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

nextrec/models/match/sdm.py

@@ -6,6 +6,7 @@ Reference:
     [1] Ying H, Zhuang F, Zhang F, et al. Sequential recommender system based on hierarchical attention networks[C]
         //IJCAI. 2018: 3926-3932.
 """
+
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -20,46 +21,53 @@ class SDM(BaseMatchModel):
     @property
     def model_name(self) -> str:
         return "SDM"
-    
+
     @property
     def support_training_modes(self) -> list[str]:
-        return ['pointwise']
-    
-    def __init__(self,
-                 user_dense_features: list[DenseFeature] | None = None,
-                 user_sparse_features: list[SparseFeature] | None = None,
-                 user_sequence_features: list[SequenceFeature] | None = None,
-                 item_dense_features: list[DenseFeature] | None = None,
-                 item_sparse_features: list[SparseFeature] | None = None,
-                 item_sequence_features: list[SequenceFeature] | None = None,
-                 embedding_dim: int = 64,
-                 rnn_type: Literal['GRU', 'LSTM'] = 'GRU',
-                 rnn_hidden_size: int = 64,
-                 rnn_num_layers: int = 1,
-                 rnn_dropout: float = 0.0,
-                 use_short_term: bool = True,
-                 use_long_term: bool = True,
-                 item_dnn_hidden_units: list[int] = [256, 128],
-                 dnn_activation: str = 'relu',
-                 dnn_dropout: float = 0.0,
-                 training_mode: Literal['pointwise', 'pairwise', 'listwise'] = 'pointwise',
-                 num_negative_samples: int = 4,
-                 temperature: float = 1.0,
-                 similarity_metric: Literal['dot', 'cosine', 'euclidean'] = 'dot',
-                 device: str = 'cpu',
-                 embedding_l1_reg: float = 0.0,
-                 dense_l1_reg: float = 0.0,
-                 embedding_l2_reg: float = 0.0,
-                 dense_l2_reg: float = 0.0,
-                 early_stop_patience: int = 20,
-                 optimizer: str | torch.optim.Optimizer = "adam",
-                 optimizer_params: dict | None = None,
-                 scheduler: str | torch.optim.lr_scheduler._LRScheduler | type[torch.optim.lr_scheduler._LRScheduler] | None = None,
-                 scheduler_params: dict | None = None,
-                 loss: str | nn.Module | list[str | nn.Module] | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 **kwargs):
-        
+        return ["pointwise"]
+
+    def __init__(
+        self,
+        user_dense_features: list[DenseFeature] | None = None,
+        user_sparse_features: list[SparseFeature] | None = None,
+        user_sequence_features: list[SequenceFeature] | None = None,
+        item_dense_features: list[DenseFeature] | None = None,
+        item_sparse_features: list[SparseFeature] | None = None,
+        item_sequence_features: list[SequenceFeature] | None = None,
+        embedding_dim: int = 64,
+        rnn_type: Literal["GRU", "LSTM"] = "GRU",
+        rnn_hidden_size: int = 64,
+        rnn_num_layers: int = 1,
+        rnn_dropout: float = 0.0,
+        use_short_term: bool = True,
+        use_long_term: bool = True,
+        item_dnn_hidden_units: list[int] = [256, 128],
+        dnn_activation: str = "relu",
+        dnn_dropout: float = 0.0,
+        training_mode: Literal["pointwise", "pairwise", "listwise"] = "pointwise",
+        num_negative_samples: int = 4,
+        temperature: float = 1.0,
+        similarity_metric: Literal["dot", "cosine", "euclidean"] = "dot",
+        device: str = "cpu",
+        embedding_l1_reg: float = 0.0,
+        dense_l1_reg: float = 0.0,
+        embedding_l2_reg: float = 0.0,
+        dense_l2_reg: float = 0.0,
+        early_stop_patience: int = 20,
+        optimizer: str | torch.optim.Optimizer = "adam",
+        optimizer_params: dict | None = None,
+        scheduler: (
+            str
+            | torch.optim.lr_scheduler._LRScheduler
+            | type[torch.optim.lr_scheduler._LRScheduler]
+            | None
+        ) = None,
+        scheduler_params: dict | None = None,
+        loss: str | nn.Module | list[str | nn.Module] | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        **kwargs,
+    ):
+
         super(SDM, self).__init__(
             user_dense_features=user_dense_features,
