nextrec 0.4.1__py3-none-any.whl → 0.4.3__py3-none-any.whl

nextrec/models/multi_task/poso.py

@@ -46,7 +46,8 @@ from nextrec.basic.features import DenseFeature, SequenceFeature, SparseFeature
 from nextrec.basic.layers import EmbeddingLayer, MLP, PredictionLayer
 from nextrec.basic.activation import activation_layer
 from nextrec.basic.model import BaseModel
-from nextrec.utils.model import merge_features
+
+from nextrec.utils.model import select_features
 
 
 class POSOGate(nn.Module):
@@ -79,7 +80,7 @@ class POSOGate(nn.Module):
         h = self.act(self.fc1(pc))
         g = torch.sigmoid(self.fc2(h))  # (B, out_dim) in (0,1)
         return self.scale_factor * g
-    
+
 
 class POSOFC(nn.Module):
     """
@@ -116,10 +117,10 @@ class POSOFC(nn.Module):
         pc: (B, pc_dim)
         return: (B, out_dim)
         """
-        h = self.act(self.linear(x))          # Standard FC with activation
-        g = self.gate(pc)                     # (B, out_dim)
-        return g * h                          # Element-wise gating
-    
+        h = self.act(self.linear(x))  # Standard FC with activation
+        g = self.gate(pc)  # (B, out_dim)
+        return g * h  # Element-wise gating
+
 
 class POSOMLP(nn.Module):
     """
@@ -173,7 +174,7 @@ class POSOMLP(nn.Module):
             if self.dropout is not None:
                 h = self.dropout(h)
         return h
-    
+
 
 class POSOMMoE(nn.Module):
     """
@@ -183,7 +184,7 @@ class POSOMMoE(nn.Module):
         - Task gates aggregate the PC-masked expert outputs
 
     Concretely:
-        h_e = expert_e(x)                 # (B, D) 
+        h_e = expert_e(x)                 # (B, D)
         g_e = POSOGate(pc) in (0, C)^{D}  # (B, D)
         h_e_tilde = g_e ⊙ h_e            # (B, D)
         z_t = Σ_e gate_t,e(x) * h_e_tilde
@@ -192,14 +193,14 @@ class POSOMMoE(nn.Module):
     def __init__(
         self,
         input_dim: int,
-        pc_dim: int,                   # for poso feature dimension
+        pc_dim: int,  # for poso feature dimension
         num_experts: int,
         expert_hidden_dims: list[int],
         num_tasks: int,
         activation: str = "relu",
         expert_dropout: float = 0.0,
-        gate_hidden_dim: int = 32,     # for poso gate hidden dimension
-        scale_factor: float = 2.0,     # for poso gate scale factor
+        gate_hidden_dim: int = 32,  # for poso gate hidden dimension
+        scale_factor: float = 2.0,  # for poso gate scale factor
         gate_use_softmax: bool = True,
     ) -> None:
         super().__init__()
@@ -207,15 +208,41 @@ class POSOMMoE(nn.Module):
         self.num_tasks = num_tasks
 
         # Experts built with framework MLP, same as standard MMoE
-        self.experts = nn.ModuleList([MLP(input_dim=input_dim, output_layer=False, dims=expert_hidden_dims, activation=activation, dropout=expert_dropout,) for _ in range(num_experts)])
-        self.expert_output_dim = expert_hidden_dims[-1] if expert_hidden_dims else input_dim
+        self.experts = nn.ModuleList(
+            [
+                MLP(
+                    input_dim=input_dim,
+                    output_layer=False,
+                    dims=expert_hidden_dims,
+                    activation=activation,
+                    dropout=expert_dropout,
+                )
+                for _ in range(num_experts)
+            ]
+        )
+        self.expert_output_dim = (
+            expert_hidden_dims[-1] if expert_hidden_dims else input_dim
+        )
 
         # Task-specific gates: gate_t(x) over experts
-        self.gates = nn.ModuleList([nn.Linear(input_dim, num_experts) for _ in range(num_tasks)])
+        self.gates = nn.ModuleList(
+            [nn.Linear(input_dim, num_experts) for _ in range(num_tasks)]
+        )
         self.gate_use_softmax = gate_use_softmax
 
