pyg-nightly 2.7.0.dev20250406__py3-none-any.whl → 2.7.0.dev20250407__py3-none-any.whl

{pyg_nightly-2.7.0.dev20250406.dist-info → pyg_nightly-2.7.0.dev20250407.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyg-nightly
-Version: 2.7.0.dev20250406
+Version: 2.7.0.dev20250407
 Summary: Graph Neural Network Library for PyTorch
 Keywords: deep-learning,pytorch,geometric-deep-learning,graph-neural-networks,graph-convolutional-networks
 Author-email: Matthias Fey <matthias@pyg.org>

{pyg_nightly-2.7.0.dev20250406.dist-info → pyg_nightly-2.7.0.dev20250407.dist-info}/RECORD

@@ -1,4 +1,4 @@
-torch_geometric/__init__.py,sha256=EFUlgJy_cHoHOgqO8KCynWIfRJFW8DFqG7O5v9DFOzI,1978
+torch_geometric/__init__.py,sha256=peFG3sVQB1R7kf4erqOCMV6_UO_k1PvDHdQg1HBaqog,1978
 torch_geometric/_compile.py,sha256=f-WQeH4VLi5Hn9lrgztFUCSrN_FImjhQa6BxFzcYC38,1338
 torch_geometric/_onnx.py,sha256=V9ffrIKSqhDw6xUZ12lkuSfNs48cQp2EeJ6Z19GfnVw,349
 torch_geometric/backend.py,sha256=lVaf7aLoVaB3M-UcByUJ1G4T4FOK6LXAg0CF4W3E8jo,1575
@@ -200,7 +200,7 @@ torch_geometric/explain/algorithm/base.py,sha256=wwJcREUFKDLFUDjRa9o4X3DWqQgMvhS
 torch_geometric/explain/algorithm/captum.py,sha256=k6hNgC5Kn9lVirOYVJzej8-hRuf5C2mPFUXFLd2wWsY,12857
 torch_geometric/explain/algorithm/captum_explainer.py,sha256=oz-c40hvdzii4_chEQPHzQo_dFjHr9HLuJhDLsqRIVU,7346
 torch_geometric/explain/algorithm/dummy_explainer.py,sha256=jvcVQmfngmUWgoKa5p7CXzju2HM5D5DfieJhZW3gbLc,2872
-torch_geometric/explain/algorithm/gnn_explainer.py,sha256=TRGwaKYn9nLn3fp0rSSzeGs9uHj2rZzfomMseDfq8O8,12454
+torch_geometric/explain/algorithm/gnn_explainer.py,sha256=iu45fGWdd4c6wNczWEAT-29HCAz7ncuoaS6cpx-xDJM,24660
 torch_geometric/explain/algorithm/graphmask_explainer.py,sha256=T2B081dK-JSpaQmutnkQd5xF3JF49_dPZCOgwqIKJDk,21367
 torch_geometric/explain/algorithm/pg_explainer.py,sha256=zPsl0tT9ISSWK1xP1KKpe1ZjUarhSB736WTtqwcmDIo,10372
 torch_geometric/explain/algorithm/utils.py,sha256=eh0ARPG41V7piVw5jdMYpV0p7WjTlpehnY-bWqPV_zg,2564
@@ -636,7 +636,7 @@ torch_geometric/utils/undirected.py,sha256=H_nfpI0_WluOG6VfjPyldvcjL4w5USAKWu2x5
 torch_geometric/visualization/__init__.py,sha256=PyR_4K5SafsJrBr6qWrkjKr6GBL1b7FtZybyXCDEVwY,154
 torch_geometric/visualization/graph.py,sha256=ZuLPL92yGRi7lxlqsUPwL_EVVXF7P2kMcveTtW79vpA,4784
 torch_geometric/visualization/influence.py,sha256=CWMvuNA_Nf1sfbJmQgn58yS4OFpeKXeZPe7kEuvkUBw,477
-pyg_nightly-2.7.0.dev20250406.dist-info/licenses/LICENSE,sha256=ic-27cMJc1kWoMEYncz3Ya3Ur2Bi3bNLWib2DT763-o,1067
-pyg_nightly-2.7.0.dev20250406.dist-info/WHEEL,sha256=G2gURzTEtmeR8nrdXUJfNiB3VYVxigPQ-bEQujpNiNs,82
-pyg_nightly-2.7.0.dev20250406.dist-info/METADATA,sha256=csAfUo5zCWohsFtqQsynpzuSflZRg7_4f9DB3JWSVWE,63021
-pyg_nightly-2.7.0.dev20250406.dist-info/RECORD,,
+pyg_nightly-2.7.0.dev20250407.dist-info/licenses/LICENSE,sha256=ic-27cMJc1kWoMEYncz3Ya3Ur2Bi3bNLWib2DT763-o,1067
+pyg_nightly-2.7.0.dev20250407.dist-info/WHEEL,sha256=G2gURzTEtmeR8nrdXUJfNiB3VYVxigPQ-bEQujpNiNs,82
+pyg_nightly-2.7.0.dev20250407.dist-info/METADATA,sha256=uPWrrVVadg-GQ1_t4bwTD4LTyNy8r0m9_g5BaKW1AVs,63021
+pyg_nightly-2.7.0.dev20250407.dist-info/RECORD,,

