nextrec 0.3.6__py3-none-any.whl → 0.4.2__py3-none-any.whl

nextrec/models/ranking/afm.py

@@ -1,17 +1,46 @@
 """
 Date: create on 09/11/2025
-Author:
-    Yang Zhou,zyaztec@gmail.com
+Checkpoint: edit on 06/12/2025
+Author: Yang Zhou,zyaztec@gmail.com
 Reference:
-    [1] Xiao J, Ye H, He X, et al. Attentional factorization machines: Learning the weight of
-        feature interactions via attention networks[C]//IJCAI. 2017: 3119-3125.
+[1] Xiao J, Ye H, He X, et al. Attentional factorization machines: Learning the weight of
+feature interactions via attention networks[C]//IJCAI. 2017: 3119-3125.
+
+Attentional Factorization Machine (AFM) builds on FM by learning an importance
+weight for every second-order interaction instead of treating all pairs equally.
+It retains FM’s linear (first-order) component for sparsity-friendly modeling,
+while using an attention network to reweight the element-wise product of field
+embeddings before aggregation.
+
+In each forward pass:
+  (1) Embed each field and compute pairwise element-wise products v_i ⊙ v_j
+  (2) Pass interactions through an attention MLP (ReLU + projection) to score them
+  (3) Softmax-normalize scores to obtain interaction weights
+  (4) Weighted sum of interactions -> linear projection -> add FM first-order term
+
+Key Advantages:
+- Learns which feature pairs contribute most via attention weights
+- Keeps FM efficiency and interpretability by preserving first-order terms
+- Softmax-normalized reweighting reduces noise from uninformative interactions
+
+AFM 在 FM 的二阶交互上引入注意力，为每个特征对学习重要性权重；同时保留 FM 的一阶项，
+保持对稀疏特征的友好与可解释性。具体流程：
+  (1) 对各字段做 embedding，并计算所有特征对的元素积 v_i ⊙ v_j
+  (2) 经由注意力 MLP（ReLU + 线性映射）得到交互得分
+  (3) 通过 softmax 归一化交互得分，得到权重
+  (4) 将加权交互求和、线性映射，再与一阶项相加得到最终预测
+
+主要优点：
+- 注意力显式告诉哪些特征对更重要
+- 保留 FM 的效率和可解释性
+- softmax 归一化减弱噪声交互的影响
 """
 
 import torch
 import torch.nn as nn
 
 from nextrec.basic.model import BaseModel
-from nextrec.basic.layers import EmbeddingLayer, LR, PredictionLayer
+from nextrec.basic.layers import EmbeddingLayer, PredictionLayer, InputMask
 from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature
 
 
@@ -21,69 +50,113 @@ class AFM(BaseModel):
         return "AFM"
 
     @property
-    def task_type(self):
+    def default_task(self):
         return "binary"
-    
-    def __init__(self,
-                 dense_features: list[DenseFeature] | list = [],
-                 sparse_features: list[SparseFeature] | list = [],
-                 sequence_features: list[SequenceFeature] | list = [],
-                 attention_dim: int = 32,
-                 attention_dropout: float = 0.0,
-                 target: list[str] | list = [],
-                 optimizer: str = "adam",
-                 optimizer_params: dict = {},
-                 loss: str | nn.Module | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 device: str = 'cpu',
-                 embedding_l1_reg=1e-6,
-                 dense_l1_reg=1e-5,
-                 embedding_l2_reg=1e-5,
-                 dense_l2_reg=1e-4, **kwargs):
-        
+
+    def __init__(
+        self,
+        dense_features: list[DenseFeature] | list = [],
+        sparse_features: list[SparseFeature] | list = [],
+        sequence_features: list[SequenceFeature] | list = [],
+        attention_dim: int = 32,
+        attention_dropout: float = 0.0,
+        target: list[str] | list = [],
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict = {},
+        loss: str | nn.Module | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        device: str = "cpu",
+        embedding_l1_reg=1e-6,
+        dense_l1_reg=1e-5,
+        embedding_l2_reg=1e-5,
+        dense_l2_reg=1e-4,
+        **kwargs,
+    ):
+
         super(AFM, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
             sequence_features=sequence_features,
             target=target,
-            task=self.task_type,
+            task=task or self.default_task,
             device=device,
             embedding_l1_reg=embedding_l1_reg,
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            early_stop_patience=20,
-            **kwargs
+            **kwargs,
         )
 