         # PC gate per expert: g_e(pc) ∈ R^D
-        self.expert_pc_gates = nn.ModuleList([POSOGate(pc_dim=pc_dim, out_dim=self.expert_output_dim, hidden_dim=gate_hidden_dim, scale_factor=scale_factor, activation=activation,) for _ in range(num_experts)])
+        self.expert_pc_gates = nn.ModuleList(
+            [
+                POSOGate(
+                    pc_dim=pc_dim,
+                    out_dim=self.expert_output_dim,
+                    hidden_dim=gate_hidden_dim,
+                    scale_factor=scale_factor,
+                    activation=activation,
+                )
+                for _ in range(num_experts)
+            ]
+        )
 
     def forward(self, x: torch.Tensor, pc: torch.Tensor) -> list[torch.Tensor]:
         """
@@ -226,9 +253,9 @@ class POSOMMoE(nn.Module):
         # 1) Expert outputs with POSO PC gate
         masked_expert_outputs = []
         for e, expert in enumerate(self.experts):
-            h_e = expert(x)                         # (B, D)
-            g_e = self.expert_pc_gates[e](pc)       # (B, D)
-            h_e_tilde = g_e * h_e                   # (B, D)
+            h_e = expert(x)  # (B, D)
+            g_e = self.expert_pc_gates[e](pc)  # (B, D)
+            h_e_tilde = g_e * h_e  # (B, D)
             masked_expert_outputs.append(h_e_tilde)
 
         masked_expert_outputs = torch.stack(masked_expert_outputs, dim=1)  # (B, E, D)
@@ -236,13 +263,13 @@ class POSOMMoE(nn.Module):
         # 2) Task gates depend on x as in standard MMoE
         task_outputs: list[torch.Tensor] = []
         for t in range(self.num_tasks):
-            logits = self.gates[t](x)               # (B, E)
+            logits = self.gates[t](x)  # (B, E)
             if self.gate_use_softmax:
                 gate = F.softmax(logits, dim=1)
             else:
                 gate = logits
 
-            gate = gate.unsqueeze(-1)               # (B, E, 1)
+            gate = gate.unsqueeze(-1)  # (B, E, 1)
             z_t = torch.sum(gate * masked_expert_outputs, dim=1)  # (B, D)
             task_outputs.append(z_t)
 
@@ -269,15 +296,18 @@ class POSO(BaseModel):
 
     def __init__(
         self,
-        main_dense_features: list[DenseFeature] | None,
-        main_sparse_features: list[SparseFeature] | None,
-        main_sequence_features: list[SequenceFeature] | None,
-        pc_dense_features: list[DenseFeature] | None,
-        pc_sparse_features: list[SparseFeature] | None,
-        pc_sequence_features: list[SequenceFeature] | None,
+        dense_features: list[DenseFeature] | None,
+        sparse_features: list[SparseFeature] | None,
+        sequence_features: list[SequenceFeature] | None,
+        main_dense_features: list[str] | None,
+        main_sparse_features: list[str] | None,
+        main_sequence_features: list[str] | None,
+        pc_dense_features: list[str] | None,
+        pc_sparse_features: list[str] | None,
+        pc_sequence_features: list[str] | None,
         tower_params_list: list[dict],
         target: list[str],
-        task: str | list[str] | None = None,
+        task: str | list[str] = "binary",
         architecture: str = "mlp",
         # POSO gating defaults
         gate_hidden_dim: int = 32,
@@ -303,28 +333,32 @@ class POSO(BaseModel):
         dense_l2_reg: float = 1e-4,
         **kwargs,
     ):
-        # Keep explicit copies of main and PC features
-        self.main_dense_features = list(main_dense_features or [])
-        self.main_sparse_features = list(main_sparse_features or [])
-        self.main_sequence_features = list(main_sequence_features or [])
-        self.pc_dense_features = list(pc_dense_features or [])
-        self.pc_sparse_features = list(pc_sparse_features or [])
-        self.pc_sequence_features = list(pc_sequence_features or [])
         self.num_tasks = len(target)
 