torch_geometric/__init__.py

@@ -31,7 +31,7 @@ from .lazy_loader import LazyLoader
 contrib = LazyLoader('contrib', globals(), 'torch_geometric.contrib')
 graphgym = LazyLoader('graphgym', globals(), 'torch_geometric.graphgym')
 
-__version__ = '2.7.0.dev20250406'
+__version__ = '2.7.0.dev20250407'
 
 __all__ = [
     'Index',

torch_geometric/explain/algorithm/gnn_explainer.py

@@ -1,14 +1,24 @@
 from math import sqrt
-from typing import Optional, Tuple, Union
+from typing import Dict, Optional, Tuple, Union, overload
 
 import torch
 from torch import Tensor
 from torch.nn.parameter import Parameter
 
-from torch_geometric.explain import ExplainerConfig, Explanation, ModelConfig
+from torch_geometric.explain import (
+    ExplainerConfig,
+    Explanation,
+    HeteroExplanation,
+    ModelConfig,
+)
 from torch_geometric.explain.algorithm import ExplainerAlgorithm
-from torch_geometric.explain.algorithm.utils import clear_masks, set_masks
+from torch_geometric.explain.algorithm.utils import (
+    clear_masks,
+    set_hetero_masks,
+    set_masks,
+)
 from torch_geometric.explain.config import MaskType, ModelMode, ModelTaskLevel
+from torch_geometric.typing import EdgeType, NodeType
 
 
 class GNNExplainer(ExplainerAlgorithm):
@@ -69,7 +79,9 @@ class GNNExplainer(ExplainerAlgorithm):
 
         self.node_mask = self.hard_node_mask = None
         self.edge_mask = self.hard_edge_mask = None
+        self.is_hetero = False
 
+    @overload
     def forward(
         self,
         model: torch.nn.Module,
@@ -80,30 +92,87 @@ class GNNExplainer(ExplainerAlgorithm):
         index: Optional[Union[int, Tensor]] = None,
         **kwargs,
     ) -> Explanation:
-        if isinstance(x, dict) or isinstance(edge_index, dict):
-            raise ValueError(f"Heterogeneous graphs not yet supported in "
-                             f"'{self.__class__.__name__}'")
+        ...
 
-        self._train(model, x, edge_index, target=target, index=index, **kwargs)
-
-        node_mask = self._post_process_mask(
-            self.node_mask,
-            self.hard_node_mask,
-            apply_sigmoid=True,
-        )
-        edge_mask = self._post_process_mask(
-            self.edge_mask,
-            self.hard_edge_mask,
-            apply_sigmoid=True,
-        )
+    @overload
+    def forward(
+        self,
+        model: torch.nn.Module,
+        x: Dict[NodeType, Tensor],
+        edge_index: Dict[EdgeType, Tensor],
+        *,
+        target: Tensor,
+        index: Optional[Union[int, Tensor]] = None,
+        **kwargs,
+    ) -> HeteroExplanation:
+        ...
 