             user_sparse_features=user_sparse_features,
@@ -76,17 +84,16 @@ class SDM(BaseMatchModel):
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            early_stop_patience=early_stop_patience,
-            **kwargs
+            **kwargs,
         )
-        
+
         self.embedding_dim = embedding_dim
         self.rnn_type = rnn_type
         self.rnn_hidden_size = rnn_hidden_size
         self.use_short_term = use_short_term
         self.use_long_term = use_long_term
         self.item_dnn_hidden_units = item_dnn_hidden_units
-        
+
         # User tower
         user_features = []
         if user_dense_features:
@@ -95,54 +102,54 @@ class SDM(BaseMatchModel):
             user_features.extend(user_sparse_features)
         if user_sequence_features:
             user_features.extend(user_sequence_features)
-        
+
         if len(user_features) > 0:
             self.user_embedding = EmbeddingLayer(user_features)
-            
+
             if not user_sequence_features or len(user_sequence_features) == 0:
                 raise ValueError("SDM requires at least one user sequence feature")
-            
+
             seq_emb_dim = user_sequence_features[0].embedding_dim
-            
-            if rnn_type == 'GRU':
+
+            if rnn_type == "GRU":
                 self.rnn = nn.GRU(
                     input_size=seq_emb_dim,
                     hidden_size=rnn_hidden_size,
                     num_layers=rnn_num_layers,
                     batch_first=True,
-                    dropout=rnn_dropout if rnn_num_layers > 1 else 0.0
+                    dropout=rnn_dropout if rnn_num_layers > 1 else 0.0,
                 )
-            elif rnn_type == 'LSTM':
+            elif rnn_type == "LSTM":
                 self.rnn = nn.LSTM(
                     input_size=seq_emb_dim,
                     hidden_size=rnn_hidden_size,
                     num_layers=rnn_num_layers,
                     batch_first=True,
-                    dropout=rnn_dropout if rnn_num_layers > 1 else 0.0
+                    dropout=rnn_dropout if rnn_num_layers > 1 else 0.0,
                 )
             else:
                 raise ValueError(f"Unknown RNN type: {rnn_type}")
-            
+
             user_final_dim = 0
             if use_long_term:
-                user_final_dim += rnn_hidden_size  
+                user_final_dim += rnn_hidden_size
             if use_short_term:
-                user_final_dim += seq_emb_dim  
-            
+                user_final_dim += seq_emb_dim
+
             for feat in user_dense_features or []:
                 user_final_dim += 1
             for feat in user_sparse_features or []:
                 user_final_dim += feat.embedding_dim
-            
+
             # User DNN to final embedding
             self.user_dnn = MLP(
                 input_dim=user_final_dim,
                 dims=[rnn_hidden_size * 2, embedding_dim],
                 output_layer=False,
                 dropout=dnn_dropout,
-                activation=dnn_activation
+                activation=dnn_activation,
             )
-        
+
         # Item tower
         item_features = []
         if item_dense_features:
@@ -151,10 +158,10 @@ class SDM(BaseMatchModel):
             item_features.extend(item_sparse_features)
         if item_sequence_features:
             item_features.extend(item_sequence_features)
-        
+
         if len(item_features) > 0:
             self.item_embedding = EmbeddingLayer(item_features)
-            
+
             item_input_dim = 0
             for feat in item_dense_features or []:
                 item_input_dim += 1
@@ -162,7 +169,7 @@ class SDM(BaseMatchModel):
                 item_input_dim += feat.embedding_dim
             for feat in item_sequence_features or []:
                 item_input_dim += feat.embedding_dim
-            
+
             # Item DNN
             if len(item_dnn_hidden_units) > 0:
                 item_dnn_units = item_dnn_hidden_units + [embedding_dim]
@@ -171,20 +178,19 @@ class SDM(BaseMatchModel):
                     dims=item_dnn_units,
                     output_layer=False,
                     dropout=dnn_dropout,
-                    activation=dnn_activation
+                    activation=dnn_activation,
                 )
             else:
                 self.item_dnn = None
-        
+
         self.register_regularization_weights(