-        if not self.pc_dense_features and not self.pc_sparse_features and not self.pc_sequence_features:
-            raise ValueError("POSO requires at least one PC feature for personalization.")
+        # Normalize task to match num_tasks
+        resolved_task = task
+        if resolved_task is None:
+            resolved_task = self.default_task
+        elif isinstance(resolved_task, str):
+            resolved_task = [resolved_task] * self.num_tasks
+        elif len(resolved_task) == 1 and self.num_tasks > 1:
+            resolved_task = resolved_task * self.num_tasks
+        elif len(resolved_task) != self.num_tasks:
+            raise ValueError(
+                f"Length of task ({len(resolved_task)}) must match number of targets ({self.num_tasks})."
+            )
 
-        dense_features = merge_features(self.main_dense_features, self.pc_dense_features)
-        sparse_features = merge_features(self.main_sparse_features, self.pc_sparse_features)
-        sequence_features = merge_features(self.main_sequence_features, self.pc_sequence_features)
+        if len(tower_params_list) != self.num_tasks:
+            raise ValueError(
+                f"Number of tower params ({len(tower_params_list)}) must match number of tasks ({self.num_tasks})"
+            )
 
         super().__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
             sequence_features=sequence_features,
             target=target,
-            task=task or self.default_task,
+            task=resolved_task,
             device=device,
             embedding_l1_reg=embedding_l1_reg,
             dense_l1_reg=dense_l1_reg,
@@ -333,15 +367,58 @@ class POSO(BaseModel):
             **kwargs,
         )
 
-        self.loss = loss if loss is not None else "bce"
+        self.main_dense_feature_names = list(main_dense_features or [])
+        self.main_sparse_feature_names = list(main_sparse_features or [])
+        self.main_sequence_feature_names = list(main_sequence_features or [])
+        self.pc_dense_feature_names = list(pc_dense_features or [])
+        self.pc_sparse_feature_names = list(pc_sparse_features or [])
+        self.pc_sequence_feature_names = list(pc_sequence_features or [])
+
+        if loss is None:
+            self.loss = "bce"
+        self.loss = loss
+
         optimizer_params = optimizer_params or {}
 
-        self.num_tasks = len(target)
-        if len(tower_params_list) != self.num_tasks:
-            raise ValueError(f"Number of tower params ({len(tower_params_list)}) must match number of tasks ({self.num_tasks})")
+        self.main_dense_features = select_features(
+            self.dense_features, self.main_dense_feature_names, "main_dense_features"
+        )
+        self.main_sparse_features = select_features(
+            self.sparse_features, self.main_sparse_feature_names, "main_sparse_features"
+        )
+        self.main_sequence_features = select_features(
+            self.sequence_features,
+            self.main_sequence_feature_names,
+            "main_sequence_features",
+        )
+
+        self.pc_dense_features = select_features(
+            self.dense_features, self.pc_dense_feature_names, "pc_dense_features"
+        )
+        self.pc_sparse_features = select_features(
+            self.sparse_features, self.pc_sparse_feature_names, "pc_sparse_features"
+        )
+        self.pc_sequence_features = select_features(
+            self.sequence_features,
+            self.pc_sequence_feature_names,
+            "pc_sequence_features",
+        )
 
-        self.main_features = self.main_dense_features + self.main_sparse_features + self.main_sequence_features
-        self.pc_features = self.pc_dense_features + self.pc_sparse_features + self.pc_sequence_features
+        self.main_features = (
+            self.main_dense_features
+            + self.main_sparse_features
+            + self.main_sequence_features
+        )
+        self.pc_features = (
+            self.pc_dense_features + self.pc_sparse_features + self.pc_sequence_features
+        )
+
+        if not self.main_features:
+            raise ValueError("POSO requires at least one main feature.")
+        if not self.pc_features:
+            raise ValueError(
+                "POSO requires at least one PC feature for personalization."
+            )
 
         self.embedding = EmbeddingLayer(features=self.all_features)
         self.main_input_dim = self.embedding.get_input_dim(self.main_features)
@@ -349,7 +426,9 @@ class POSO(BaseModel):
 
         self.architecture = architecture.lower()
         if self.architecture not in {"mlp", "mmoe"}:
-            raise ValueError(f"Unsupported architecture '{architecture}', choose from ['mlp', 'mmoe'].")
+            raise ValueError(
+                f"Unsupported architecture '{architecture}', choose from ['mlp', 'mmoe']."
+            )
 