+    def forward(
+        self,
+        model: torch.nn.Module,
+        x: Union[Tensor, Dict[NodeType, Tensor]],
+        edge_index: Union[Tensor, Dict[EdgeType, Tensor]],
+        *,
+        target: Tensor,
+        index: Optional[Union[int, Tensor]] = None,
+        **kwargs,
+    ) -> Union[Explanation, HeteroExplanation]:
+        self.is_hetero = isinstance(x, dict)
+        self._train(model, x, edge_index, target=target, index=index, **kwargs)
+        explanation = self._create_explanation()
         self._clean_model(model)
+        return explanation
+
+    def _create_explanation(self) -> Union[Explanation, HeteroExplanation]:
+        """Create an explanation object from the current masks."""
+        if self.is_hetero:
+            # For heterogeneous graphs, process each type separately
+            node_mask_dict = {}
+            edge_mask_dict = {}
+
+            for node_type, mask in self.node_mask.items():
+                if mask is not None:
+                    node_mask_dict[node_type] = self._post_process_mask(
+                        mask,
+                        self.hard_node_mask[node_type],
+                        apply_sigmoid=True,
+                    )
+
+            for edge_type, mask in self.edge_mask.items():
+                if mask is not None:
+                    edge_mask_dict[edge_type] = self._post_process_mask(
+                        mask,
+                        self.hard_edge_mask[edge_type],
+                        apply_sigmoid=True,
+                    )
+
+            # Create heterogeneous explanation
+            explanation = HeteroExplanation()
+            explanation.set_value_dict('node_mask', node_mask_dict)
+            explanation.set_value_dict('edge_mask', edge_mask_dict)
 
-        return Explanation(node_mask=node_mask, edge_mask=edge_mask)
+        else:
+            # For homogeneous graphs, process single masks
+            node_mask = self._post_process_mask(
+                self.node_mask,
+                self.hard_node_mask,
+                apply_sigmoid=True,
+            )
+            edge_mask = self._post_process_mask(
+                self.edge_mask,
+                self.hard_edge_mask,
+                apply_sigmoid=True,
+            )
+
+            # Create homogeneous explanation
+            explanation = Explanation(node_mask=node_mask, edge_mask=edge_mask)
+
+        return explanation
 
     def supports(self) -> bool:
         return True
 
+    @overload
     def _train(
         self,
         model: torch.nn.Module,
@@ -113,57 +182,222 @@ class GNNExplainer(ExplainerAlgorithm):
         target: Tensor,
         index: Optional[Union[int, Tensor]] = None,
         **kwargs,
-    ):
+    ) -> None:
+        ...
+
+    @overload
+    def _train(
+        self,
+        model: torch.nn.Module,
+        x: Dict[NodeType, Tensor],
+        edge_index: Dict[EdgeType, Tensor],
+        *,
+        target: Tensor,
+        index: Optional[Union[int, Tensor]] = None,
+        **kwargs,
+    ) -> None:
+        ...
+
+    def _train(
+        self,
+        model: torch.nn.Module,
+        x: Union[Tensor, Dict[NodeType, Tensor]],
+        edge_index: Union[Tensor, Dict[EdgeType, Tensor]],
+        *,
+        target: Tensor,
+        index: Optional[Union[int, Tensor]] = None,
+        **kwargs,
+    ) -> None:
+        # Initialize masks based on input type
         self._initialize_masks(x, edge_index)
 
-        parameters = []
-        if self.node_mask is not None:
-            parameters.append(self.node_mask)
-        if self.edge_mask is not None:
-            set_masks(model, self.edge_mask, edge_index, apply_sigmoid=True)
-            parameters.append(self.edge_mask)
+        # Collect parameters for optimization
+        parameters = self._collect_parameters(model, edge_index)
 
+        # Create optimizer
         optimizer = torch.optim.Adam(parameters, lr=self.lr)
 
+        # Training loop
         for i in range(self.epochs):
             optimizer.zero_grad()
 
-            h = x if self.node_mask is None else x * self.node_mask.sigmoid()
-            y_hat, y = model(h, edge_index, **kwargs), target
+            # Forward pass with masked inputs
+            y_hat = self._forward_with_masks(model, x, edge_index, **kwargs)
+            y = target
 
+            # Handle index if provided
             if index is not None:
                 y_hat, y = y_hat[index], y[index]
 
+            # Calculate loss
             loss = self._loss(y_hat, y)
 
+            # Backward pass
             loss.backward()
             optimizer.step()
 
