Add dcnv2 and xdeepfm net

docs/source/models/dcn.md

@@ -35,13 +35,13 @@ model_config {
         group_type: DEEP
     }
     dcn_v1 {
-        cross_tower {
+        cross {
             cross_num: 3
         }
-        deep_tower {
+        deep {
             hidden_units: [256, 128]
         }
-        final_dnn {
+        final {
             hidden_units: [64, 32]
         }
     }
@@ -57,11 +57,11 @@ model_config {
 }
 ```
 
-- cross_tower
+- cross
   - cross_num: 交叉层层数，默认为3
-- deep_tower
+- deep
   - hidden_units: dnn每一层的channel数目，即神经元的数目
-- final_dnn: 整合cross_tower, deep_tower的全连接层
+- final: 整合cross层, deep层的全连接层
 
 ## 参考论文
 

docs/source/models/dcn_v2.md

@@ -0,0 +1,83 @@
+# DCN V2
+
+## 简介
+
+### DCN v2
+
+相对于DCN v1模型，主要的改进点在于：
+
+1. Wide侧-Cross Network中用矩阵替代向量,方阵可以分解成2个低维矩阵;
+1. 提出2种模型结构，传统的Wide&Deep并行 + Wide&Deep串行。
+   ![dcn_v2](../../images/models/dcn_v2.jpg)
+   ![dcn_v2_cross](../../images/models/dcn_v2_cross.jpg)
+
+## 配置说明
+
+### DCNV2
+
+```
+model_config {
+    feature_groups {
+        group_name: "features"
+        feature_names: "user_id"
+        feature_names: "cms_segid"
+        feature_names: "cms_group_id"
+        feature_names: "final_gender_code"
+        feature_names: "age_level"
+        feature_names: "pvalue_level"
+        feature_names: "shopping_level"
+        feature_names: "occupation"
+        feature_names: "new_user_class_level"
+        feature_names: "pid"
+        feature_names: "adgroup_id"
+        feature_names: "cate_id"
+        feature_names: "campaign_id"
+        feature_names: "customer"
+        feature_names: "brand"
+        feature_names: "price"
+        group_type: DEEP
+    }
+    dcn_v2 {
+        backbone {
+            hidden_units: 512
+            hidden_units: 256
+            hidden_units: 128
+        }
+        cross {
+            cross_num: 2
+            low_rank: 32
+        }
+        deep {
+            hidden_units: 512
+            hidden_units: 256
+        }
+        final {
+            hidden_units: 128
+            hidden_units: 32
+        }
+    }
+    num_class: 1
+    metrics {
+        auc {}
+    }
+    losses {
+        binary_cross_entropy {}
+    }
+}
+```
+
+- backbone: dnn层,可选配置,特征在进入cross层的时候是否要经过dnn层的处理
+- cross
+  - cross_num: 交叉层层数，默认为3
+  - low_rank: cross层中大矩阵分解成2个低维矩阵的维度
+- deep
+  - hidden_units: dnn每一层的channel数目，即神经元的数目
+- final: 整合cross层, deep层的全连接层
+
+### 示例Config
+
+[dcn_v2_demo.config](https://tzrec.oss-cn-beijing.aliyuncs.com/config/models/dcn_v2_criteo.config)
+
+### 参考论文
+
+[DCN v2](https://arxiv.org/abs/2008.13535)

docs/source/models/xdeepfm.md

@@ -0,0 +1,90 @@
+# xDeepFM
+
+### 简介
+
+xDeepFM模型延续了deep&cross network(参考[DCN](https://easyrec.readthedocs.io/en/latest/models/dcn.html))模型的建模思想。不过，在建模显式高阶交叉特征时，采用了不同于deep&cross network的方式，文章称为compressed interaction network(CIN)，并将CIN网络与深度神经网络结合，最后输入到输出层。
+
+![xdeepfm.png](../../images/models/xdeepfm.png)
+
+### 配置说明
+
+组件化模型
+
+```protobuf
+
+model_config {
+    feature_groups {
+        group_name: "wide"
+        feature_names: 'user_id'
+        feature_names: 'cms_segid'
+        feature_names: 'cms_group_id'
+        feature_names: 'age_level'
+        feature_names: 'pvalue_level'
+        feature_names: 'shopping_level'
+        feature_names: 'occupation'
+        feature_names: 'new_user_class_level'
+        feature_names: 'tag_category_list'
+        feature_names: 'tag_brand_list'
+        feature_names: 'adgroup_id'
+        feature_names: 'cate_id'
+        feature_names: 'campaign_id'
+        feature_names: 'customer'
+        feature_names: 'brand'
+        feature_names: 'price'
+        feature_names: 'pid'
+        group_type: WIDE
+    }
+    feature_groups {
+        group_name: "deep"
+        feature_names: 'user_id'
+        feature_names: 'cms_segid'
+        feature_names: 'cms_group_id'
+        feature_names: 'age_level'
+        feature_names: 'pvalue_level'
+        feature_names: 'shopping_level'
+        feature_names: 'occupation'
+        feature_names: 'new_user_class_level'
+        feature_names: 'tag_category_list'
+        feature_names: 'tag_brand_list'
+        feature_names: 'adgroup_id'
+        feature_names: 'cate_id'
+        feature_names: 'campaign_id'
+        feature_names: 'customer'
+        feature_names: 'brand'
+        feature_names: 'price'
+        feature_names: 'pid'
+        group_type: DEEP
+    }
+    xdeepfm {
+        cin {
+            cin_layer_size: [64, 64]
+        }
+        deep {
+            hidden_units: [128, 64]
+        }
+        final {
+            hidden_units: [64, 32]
+        }
+    }
+    metrics {
+        auc {}
+    }
+    losses {
+        binary_cross_entropy {}
+    }
+}
+```
+
+- cin: 特征交叉层
+  - cin_layer_size: cin每层要输出的维度
+- deep
+  - hidden_units: dnn每一层的channel数目，即神经元的数目
+- final: 整合cross层, deep层的全连接层
+
+### 示例Config
+
+[xdeepfm_criteo.config](https://tzrec.oss-cn-beijing.aliyuncs.com/config/models/xdeepfm_criteo.config)
+
+### 参考论文
+