         # Build backbones
         if self.architecture == "mlp":
@@ -358,13 +437,17 @@ class POSO(BaseModel):
             for tower_params in tower_params_list:
                 dims = tower_params.get("dims")
                 if not dims:
-                    raise ValueError("tower_params must include a non-empty 'dims' list for POSO-MLP towers.")
+                    raise ValueError(
+                        "tower_params must include a non-empty 'dims' list for POSO-MLP towers."
+                    )
                 dropout = tower_params.get("dropout", 0.0)
                 tower = POSOMLP(
                     input_dim=self.main_input_dim,
                     pc_dim=self.pc_input_dim,
                     dims=dims,
-                    gate_hidden_dim=tower_params.get("gate_hidden_dim", gate_hidden_dim),
+                    gate_hidden_dim=tower_params.get(
+                        "gate_hidden_dim", gate_hidden_dim
+                    ),
                     scale_factor=tower_params.get("scale_factor", gate_scale_factor),
                     activation=tower_params.get("activation", gate_activation),
                     use_bias=tower_params.get("use_bias", gate_use_bias),
@@ -375,7 +458,9 @@ class POSO(BaseModel):
                 self.tower_heads.append(nn.Linear(tower_output_dim, 1))
         else:
             if expert_hidden_dims is None or not expert_hidden_dims:
-                raise ValueError("expert_hidden_dims must be provided for MMoE architecture.")
+                raise ValueError(
+                    "expert_hidden_dims must be provided for MMoE architecture."
+                )
             self.mmoe = POSOMMoE(
                 input_dim=self.main_input_dim,
                 pc_dim=self.pc_input_dim,
@@ -388,12 +473,35 @@ class POSO(BaseModel):
                 scale_factor=expert_gate_scale_factor,
                 gate_use_softmax=gate_use_softmax,
             )
-            self.towers = nn.ModuleList([MLP(input_dim=self.mmoe.expert_output_dim, output_layer=True, **tower_params,) for tower_params in tower_params_list])
+            self.towers = nn.ModuleList(
+                [
+                    MLP(
+                        input_dim=self.mmoe.expert_output_dim,
+                        output_layer=True,
+                        **tower_params,
+                    )
+                    for tower_params in tower_params_list
+                ]
+            )
             self.tower_heads = None
-        self.prediction_layer = PredictionLayer(task_type=self.default_task, task_dims=[1] * self.num_tasks,)
-        include_modules = ["towers", "tower_heads"] if self.architecture == "mlp" else ["mmoe", "towers"]
-        self.register_regularization_weights(embedding_attr="embedding", include_modules=include_modules)
-        self.compile(optimizer=optimizer, optimizer_params=optimizer_params, loss=loss, loss_params=loss_params)
+        self.prediction_layer = PredictionLayer(
+            task_type=self.default_task,
+            task_dims=[1] * self.num_tasks,
+        )
+        include_modules = (
+            ["towers", "tower_heads"]
+            if self.architecture == "mlp"
+            else ["mmoe", "towers"]
+        )
+        self.register_regularization_weights(
+            embedding_attr="embedding", include_modules=include_modules
+        )
+        self.compile(
+            optimizer=optimizer,
+            optimizer_params=optimizer_params,
+            loss=loss,
+            loss_params=loss_params,
+        )
 
     def forward(self, x):
         # Embed main and PC features separately so PC can gate hidden units

nextrec/models/multi_task/share_bottom.py

@@ -56,42 +56,58 @@ class ShareBottom(BaseModel):
     def default_task(self):
         num_tasks = getattr(self, "num_tasks", None)
         if num_tasks is not None and num_tasks > 0:
-            return ['binary'] * num_tasks
-        return ['binary']
-    
-    def __init__(self,
-                 dense_features: list[DenseFeature],
-                 sparse_features: list[SparseFeature],
-                 sequence_features: list[SequenceFeature],
-                 bottom_params: dict,
-                 tower_params_list: list[dict],
-                 target: list[str],
-                 task: str | list[str] | None = None,
-                 optimizer: str = "adam",
-                 optimizer_params: dict = {},
-                 loss: str | nn.Module | list[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):
-        
+            return ["binary"] * num_tasks
+        return ["binary"]
+
+    def __init__(
+        self,
+        dense_features: list[DenseFeature],
+        sparse_features: list[SparseFeature],
+        sequence_features: list[SequenceFeature],
+        bottom_params: dict,
+        tower_params_list: list[dict],
+        target: list[str],
+        task: str | list[str] | None = None,
+        optimizer: str = "adam",
+        optimizer_params: dict | None = None,
+        loss: str | nn.Module | list[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,
+    ):
+
+        optimizer_params = optimizer_params or {}
+
         self.num_tasks = len(target)
 