-        self.loss = loss
-        if self.loss is None:
-            self.loss = "bce"
-            
+        if target is None:
+            target = []
+        if optimizer_params is None:
+            optimizer_params = {}
+        if loss is None:
+            loss = "bce"
+
         self.fm_features = sparse_features + sequence_features
         if len(self.fm_features) < 2:
-            raise ValueError("AFM requires at least two sparse/sequence features to build pairwise interactions.")
+            raise ValueError(
+                "AFM requires at least two sparse/sequence features to build pairwise interactions."
+            )
 
-        # Assume uniform embedding dimension across FM fields
+        # make sure all embedding dimension are the same for FM features
         self.embedding_dim = self.fm_features[0].embedding_dim
         if any(f.embedding_dim != self.embedding_dim for f in self.fm_features):
-            raise ValueError("All FM features must share the same embedding_dim for AFM.")
-
-        self.embedding = EmbeddingLayer(features=self.fm_features)
+            raise ValueError(
+                "All FM features must share the same embedding_dim for AFM."
+            )
+
+        self.embedding = EmbeddingLayer(
+            features=self.fm_features
+        )  # [Batch, Field, Dim ]
+
+        # First-order terms: dense linear + one hot embeddings
+        self.dense_features = list(dense_features)
+        dense_input_dim = sum([f.input_dim for f in self.dense_features])
+        self.linear_dense = (
+            nn.Linear(dense_input_dim, 1, bias=True) if dense_input_dim > 0 else None
+        )
 
-        fm_input_dim = sum([f.embedding_dim for f in self.fm_features])
-        self.linear = LR(fm_input_dim)
+        # First-order term: sparse/sequence features one-hot
+        # **INFO**: source paper does not contain sequence features in experiments,
+        # but we implement it here for completeness. if you want follow the paper strictly,
+        # remove sequence features from fm_features.
+        self.first_order_embeddings = nn.ModuleDict()
+        for feature in self.fm_features:
+            if (
+                feature.embedding_name in self.first_order_embeddings
+            ):  # shared embedding
+                continue
+            emb = nn.Embedding(
+                num_embeddings=feature.vocab_size,
+                embedding_dim=1,
+                padding_idx=feature.padding_idx,
+            )  # equal to one-hot encoding weight
+            # nn.init.zeros_(emb.weight)
+            self.first_order_embeddings[feature.embedding_name] = emb
 
         self.attention_linear = nn.Linear(self.embedding_dim, attention_dim)
         self.attention_p = nn.Linear(attention_dim, 1, bias=False)
         self.attention_dropout = nn.Dropout(attention_dropout)
         self.output_projection = nn.Linear(self.embedding_dim, 1, bias=False)
-        self.prediction_layer = PredictionLayer(task_type=self.task_type)
+        self.prediction_layer = PredictionLayer(task_type=self.default_task)
+        self.input_mask = InputMask()
 
         # Register regularization weights
         self.register_regularization_weights(
-            embedding_attr='embedding',
-            include_modules=['linear', 'attention_linear', 'attention_p', 'output_projection']
+            embedding_attr="embedding",
+            include_modules=[
+                "linear_dense",
+                "attention_linear",
+                "attention_p",
+                "output_projection",
+            ],
+        )
+        # add first-order embeddings to embedding regularization list
+        self.embedding_params.extend(
+            emb.weight for emb in self.first_order_embeddings.values()
         )
 
         self.compile(
@@ -94,11 +167,65 @@ class AFM(BaseModel):
         )
 