-            # In the first iteration, we collect the nodes and edges that are
-            # involved into making the prediction. These are all the nodes and
-            # edges with gradient != 0 (without regularization applied).
-            if i == 0 and self.node_mask is not None:
-                if self.node_mask.grad is None:
-                    raise ValueError("Could not compute gradients for node "
-                                     "features. Please make sure that node "
-                                     "features are used inside the model or "
-                                     "disable it via `node_mask_type=None`.")
-                self.hard_node_mask = self.node_mask.grad != 0.0
-            if i == 0 and self.edge_mask is not None:
-                if self.edge_mask.grad is None:
-                    raise ValueError("Could not compute gradients for edges. "
-                                     "Please make sure that edges are used "
-                                     "via message passing inside the model or "
-                                     "disable it via `edge_mask_type=None`.")
-                self.hard_edge_mask = self.edge_mask.grad != 0.0
-
-    def _initialize_masks(self, x: Tensor, edge_index: Tensor):
+            # In the first iteration, collect gradients to identify important
+            # nodes/edges
+            if i == 0:
+                self._collect_gradients()
+
+    def _collect_parameters(self, model, edge_index):
+        """Collect parameters for optimization."""
+        parameters = []
+
+        if self.is_hetero:
+            # For heterogeneous graphs, collect parameters from all types
+            for mask in self.node_mask.values():
+                if mask is not None:
+                    parameters.append(mask)
+            if any(v is not None for v in self.edge_mask.values()):
+                set_hetero_masks(model, self.edge_mask, edge_index)
+            for mask in self.edge_mask.values():
+                if mask is not None:
+                    parameters.append(mask)
+        else:
+            # For homogeneous graphs, collect single parameters
+            if self.node_mask is not None:
+                parameters.append(self.node_mask)
+            if self.edge_mask is not None:
+                set_masks(model, self.edge_mask, edge_index,
+                          apply_sigmoid=True)
+                parameters.append(self.edge_mask)
+
+        return parameters
+
+    @overload
+    def _forward_with_masks(
+        self,
+        model: torch.nn.Module,
+        x: Tensor,
+        edge_index: Tensor,
+        **kwargs,
+    ) -> Tensor:
+        ...
+
+    @overload
+    def _forward_with_masks(
+        self,
+        model: torch.nn.Module,
+        x: Dict[NodeType, Tensor],
+        edge_index: Dict[EdgeType, Tensor],
+        **kwargs,
+    ) -> Tensor:
+        ...
+
+    def _forward_with_masks(
+        self,
+        model: torch.nn.Module,
+        x: Union[Tensor, Dict[NodeType, Tensor]],
+        edge_index: Union[Tensor, Dict[EdgeType, Tensor]],
+        **kwargs,
+    ) -> Tensor:
+        """Forward pass with masked inputs."""
+        if self.is_hetero:
+            # Apply masks to heterogeneous inputs
+            h_dict = {}
+            for node_type, features in x.items():
+                if node_type in self.node_mask and self.node_mask[
+                        node_type] is not None:
+                    h_dict[node_type] = features * self.node_mask[
+                        node_type].sigmoid()
+                else:
+                    h_dict[node_type] = features
+
+            # Forward pass with masked features
+            return model(h_dict, edge_index, **kwargs)
+        else:
+            # Apply mask to homogeneous input
+            h = x if self.node_mask is None else x * self.node_mask.sigmoid()
+
+            # Forward pass with masked features
+            return model(h, edge_index, **kwargs)
+
+    def _initialize_masks(
+        self,
+        x: Union[Tensor, Dict[NodeType, Tensor]],
+        edge_index: Union[Tensor, Dict[EdgeType, Tensor]],
+    ) -> None:
         node_mask_type = self.explainer_config.node_mask_type
         edge_mask_type = self.explainer_config.edge_mask_type
 
-        device = x.device
-        (N, F), E = x.size(), edge_index.size(1)
+        if self.is_hetero:
+            # Initialize dictionaries for heterogeneous masks
+            self.node_mask = {}
+            self.hard_node_mask = {}
+            self.edge_mask = {}
+            self.hard_edge_mask = {}
+
+            # Initialize node masks for each node type
+            for node_type, features in x.items():
+                device = features.device
+                N, F = features.size()
+                self._initialize_node_mask(node_mask_type, node_type, N, F,
+                                           device)
+
+            # Initialize edge masks for each edge type
+            for edge_type, indices in edge_index.items():
+                device = indices.device
+                E = indices.size(1)
+                N = max(indices.max().item() + 1,
+                        max(feat.size(0) for feat in x.values()))
+                self._initialize_edge_mask(edge_mask_type, edge_type, E, N,
+                                           device)
+        else:
+            # Initialize masks for homogeneous graph
+            device = x.device
+            (N, F), E = x.size(), edge_index.size(1)
+
+            # Initialize homogeneous node and edge masks
+            self._initialize_homogeneous_masks(node_mask_type, edge_mask_type,
+                                               N, F, E, device)
+
+    def _initialize_node_mask(
+        self,
+        node_mask_type,
+        node_type,
+        N,
+        F,
+        device,
+    ) -> None:
+        """Initialize node mask for a specific node type."""
+        std = 0.1
+        if node_mask_type is None:
+            self.node_mask[node_type] = None
+            self.hard_node_mask[node_type] = None
+        elif node_mask_type == MaskType.object:
+            self.node_mask[node_type] = Parameter(
+                torch.randn(N, 1, device=device) * std)
+            self.hard_node_mask[node_type] = None
+        elif node_mask_type == MaskType.attributes:
+            self.node_mask[node_type] = Parameter(
+                torch.randn(N, F, device=device) * std)
+            self.hard_node_mask[node_type] = None
+        elif node_mask_type == MaskType.common_attributes:
+            self.node_mask[node_type] = Parameter(
+                torch.randn(1, F, device=device) * std)
+            self.hard_node_mask[node_type] = None
+        else:
+            raise ValueError(f"Invalid node mask type: {node_mask_type}")
+
+    def _initialize_edge_mask(self, edge_mask_type, edge_type, E, N, device):
+        """Initialize edge mask for a specific edge type."""
+        if edge_mask_type is None:
+            self.edge_mask[edge_type] = None
+            self.hard_edge_mask[edge_type] = None
+        elif edge_mask_type == MaskType.object:
+            std = torch.nn.init.calculate_gain('relu') * sqrt(2.0 / (2 * N))
+            self.edge_mask[edge_type] = Parameter(
+                torch.randn(E, device=device) * std)
+            self.hard_edge_mask[edge_type] = None
+        else:
+            raise ValueError(f"Invalid edge mask type: {edge_mask_type}")
 