-            embedding_attr='user_embedding',
-            include_modules=['rnn', 'user_dnn']
+            embedding_attr="user_embedding", include_modules=["rnn", "user_dnn"]
         )
         self.register_regularization_weights(
-            embedding_attr='item_embedding',
-            include_modules=['item_dnn'] if self.item_dnn else []
+            embedding_attr="item_embedding",
+            include_modules=["item_dnn"] if self.item_dnn else [],
         )
-        
+
         self.compile(
             optimizer=optimizer,
             optimizer_params=optimizer_params,
@@ -195,38 +201,44 @@ class SDM(BaseMatchModel):
         )
 
         self.to(device)
-    
+
     def user_tower(self, user_input: dict) -> torch.Tensor:
         seq_feature = self.user_sequence_features[0]
         seq_input = user_input[seq_feature.name]
-        
+
         embed = self.user_embedding.embed_dict[seq_feature.embedding_name]
         seq_emb = embed(seq_input.long())  # [batch_size, seq_len, seq_emb_dim]
-        
-        if self.rnn_type == 'GRU':
-            rnn_output, hidden = self.rnn(seq_emb)  # hidden: [num_layers, batch, hidden_size]
-        elif self.rnn_type == 'LSTM':
+
+        if self.rnn_type == "GRU":
+            rnn_output, hidden = self.rnn(
+                seq_emb
+            )  # hidden: [num_layers, batch, hidden_size]
+        elif self.rnn_type == "LSTM":
             rnn_output, (hidden, cell) = self.rnn(seq_emb)
-        
+
         features_list = []
-        
+
         if self.use_long_term:
             if self.rnn.num_layers > 1:
                 long_term = hidden[-1, :, :]  # [batch_size, hidden_size]
             else:
                 long_term = hidden.squeeze(0)  # [batch_size, hidden_size]
             features_list.append(long_term)
-        
+
         if self.use_short_term:
-            mask = (seq_input != seq_feature.padding_idx).float()  # [batch_size, seq_len]
+            mask = (
+                seq_input != seq_feature.padding_idx
+            ).float()  # [batch_size, seq_len]
             seq_lengths = mask.sum(dim=1).long() - 1  # [batch_size]
             seq_lengths = torch.clamp(seq_lengths, min=0)
-            
+
             batch_size = seq_emb.size(0)
             batch_indices = torch.arange(batch_size, device=seq_emb.device)
-            short_term = seq_emb[batch_indices, seq_lengths, :]  # [batch_size, seq_emb_dim]
+            short_term = seq_emb[
+                batch_indices, seq_lengths, :
+            ]  # [batch_size, seq_emb_dim]
             features_list.append(short_term)
-        
+
         if self.user_dense_features:
             dense_features = []
             for feat in self.user_dense_features:
@@ -237,7 +249,7 @@ class SDM(BaseMatchModel):
                     dense_features.append(val)
             if dense_features:
                 features_list.append(torch.cat(dense_features, dim=1))
-        
+
         if self.user_sparse_features:
             sparse_features = []
             for feat in self.user_sparse_features:
@@ -247,22 +259,26 @@ class SDM(BaseMatchModel):
                     sparse_features.append(sparse_emb)
             if sparse_features:
                 features_list.append(torch.cat(sparse_features, dim=1))
-        
+
         user_features = torch.cat(features_list, dim=1)
         user_emb = self.user_dnn(user_features)
         user_emb = F.normalize(user_emb, p=2, dim=1)
-        
+
         return user_emb
-    
+
     def item_tower(self, item_input: dict) -> torch.Tensor:
         """Item tower"""
-        all_item_features = self.item_dense_features + self.item_sparse_features + self.item_sequence_features
+        all_item_features = (
+            self.item_dense_features
+            + self.item_sparse_features
+            + self.item_sequence_features
+        )
         item_emb = self.item_embedding(item_input, all_item_features, squeeze_dim=True)
-        
+
         if self.item_dnn is not None:
             item_emb = self.item_dnn(item_emb)
-        
+
         # L2 normalization
         item_emb = F.normalize(item_emb, p=2, dim=1)
-        
+
         return item_emb

nextrec/models/match/youtube_dnn.py

@@ -6,13 +6,14 @@ Reference:
     [1] Covington P, Adams J, Sargin E. Deep neural networks for youtube recommendations[C]
         //Proceedings of the 10th ACM conference on recommender systems. 2016: 191-198.
