha256=Op-Z_BTJ7BdqA_9gTnruomN2-rKtrkPCt1Zq1iCzxr0,1758
+explainiverse/explainers/__init__.py,sha256=3yhamu1E2hpb0vE_hg3xK621YJdZYcy7gsSGgCT4Km4,1962
 explainiverse/explainers/attribution/__init__.py,sha256=YeVs9bS_IWDtqGbp6T37V6Zp5ZDWzLdAXHxxyFGpiQM,431
 explainiverse/explainers/attribution/lime_wrapper.py,sha256=OnXIV7t6yd-vt38sIi7XmHFbgzlZfCEbRlFyGGd5XiE,3245
 explainiverse/explainers/attribution/shap_wrapper.py,sha256=tKie5AvN7mb55PWOYdMvW0lUAYjfHPzYosEloEY2ZzI,3210
@@ -22,9 +23,11 @@ explainiverse/explainers/global_explainers/ale.py,sha256=tgG3XTppCf8LiD7uKzBt4DI
 explainiverse/explainers/global_explainers/partial_dependence.py,sha256=dH6yMjpwZads3pACR3rSykTbssLGHH7e6HfMlpl-S3I,6745
 explainiverse/explainers/global_explainers/permutation_importance.py,sha256=bcgKz1S_D3lrBMgpqEF_Z6qw8Knxl_cfR50hrSO2tBc,4410
 explainiverse/explainers/global_explainers/sage.py,sha256=57Xw1SK529x5JXWt0TVrcFYUUP3C65LfUwgoM-Z3gaw,5839
+explainiverse/explainers/gradient/__init__.py,sha256=Z4uSZcBhnHGp7DCd7bhcIMj_3f_uuCFw5AGA1JX6myQ,350
+explainiverse/explainers/gradient/integrated_gradients.py,sha256=feBgY3Vw2rDti7fxRZtLkxse75m2dbP_R05ARqo2BRM,13367
 explainiverse/explainers/rule_based/__init__.py,sha256=gKzlFCAzwurAMLJcuYgal4XhDj1thteBGcaHWmN7iWk,243
 explainiverse/explainers/rule_based/anchors_wrapper.py,sha256=ML7W6aam-eMGZHy5ilol8qupZvNBJpYAFatEEPnuMyo,13254
-explainiverse-0.2.1.dist-info/LICENSE,sha256=28rbHe8rJgmUlRdxJACfq1Sj-MtCEhyHxkJedQd1ZYA,1070
-explainiverse-0.2.1.dist-info/METADATA,sha256=k6JKGA2LJZaxxPVj_7e4cP99PoFMPEU0jBAY-fk2m3U,7731
-explainiverse-0.2.1.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-explainiverse-0.2.1.dist-info/RECORD,,
+explainiverse-0.2.3.dist-info/LICENSE,sha256=28rbHe8rJgmUlRdxJACfq1Sj-MtCEhyHxkJedQd1ZYA,1070
+explainiverse-0.2.3.dist-info/METADATA,sha256=TGuHUB9HZEcTbkQ7vmXl6ygm9arV5tlzufCHMoFqmdk,10465
+explainiverse-0.2.3.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+explainiverse-0.2.3.dist-info/RECORD,,