     def forward(self, x):
-        field_emb = self.embedding(x=x, features=self.fm_features, squeeze_dim=False)  # [B, F, D]
-        input_linear = field_emb.flatten(start_dim=1)
-        y_linear = self.linear(input_linear)
+        field_emb = self.embedding(
+            x=x, features=self.fm_features, squeeze_dim=False
+        )  # [B, F, D]
+        batch_size = field_emb.size(0)
+        y_linear = torch.zeros(batch_size, 1, device=field_emb.device)
+
+        # First-order dense part
+        if self.linear_dense is not None:
+            dense_inputs = [
+                x[f.name].float().view(batch_size, -1) for f in self.dense_features
+            ]
+            dense_stack = torch.cat(dense_inputs, dim=1) if dense_inputs else None
+            if dense_stack is not None:
+                y_linear = y_linear + self.linear_dense(dense_stack)
+
+        # First-order sparse/sequence part
+        first_order_terms = []
+        for feature in self.fm_features:
+            emb = self.first_order_embeddings[feature.embedding_name]
+            if isinstance(feature, SparseFeature):
+                term = emb(x[feature.name].long())  # [B, 1]
+            else:  # SequenceFeature
+                seq_input = x[feature.name].long()  # [B, 1]
+                if feature.max_len is not None and seq_input.size(1) > feature.max_len:
+                    seq_input = seq_input[:, -feature.max_len :]
+                mask = self.input_mask(x, feature, seq_input).squeeze(1)  # [B, 1]
+                seq_weight = emb(seq_input).squeeze(-1)  # [B, L]
+                term = (seq_weight * mask).sum(dim=1, keepdim=True)  # [B, 1]
+            first_order_terms.append(term)
+        if first_order_terms:
+            y_linear = y_linear + torch.sum(
+                torch.cat(first_order_terms, dim=1), dim=1, keepdim=True
+            )
 
         interactions = []
+        feature_values = []
+        for feature in self.fm_features:
+            value = x.get(f"{feature.name}_value")
+            if value is not None:
+                value = value.float()
+                if value.dim() == 1:
+                    value = value.unsqueeze(-1)
+            else:
+                if isinstance(feature, SequenceFeature):
+                    seq_input = x[feature.name].long()
+                    if (
+                        feature.max_len is not None
+                        and seq_input.size(1) > feature.max_len
+                    ):
+                        seq_input = seq_input[:, -feature.max_len :]
+                    value = self.input_mask(x, feature, seq_input).sum(dim=2)  # [B, 1]
+                else:
+                    value = torch.ones(batch_size, 1, device=field_emb.device)
+            feature_values.append(value)
+        feature_values_tensor = torch.cat(feature_values, dim=1).unsqueeze(
+            -1
+        )  # [B, F, 1]
+        field_emb = field_emb * feature_values_tensor
+
         num_fields = field_emb.shape[1]
         for i in range(num_fields - 1):
             vi = field_emb[:, i, :]
@@ -107,7 +234,7 @@ class AFM(BaseModel):
                 interactions.append(vi * vj)
 
         pair_tensor = torch.stack(interactions, dim=1)  # [B, num_pairs, D]
-        attention_scores = torch.tanh(self.attention_linear(pair_tensor))
+        attention_scores = torch.relu(self.attention_linear(pair_tensor))
         attention_scores = self.attention_p(attention_scores)  # [B, num_pairs, 1]
         attention_weights = torch.softmax(attention_scores, dim=1)
 

nextrec/models/ranking/autoint.py

@@ -3,8 +3,8 @@ Date: create on 09/11/2025
 Checkpoint: edit on 24/11/2025
 Author: Yang Zhou,zyaztec@gmail.com
 Reference:
-[1] Song W, Shi C, Xiao Z, et al. Autoint: Automatic feature interaction learning via 
-self-attentive neural networks[C]//Proceedings of the 28th ACM international conference 
+[1] Song W, Shi C, Xiao Z, et al. Autoint: Automatic feature interaction learning via
+self-attentive neural networks[C]//Proceedings of the 28th ACM international conference
 on information and knowledge management. 2019: 1161-1170.
 (https://arxiv.org/abs/1810.11921)
 
@@ -68,91 +68,106 @@ class AutoInt(BaseModel):
         return "AutoInt"
 