+        resolved_task = task
+        if resolved_task is None:
+            resolved_task = self.default_task
+        elif isinstance(resolved_task, str):
+            resolved_task = [resolved_task] * self.num_tasks
+        elif len(resolved_task) == 1 and self.num_tasks > 1:
+            resolved_task = resolved_task * self.num_tasks
+        elif len(resolved_task) != self.num_tasks:
+            raise ValueError(
+                f"Length of task ({len(resolved_task)}) must match number of targets ({self.num_tasks})."
+            )
+
         super(ShareBottom, self).__init__(
             dense_features=dense_features,
             sparse_features=sparse_features,
             sequence_features=sequence_features,
             target=target,
-            task=task or self.default_task,
+            task=resolved_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,
-            **kwargs
+            **kwargs,
         )
 
         self.loss = loss
@@ -100,7 +116,9 @@ class ShareBottom(BaseModel):
         # Number of tasks
         self.num_tasks = len(target)
         if len(tower_params_list) != self.num_tasks:
-            raise ValueError(f"Number of tower params ({len(tower_params_list)}) must match number of tasks ({self.num_tasks})")
+            raise ValueError(
+                f"Number of tower params ({len(tower_params_list)}) must match number of tasks ({self.num_tasks})"
+            )
         # Embedding layer
         self.embedding = EmbeddingLayer(features=self.all_features)
         # Calculate input dimension
@@ -108,39 +126,48 @@ class ShareBottom(BaseModel):
         # 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
-        
+
         # Shared bottom network
         self.bottom = MLP(input_dim=input_dim, output_layer=False, **bottom_params)
-        
+
         # Get bottom output dimension
-        if 'dims' in bottom_params and len(bottom_params['dims']) > 0:
-            bottom_output_dim = bottom_params['dims'][-1]
+        if "dims" in bottom_params and len(bottom_params["dims"]) > 0:
+            bottom_output_dim = bottom_params["dims"][-1]
         else:
             bottom_output_dim = input_dim
-        
+
         # Task-specific towers
         self.towers = nn.ModuleList()
         for tower_params in tower_params_list:
             tower = MLP(input_dim=bottom_output_dim, output_layer=True, **tower_params)
             self.towers.append(tower)
-        self.prediction_layer = PredictionLayer(task_type=self.default_task, task_dims=[1] * self.num_tasks)
+        self.prediction_layer = PredictionLayer(
+            task_type=self.default_task, task_dims=[1] * self.num_tasks
+        )
         # Register regularization weights
-        self.register_regularization_weights(embedding_attr='embedding', include_modules=['bottom', 'towers'])
-        self.compile(optimizer=optimizer, optimizer_params=optimizer_params, loss=loss, loss_params=loss_params)
+        self.register_regularization_weights(
+            embedding_attr="embedding", include_modules=["bottom", "towers"]
+        )
+        self.compile(
+            optimizer=optimizer,
+            optimizer_params=optimizer_params,
+            loss=loss,
+            loss_params=loss_params,
+        )
 
     def forward(self, x):
         # Get all embeddings and flatten
         input_flat = self.embedding(x=x, features=self.all_features, squeeze_dim=True)
-        
+
         # Shared bottom
         bottom_output = self.bottom(input_flat)  # [B, bottom_dim]
-        
+
         # Task-specific towers
         task_outputs = []
         for tower in self.towers:
             tower_output = tower(bottom_output)  # [B, 1]
             task_outputs.append(tower_output)
-        
+
         # Stack outputs: [B, num_tasks]
         y = torch.cat(task_outputs, dim=1)
         return self.prediction_layer(y)