+    def _initialize_homogeneous_masks(self, node_mask_type, edge_mask_type, N,
+                                      F, E, device):
+        """Initialize masks for homogeneous graph."""
+        # Initialize node mask
         std = 0.1
         if node_mask_type is None:
             self.node_mask = None
@@ -174,43 +408,145 @@ class GNNExplainer(ExplainerAlgorithm):
         elif node_mask_type == MaskType.common_attributes:
             self.node_mask = Parameter(torch.randn(1, F, device=device) * std)
         else:
-            assert False
+            raise ValueError(f"Invalid node mask type: {node_mask_type}")
 
+        # Initialize edge mask
         if edge_mask_type is None:
             self.edge_mask = None
         elif edge_mask_type == MaskType.object:
             std = torch.nn.init.calculate_gain('relu') * sqrt(2.0 / (2 * N))
             self.edge_mask = Parameter(torch.randn(E, device=device) * std)
         else:
-            assert False
+            raise ValueError(f"Invalid edge mask type: {edge_mask_type}")
+
+    def _collect_gradients(self) -> None:
+        if self.is_hetero:
+            self._collect_hetero_gradients()
+        else:
+            self._collect_homo_gradients()
+
+    def _collect_hetero_gradients(self):
+        """Collect gradients for heterogeneous graph."""
+        for node_type, mask in self.node_mask.items():
+            if mask is not None:
+                if mask.grad is None:
+                    raise ValueError(
+                        f"Could not compute gradients for node masks of type "
+                        f"'{node_type}'. Please make sure that node masks are "
+                        f"used inside the model or disable it via "
+                        f"`node_mask_type=None`.")
+
+                self.hard_node_mask[node_type] = mask.grad != 0.0
+
+        for edge_type, mask in self.edge_mask.items():
+            if mask is not None:
+                if mask.grad is None:
+                    raise ValueError(
+                        f"Could not compute gradients for edge masks of type "
+                        f"'{edge_type}'. Please make sure that edge masks are "
+                        f"used inside the model or disable it via "
+                        f"`edge_mask_type=None`.")
+                self.hard_edge_mask[edge_type] = mask.grad != 0.0
+
+    def _collect_homo_gradients(self):
+        """Collect gradients for homogeneous graph."""
+        if self.node_mask is not None:
+            if self.node_mask.grad is None:
+                raise ValueError("Could not compute gradients for node "
+                                 "features. Please make sure that node "
+                                 "features are used inside the model or "
+                                 "disable it via `node_mask_type=None`.")
+            self.hard_node_mask = self.node_mask.grad != 0.0
+
+        if self.edge_mask is not None:
+            if self.edge_mask.grad is None:
+                raise ValueError("Could not compute gradients for edges. "
+                                 "Please make sure that edges are used "
+                                 "via message passing inside the model or "
+                                 "disable it via `edge_mask_type=None`.")
+            self.hard_edge_mask = self.edge_mask.grad != 0.0
 