     @property
-    def task_type(self):
+    def default_task(self):
         return "binary"
-    
-    def __init__(self,
-                 dense_features: list[DenseFeature],
-                 sparse_features: list[SparseFeature],
-                 sequence_features: list[SequenceFeature],
-                 att_layer_num: int = 3,
-                 att_embedding_dim: int = 8,
-                 att_head_num: int = 2,
-                 att_dropout: float = 0.0,
-                 att_use_residual: bool = True,
-                 target: list[str] = [],
-                 optimizer: str = "adam",
-                 optimizer_params: dict = {},
-                 loss: str | nn.Module | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 device: str = 'cpu',
-                 embedding_l1_reg=1e-6,
-                 dense_l1_reg=1e-5,
-                 embedding_l2_reg=1e-5,
-                 dense_l2_reg=1e-4,
-                 **kwargs):
-        
+
+    def __init__(
+        self,
+        dense_features: list[DenseFeature],
+        sparse_features: list[SparseFeature],
+        sequence_features: list[SequenceFeature],
+        att_layer_num: int = 3,
+        att_embedding_dim: int = 8,
+        att_head_num: int = 2,
+        att_dropout: float = 0.0,
+        att_use_residual: bool = True,
+        target: list[str] | None = None,
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict | None = None,
+        loss: str | nn.Module | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        device: str = "cpu",
+        embedding_l1_reg=1e-6,
+        dense_l1_reg=1e-5,
+        embedding_l2_reg=1e-5,
+        dense_l2_reg=1e-4,
+        **kwargs,
+    ):
+
         super(AutoInt, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
             sequence_features=sequence_features,
             target=target,
-            task=self.task_type,
+            task=task or self.default_task,
             device=device,
             embedding_l1_reg=embedding_l1_reg,
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            early_stop_patience=20,
-            **kwargs
+            **kwargs,
         )
 
-        self.loss = loss
-        if self.loss is None:
-            self.loss = "bce"
-            
+        if target is None:
+            target = []
+        if optimizer_params is None:
+            optimizer_params = {}
+        if loss is None:
+            loss = "bce"
+
         self.att_layer_num = att_layer_num
         self.att_embedding_dim = att_embedding_dim
-        
+
         # Use sparse and sequence features for interaction
-        self.interaction_features = dense_features + sparse_features + sequence_features 
-        
+        # **INFO**: this is different from the original paper, we also include dense features
+        # if you want to follow the paper strictly, set dense_features=[]
+        # or modify the code accordingly
+        self.interaction_features = dense_features + sparse_features + sequence_features
+
         # All features for embedding
         self.all_features = dense_features + sparse_features + sequence_features
 
         # Embedding layer
         self.embedding = EmbeddingLayer(features=self.all_features)
-        
+
         # Project embeddings to attention embedding dimension
         num_fields = len(self.interaction_features)
-        
+
         # If embeddings have different dimensions, project them to att_embedding_dim
-        self.need_projection = not all(f.embedding_dim == att_embedding_dim for f in self.interaction_features)
+        self.need_projection = not all(
+            f.embedding_dim == att_embedding_dim for f in self.interaction_features
+        )
         self.projection_layers = None
         if self.need_projection:
-            self.projection_layers = nn.ModuleList([
-                nn.Linear(f.embedding_dim, att_embedding_dim, bias=False) 
-                for f in self.interaction_features
-            ])
-        
+            self.projection_layers = nn.ModuleList(
+                [
+                    nn.Linear(f.embedding_dim, att_embedding_dim, bias=False)
+                    for f in self.interaction_features
+                ]
+            )
+
         # Multi-head self-attention layers
-        self.attention_layers = nn.ModuleList([
-            MultiHeadSelfAttention(
-                embedding_dim=att_embedding_dim,
-                num_heads=att_head_num,
-                dropout=att_dropout,
-                use_residual=att_use_residual
-            ) for _ in range(att_layer_num)
-        ])
-        
+        self.attention_layers = nn.ModuleList(
+            [
+                MultiHeadSelfAttention(
+                    embedding_dim=att_embedding_dim,
+                    num_heads=att_head_num,
+                    dropout=att_dropout,
+                    use_residual=att_use_residual,
+                )
+                for _ in range(att_layer_num)
+            ]
+        )
+
         # Final prediction layer
         self.fc = nn.Linear(num_fields * att_embedding_dim, 1)
-        self.prediction_layer = PredictionLayer(task_type=self.task_type)
+        self.prediction_layer = PredictionLayer(task_type=self.default_task)
 