+[xDeepFM](https://arxiv.org/pdf/1803.05170)

tzrec/models/dcn.py

@@ -46,19 +46,18 @@ def __init__(
         self.group_name = self.embedding_group.group_names()[0]
         feature_dim = self.embedding_group.group_total_dim(self.group_name)
         self.cross = Cross(
-            input_dim=feature_dim, **config_to_kwargs(self._model_config.cross_tower)
+            input_dim=feature_dim, **config_to_kwargs(self._model_config.cross)
         )
         self.deep = MLP(
-            in_features=feature_dim, **config_to_kwargs(self._model_config.deep_tower)
+            in_features=feature_dim, **config_to_kwargs(self._model_config.deep)
         )
-        deep_output_dim = self._model_config.deep_tower.hidden_units[-1]
-        final_dnn_input_dim = feature_dim + deep_output_dim
+        final_dnn_input_dim = self.cross.output_dim() + self.deep.output_dim()
         self.final_dnn = MLP(
             in_features=final_dnn_input_dim,
-            **config_to_kwargs(self._model_config.final_dnn),
+            **config_to_kwargs(self._model_config.final),
         )
         self.output_linear = nn.Linear(
-            self._model_config.final_dnn.hidden_units[-1], self._num_class, bias=False
+            self.final_dnn.output_dim(), self._num_class, bias=False
         )
 
     def predict(self, batch: Batch) -> Dict[str, torch.Tensor]:

tzrec/models/dcn_test.py

@@ -56,9 +56,9 @@ def test_DCNV1(self, graph_type) -> None:
         model_config = model_pb2.ModelConfig(
             feature_groups=feature_groups,
             dcn_v1=rank_model_pb2.DCNV1(
-                cross_tower=module_pb2.Cross(cross_num=3),
-                deep_tower=module_pb2.MLP(hidden_units=[8, 4]),
-                final_dnn=module_pb2.MLP(hidden_units=[2]),
+                cross=module_pb2.Cross(cross_num=3),
+                deep=module_pb2.MLP(hidden_units=[8, 4]),
+                final=module_pb2.MLP(hidden_units=[2]),
             ),
             losses=[
                 loss_pb2.LossConfig(binary_cross_entropy=loss_pb2.BinaryCrossEntropy())

tzrec/models/dcn_v2.py

@@ -0,0 +1,88 @@
+# Copyright (c) 2025, Alibaba Group;
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#    http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, List, Optional
+
+import torch
+from torch import nn
+
+from tzrec.datasets.utils import Batch
+from tzrec.features.feature import BaseFeature
+from tzrec.models.rank_model import RankModel
+from tzrec.modules.interaction import CrossV2
+from tzrec.modules.mlp import MLP
+from tzrec.protos.model_pb2 import ModelConfig
+from tzrec.utils.config_util import config_to_kwargs
+
+
+class DCNV2(RankModel):
+    """Deep cross network v2.
+
+    Args:
+        model_config (ModelConfig): an instance of ModelConfig.
+        features (list): list of features.
+        labels (list): list of label names.
+        sample_weights (list): sample weight names.
+    """
+
+    def __init__(
+        self,
+        model_config: ModelConfig,
+        features: List[BaseFeature],
+        labels: List[str],
+        sample_weights: Optional[List[str]] = None,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(model_config, features, labels, sample_weights, **kwargs)
+        self.init_input()
+        self.group_name = self.embedding_group.group_names()[0]
+        feature_dim = self.embedding_group.group_total_dim(self.group_name)
+
+        self.backbone = None
+        if self._model_config.HasField("backbone"):
+            self.backbone = MLP(
+                in_features=feature_dim, **config_to_kwargs(self._model_config.backbone)
+            )
+            feature_dim = self.backbone.output_dim()
+
+        self.cross = CrossV2(
+            input_dim=feature_dim, **config_to_kwargs(self._model_config.cross)
+        )
+        final_input_dim = self.cross.output_dim()
+        self.deep = None
+        if self._model_config.HasField("deep"):
+            in_features = self.embedding_group.group_total_dim(self.group_name)
+            self.deep = MLP(
+                in_features=in_features, **config_to_kwargs(self._model_config.deep)
+            )
+            final_input_dim += self.deep.output_dim()
+        self.final = MLP(
+            in_features=final_input_dim,
+            **config_to_kwargs(self._model_config.final),
+        )
+        self.output_mlp = nn.Linear(
+            self.final.output_dim(), self._num_class, bias=False
+        )
+
+    def predict(self, batch: Batch) -> Dict[str, torch.Tensor]:
+        """Forward method."""
+        feature_dict = self.build_input(batch)
+        features = feature_dict[self.group_name]
+        if self.backbone:
+            net = self.backbone(features)
+        else:
+            net = features
+        net = self.cross(net)
+        if self.deep:
+            deep_net = self.deep(features)
+            net = torch.concat([net, deep_net], dim=-1)
+        out = self.output_mlp(self.final(net))
+        return self._output_to_prediction(out)