 """
+
 import torch
 import torch.nn as nn
 from typing import Literal
 
 from nextrec.basic.model import BaseMatchModel
 from nextrec.basic.features import DenseFeature, SparseFeature, SequenceFeature
-from nextrec.basic.layers import MLP, EmbeddingLayer, AveragePooling
+from nextrec.basic.layers import MLP, EmbeddingLayer
 
 
 class YoutubeDNN(BaseMatchModel):
@@ -22,41 +23,48 @@ class YoutubeDNN(BaseMatchModel):
     Item tower: item features -> item embedding.
     Training usually uses listwise / sampled softmax style objectives.
     """
-    
+
     @property
     def model_name(self) -> str:
         return "YouTubeDNN"
-    
-    def __init__(self,
-                 user_dense_features: list[DenseFeature] | None = None,
-                 user_sparse_features: list[SparseFeature] | None = None,
-                 user_sequence_features: list[SequenceFeature] | None = None,
-                 item_dense_features: list[DenseFeature] | None = None,
-                 item_sparse_features: list[SparseFeature] | None = None,
-                 item_sequence_features: list[SequenceFeature] | None = None,
-                 user_dnn_hidden_units: list[int] = [256, 128, 64],
-                 item_dnn_hidden_units: list[int] = [256, 128, 64],
-                 embedding_dim: int = 64,
-                 dnn_activation: str = 'relu',
-                 dnn_dropout: float = 0.0,
-                 training_mode: Literal['pointwise', 'pairwise', 'listwise'] = 'listwise',
-                 num_negative_samples: int = 100,
-                 temperature: float = 1.0,
-                 similarity_metric: Literal['dot', 'cosine', 'euclidean'] = 'dot',
-                 device: str = 'cpu',
-                 embedding_l1_reg: float = 0.0,
-                 dense_l1_reg: float = 0.0,
-                 embedding_l2_reg: float = 0.0,
-                 dense_l2_reg: float = 0.0,
-                 early_stop_patience: int = 20,
-                 optimizer: str | torch.optim.Optimizer = "adam",
-                 optimizer_params: dict | None = None,
-                 scheduler: str | torch.optim.lr_scheduler._LRScheduler | type[torch.optim.lr_scheduler._LRScheduler] | None = None,
-                 scheduler_params: dict | None = None,
-                 loss: str | nn.Module | list[str | nn.Module] | None = "bce",
-                 loss_params: dict | list[dict] | None = None,
-                 **kwargs):
-        
+
+    def __init__(
+        self,
+        user_dense_features: list[DenseFeature] | None = None,
+        user_sparse_features: list[SparseFeature] | None = None,
+        user_sequence_features: list[SequenceFeature] | None = None,
+        item_dense_features: list[DenseFeature] | None = None,
+        item_sparse_features: list[SparseFeature] | None = None,
+        item_sequence_features: list[SequenceFeature] | None = None,
+        user_dnn_hidden_units: list[int] = [256, 128, 64],
+        item_dnn_hidden_units: list[int] = [256, 128, 64],
+        embedding_dim: int = 64,
+        dnn_activation: str = "relu",
+        dnn_dropout: float = 0.0,
+        training_mode: Literal["pointwise", "pairwise", "listwise"] = "listwise",
+        num_negative_samples: int = 100,
+        temperature: float = 1.0,
+        similarity_metric: Literal["dot", "cosine", "euclidean"] = "dot",
+        device: str = "cpu",
+        embedding_l1_reg: float = 0.0,
+        dense_l1_reg: float = 0.0,
+        embedding_l2_reg: float = 0.0,
+        dense_l2_reg: float = 0.0,
+        early_stop_patience: int = 20,
+        optimizer: str | torch.optim.Optimizer = "adam",
+        optimizer_params: dict | None = None,
+        scheduler: (
+            str
+            | torch.optim.lr_scheduler._LRScheduler
+            | type[torch.optim.lr_scheduler._LRScheduler]
+            | None
+        ) = None,
+        scheduler_params: dict | None = None,
+        loss: str | nn.Module | list[str | nn.Module] | None = "bce",
+        loss_params: dict | list[dict] | None = None,
+        **kwargs,
+    ):
+
         super(YoutubeDNN, self).__init__(
             user_dense_features=user_dense_features,
             user_sparse_features=user_sparse_features,
@@ -73,14 +81,13 @@ class YoutubeDNN(BaseMatchModel):
             dense_l1_reg=dense_l1_reg,
             embedding_l2_reg=embedding_l2_reg,
             dense_l2_reg=dense_l2_reg,
-            early_stop_patience=early_stop_patience,
-            **kwargs
+            **kwargs,
         )
-        
+