         # Register regularization weights
         self.register_regularization_weights(
-            embedding_attr='embedding',
-            include_modules=['projection_layers', 'attention_layers', 'fc']
+            embedding_attr="embedding",
+            include_modules=["projection_layers", "attention_layers", "fc"],
         )
 
         self.compile(
@@ -166,21 +181,29 @@ class AutoInt(BaseModel):
         # Get embeddings field-by-field so mixed dimensions can be projected safely
         field_embeddings = []
         if len(self.interaction_features) == 0:
-            raise ValueError("AutoInt requires at least one sparse or sequence feature for interactions.")
+            raise ValueError(
+                "AutoInt requires at least one sparse or sequence feature for interactions."
+            )
         for idx, feature in enumerate(self.interaction_features):
             feature_emb = self.embedding(x=x, features=[feature], squeeze_dim=False)
             feature_emb = feature_emb.squeeze(1)  # [B, embedding_dim]
             if self.need_projection and self.projection_layers is not None:
                 feature_emb = self.projection_layers[idx](feature_emb)
-            field_embeddings.append(feature_emb.unsqueeze(1))  # [B, 1, att_embedding_dim or original_dim]
+            field_embeddings.append(
+                feature_emb.unsqueeze(1)
+            )  # [B, 1, att_embedding_dim or original_dim]
         embeddings = torch.cat(field_embeddings, dim=1)
-        
+
         # Apply multi-head self-attention layers
         attention_output = embeddings
         for att_layer in self.attention_layers:
-            attention_output = att_layer(attention_output)  # [B, num_fields, att_embedding_dim]
-        
+            attention_output = att_layer(
+                attention_output
+            )  # [B, num_fields, att_embedding_dim]
+
         # Flatten and predict
-        attention_output_flat = attention_output.flatten(start_dim=1)  # [B, num_fields * att_embedding_dim]
+        attention_output_flat = attention_output.flatten(
+            start_dim=1
+        )  # [B, num_fields * att_embedding_dim]
         y = self.fc(attention_output_flat)  # [B, 1]
         return self.prediction_layer(y)

nextrec/models/ranking/dcn.py

@@ -15,24 +15,26 @@ from nextrec.basic.model import BaseModel
 from nextrec.basic.layers import EmbeddingLayer, MLP, PredictionLayer
 from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature
 
+
 class CrossNetwork(nn.Module):
     """Stacked Cross Layers from DCN (Wang et al., 2017)."""
 
     def __init__(self, input_dim, num_layers):
         super().__init__()
         self.num_layers = num_layers
-        self.w = torch.nn.ModuleList([torch.nn.Linear(input_dim, 1, bias=False) for _ in range(num_layers)])
-        self.b = torch.nn.ParameterList([torch.nn.Parameter(torch.zeros((input_dim,))) for _ in range(num_layers)])
+        self.w = torch.nn.ModuleList(
+            [torch.nn.Linear(input_dim, 1, bias=False) for _ in range(num_layers)]
+        )
+        self.b = torch.nn.ParameterList(
+            [torch.nn.Parameter(torch.zeros((input_dim,))) for _ in range(num_layers)]
+        )
 
     def forward(self, x):
-        """
-        :param x: Float tensor of size ``(batch_size, num_fields, embed_dim)``
-        """
         x0 = x
         for i in range(self.num_layers):
             xw = self.w[i](x)
             x = x0 * xw + self.b[i] + x
-        return x
+        return x  # [batch_size, input_dim]
 
 
 class DCN(BaseModel):
@@ -41,46 +43,48 @@ class DCN(BaseModel):
         return "DCN"
 