     def _loss(self, y_hat: Tensor, y: Tensor) -> Tensor:
+        # Calculate base loss based on model configuration
+        loss = self._calculate_base_loss(y_hat, y)
+
+        # Apply regularization based on graph type
+        if self.is_hetero:
+            # Apply regularization for heterogeneous graph
+            loss = self._apply_hetero_regularization(loss)
+        else:
+            # Apply regularization for homogeneous graph
+            loss = self._apply_homo_regularization(loss)
+
+        return loss
+
+    def _calculate_base_loss(self, y_hat, y):
+        """Calculate base loss based on model configuration."""
         if self.model_config.mode == ModelMode.binary_classification:
-            loss = self._loss_binary_classification(y_hat, y)
+            return self._loss_binary_classification(y_hat, y)
         elif self.model_config.mode == ModelMode.multiclass_classification:
-            loss = self._loss_multiclass_classification(y_hat, y)
+            return self._loss_multiclass_classification(y_hat, y)
         elif self.model_config.mode == ModelMode.regression:
-            loss = self._loss_regression(y_hat, y)
+            return self._loss_regression(y_hat, y)
         else:
-            assert False
+            raise ValueError(f"Invalid model mode: {self.model_config.mode}")
+
+    def _apply_hetero_regularization(self, loss):
+        """Apply regularization for heterogeneous graph."""
+        # Apply regularization for each edge type
+        for edge_type, mask in self.edge_mask.items():
+            if (mask is not None
+                    and self.hard_edge_mask[edge_type] is not None):
+                loss = self._add_mask_regularization(
+                    loss, mask, self.hard_edge_mask[edge_type],
+                    self.coeffs['edge_size'], self.coeffs['edge_reduction'],
+                    self.coeffs['edge_ent'])
+
+        # Apply regularization for each node type
+        for node_type, mask in self.node_mask.items():
+            if (mask is not None
+                    and self.hard_node_mask[node_type] is not None):
+                loss = self._add_mask_regularization(
+                    loss, mask, self.hard_node_mask[node_type],
+                    self.coeffs['node_feat_size'],
+                    self.coeffs['node_feat_reduction'],
+                    self.coeffs['node_feat_ent'])
 
+        return loss
+
+    def _apply_homo_regularization(self, loss):
+        """Apply regularization for homogeneous graph."""
+        # Apply regularization for edge mask
         if self.hard_edge_mask is not None:
             assert self.edge_mask is not None
-            m = self.edge_mask[self.hard_edge_mask].sigmoid()
-            edge_reduce = getattr(torch, self.coeffs['edge_reduction'])
-            loss = loss + self.coeffs['edge_size'] * edge_reduce(m)
-            ent = -m * torch.log(m + self.coeffs['EPS']) - (
-                1 - m) * torch.log(1 - m + self.coeffs['EPS'])
-            loss = loss + self.coeffs['edge_ent'] * ent.mean()
+            loss = self._add_mask_regularization(loss, self.edge_mask,
+                                                 self.hard_edge_mask,
+                                                 self.coeffs['edge_size'],
+                                                 self.coeffs['edge_reduction'],
+                                                 self.coeffs['edge_ent'])
 
+        # Apply regularization for node mask
         if self.hard_node_mask is not None:
             assert self.node_mask is not None
-            m = self.node_mask[self.hard_node_mask].sigmoid()
-            node_reduce = getattr(torch, self.coeffs['node_feat_reduction'])
-            loss = loss + self.coeffs['node_feat_size'] * node_reduce(m)
-            ent = -m * torch.log(m + self.coeffs['EPS']) - (
-                1 - m) * torch.log(1 - m + self.coeffs['EPS'])
-            loss = loss + self.coeffs['node_feat_ent'] * ent.mean()
+            loss = self._add_mask_regularization(
+                loss, self.node_mask, self.hard_node_mask,
+                self.coeffs['node_feat_size'],
+                self.coeffs['node_feat_reduction'],
+                self.coeffs['node_feat_ent'])
+
+        return loss
+
+    def _add_mask_regularization(self, loss, mask, hard_mask, size_coeff,
+                                 reduction_name, ent_coeff):
+        """Add size and entropy regularization for a mask."""
+        m = mask[hard_mask].sigmoid()
+        reduce_fn = getattr(torch, reduction_name)
+        # Add size regularization
+        loss = loss + size_coeff * reduce_fn(m)
+        # Add entropy regularization
+        ent = -m * torch.log(m + self.coeffs['EPS']) - (
+            1 - m) * torch.log(1 - m + self.coeffs['EPS'])
+        loss = loss + ent_coeff * ent.mean()
 
         return loss