         self.embedding_dim = embedding_dim
         self.user_dnn_hidden_units = user_dnn_hidden_units
         self.item_dnn_hidden_units = item_dnn_hidden_units
-        
+
         # User tower
         user_features = []
         if user_dense_features:
@@ -89,10 +96,10 @@ class YoutubeDNN(BaseMatchModel):
             user_features.extend(user_sparse_features)
         if user_sequence_features:
             user_features.extend(user_sequence_features)
-        
+
         if len(user_features) > 0:
             self.user_embedding = EmbeddingLayer(user_features)
-            
+
             user_input_dim = 0
             for feat in user_dense_features or []:
                 user_input_dim += 1
@@ -101,16 +108,16 @@ class YoutubeDNN(BaseMatchModel):
             for feat in user_sequence_features or []:
                 # Sequence features are pooled before entering the DNN
                 user_input_dim += feat.embedding_dim
-            
+
             user_dnn_units = user_dnn_hidden_units + [embedding_dim]
             self.user_dnn = MLP(
                 input_dim=user_input_dim,
                 dims=user_dnn_units,
                 output_layer=False,
                 dropout=dnn_dropout,
-                activation=dnn_activation
+                activation=dnn_activation,
             )
-        
+
         # Item tower
         item_features = []
         if item_dense_features:
@@ -119,10 +126,10 @@ class YoutubeDNN(BaseMatchModel):
             item_features.extend(item_sparse_features)
         if item_sequence_features:
             item_features.extend(item_sequence_features)
-        
+
         if len(item_features) > 0:
             self.item_embedding = EmbeddingLayer(item_features)
-            
+
             item_input_dim = 0
             for feat in item_dense_features or []:
                 item_input_dim += 1
@@ -130,25 +137,23 @@ class YoutubeDNN(BaseMatchModel):
                 item_input_dim += feat.embedding_dim
             for feat in item_sequence_features or []:
                 item_input_dim += feat.embedding_dim
-            
+
             item_dnn_units = item_dnn_hidden_units + [embedding_dim]
             self.item_dnn = MLP(
                 input_dim=item_input_dim,
                 dims=item_dnn_units,
                 output_layer=False,
                 dropout=dnn_dropout,
-                activation=dnn_activation
+                activation=dnn_activation,
             )
-        
+
         self.register_regularization_weights(
-            embedding_attr='user_embedding',
-            include_modules=['user_dnn']
+            embedding_attr="user_embedding", include_modules=["user_dnn"]
         )
         self.register_regularization_weights(
-            embedding_attr='item_embedding',
-            include_modules=['item_dnn']
+            embedding_attr="item_embedding", include_modules=["item_dnn"]
         )
-        
+
         self.compile(
             optimizer=optimizer,
             optimizer_params=optimizer_params,
@@ -159,27 +164,35 @@ class YoutubeDNN(BaseMatchModel):
         )
 
         self.to(device)
-    
+
     def user_tower(self, user_input: dict) -> torch.Tensor:
         """
         User tower to encode historical behavior sequences and user features.
         """
-        all_user_features = self.user_dense_features + self.user_sparse_features + self.user_sequence_features
+        all_user_features = (
+            self.user_dense_features
+            + self.user_sparse_features
+            + self.user_sequence_features
+        )
         user_emb = self.user_embedding(user_input, all_user_features, squeeze_dim=True)
         user_emb = self.user_dnn(user_emb)
-        
+
         # L2 normalization
         user_emb = torch.nn.functional.normalize(user_emb, p=2, dim=1)
-        
+
         return user_emb
-    
+
     def item_tower(self, item_input: dict) -> torch.Tensor:
         """Item tower"""
-        all_item_features = self.item_dense_features + self.item_sparse_features + self.item_sequence_features
+        all_item_features = (
+            self.item_dense_features
+            + self.item_sparse_features
+            + self.item_sequence_features
+        )
         item_emb = self.item_embedding(item_input, all_item_features, squeeze_dim=True)
         item_emb = self.item_dnn(item_emb)
-        
+
         # L2 normalization
         item_emb = torch.nn.functional.normalize(item_emb, p=2, dim=1)
-        
+
         return item_emb