     @property
-    def task_type(self):
+    def default_task(self):
         return "binary"
-    
-    def __init__(self,
-                 dense_features: list[DenseFeature],
-                 sparse_features: list[SparseFeature],
-                 sequence_features: list[SequenceFeature],
-                 cross_num: int = 3,
-                 mlp_params: dict | None = None,
-                 target: list[str] = [],
-                 optimizer: str = "adam",
-                 optimizer_params: dict = {},
-                 loss: str | nn.Module | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 device: str = 'cpu',
-                 embedding_l1_reg=1e-6,
-                 dense_l1_reg=1e-5,
-                 embedding_l2_reg=1e-5,
-                 dense_l2_reg=1e-4,
-                 **kwargs):
-        
+
+    def __init__(
+        self,
+        dense_features: list[DenseFeature],
+        sparse_features: list[SparseFeature],
+        sequence_features: list[SequenceFeature],
+        cross_num: int = 3,
+        mlp_params: dict | None = None,
+        target: list[str] = [],
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict = {},
+        loss: str | nn.Module | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        device: str = "cpu",
+        embedding_l1_reg=1e-6,
+        dense_l1_reg=1e-5,
+        embedding_l2_reg=1e-5,
+        dense_l2_reg=1e-4,
+        **kwargs,
+    ):
+
         super(DCN, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
             sequence_features=sequence_features,
             target=target,
-            task=self.task_type,
+            task=task or self.default_task,
             device=device,
             embedding_l1_reg=embedding_l1_reg,
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            early_stop_patience=20,
-            **kwargs
+            **kwargs,
         )
 
         self.loss = loss
         if self.loss is None:
             self.loss = "bce"
-            
+
         # All features
         self.all_features = dense_features + sparse_features + sequence_features
 
@@ -88,30 +92,41 @@ class DCN(BaseModel):
         self.embedding = EmbeddingLayer(features=self.all_features)
 
         # Calculate input dimension
-        emb_dim_total = sum([f.embedding_dim for f in self.all_features if not isinstance(f, DenseFeature)])
-        dense_input_dim = sum([getattr(f, "embedding_dim", 1) or 1 for f in dense_features])
+        emb_dim_total = sum(
+            [
+                f.embedding_dim
+                for f in self.all_features
+                if not isinstance(f, DenseFeature)
+            ]
+        )
+        dense_input_dim = sum(
+            [getattr(f, "embedding_dim", 1) or 1 for f in dense_features]
+        )
         input_dim = emb_dim_total + dense_input_dim
-        
+
         # Cross Network
         self.cross_network = CrossNetwork(input_dim=input_dim, num_layers=cross_num)
-        
+
         # Deep Network (optional)
         if mlp_params is not None:
             self.use_dnn = True
             self.mlp = MLP(input_dim=input_dim, **mlp_params)
+            deep_dim = self.mlp.output_dim
             # Final layer combines cross and deep
-            self.final_layer = nn.Linear(input_dim + 1, 1)  # +1 for MLP output
+            self.final_layer = nn.Linear(
+                input_dim + deep_dim, 1
+            )  # + deep_dim for MLP output
         else:
             self.use_dnn = False
             # Final layer only uses cross network output
             self.final_layer = nn.Linear(input_dim, 1)
 
-        self.prediction_layer = PredictionLayer(task_type=self.task_type)
+        self.prediction_layer = PredictionLayer(task_type=self.task)
 
         # Register regularization weights
         self.register_regularization_weights(
-            embedding_attr='embedding',
-            include_modules=['cross_network', 'mlp', 'final_layer']
+            embedding_attr="embedding",
+            include_modules=["cross_network", "mlp", "final_layer"],
         )
 
         self.compile(
@@ -124,18 +139,20 @@ class DCN(BaseModel):
     def forward(self, x):
         # Get all embeddings and flatten
         input_flat = self.embedding(x=x, features=self.all_features, squeeze_dim=True)
-        
+
         # Cross Network
         cross_output = self.cross_network(input_flat)  # [B, input_dim]
-        
+
         if self.use_dnn:
             # Deep Network
             deep_output = self.mlp(input_flat)  # [B, 1]
             # Concatenate cross and deep
-            combined = torch.cat([cross_output, deep_output], dim=-1)  # [B, input_dim + 1]
+            combined = torch.cat(
+                [cross_output, deep_output], dim=-1
+            )  # [B, input_dim + 1]
         else:
             combined = cross_output
-        
+
         # Final prediction
         y = self.final_layer(combined)
         return self.prediction_layer(y)