支持内置DCN-V2并补充DIN的Dice激活示例

docs/source/models/dcn.md

@@ -159,10 +159,42 @@ model_config {
 - embedding_regularization: 对embedding部分加regularization, 减少overfit
 
 ### 示例Config
+- 现在除了通过 backbone 组件化方式外，也支持在内置 `DCN` 模型中开启 v2 的矩阵/低秩 Cross：
+
+```protobuf
+model_config: {
+  model_class: 'DCN'
+  feature_groups: { group_name: 'all' ... }
+  dcn {
+    deep_tower { input: "all" dnn { hidden_units: [256, 128, 64] } }
+    cross_tower {
+      input: "all"
+      cross_num: 3
+      version: 2                # 开启 DCN-V2
+      projection_dim: 64        # 低秩 U*V，省略则使用 full-rank
+      diag_scale: 0.1           # 训练稳定性增强
+      preactivation: "relu"     # 与 Cross 层一致的预激活
+      use_bias: true
+    }
+    final_dnn { hidden_units: [64, 32, 16] }
+    l2_regularization: 1e-4
+  }
+  embedding_regularization: 1e-4
+}
+```
+
 
 1. DCN V1: [DCN_demo.config](https://easyrec.oss-cn-beijing.aliyuncs.com/config/dcn.config)
 1. DCN V2: [dcn_backbone_on_movielens.config](https://github.com/alibaba/EasyRec/tree/master/examples/configs/dcn_backbone_on_movielens.config)
 
+1. DCN V2（内置模型写法）：[dcn_v2_on_taobao.config](file:///mlx_devbox/users/zhaohao.9426/playground/EasyRec/samples/model_config/dcn_v2_on_taobao.config)
+
+本地直接训练命令：
+
+```bash
+python -m easy_rec.python.train_eval --pipeline_config_path samples/model_config/dcn_v2_on_taobao.config
+```
+
 ### 参考论文
 
 1. [DCN v1](https://arxiv.org/abs/1708.05123)

easy_rec/python/model/dcn.py

@@ -2,9 +2,12 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
 import tensorflow as tf
+import os
 
 from easy_rec.python.layers import dnn
 from easy_rec.python.model.rank_model import RankModel
+from easy_rec.python.layers.keras.interaction import Cross
+from easy_rec.python.layers.utils import Parameter
 
 from easy_rec.python.protos.dcn_pb2 import DCN as DCNConfig  # NOQA
 
@@ -44,6 +47,45 @@ def _cross_net(self, tensor, num_cross_layers):
       x = tf.math.add(tf.math.add(x0 * xw, b), x)
     return x
 
+  def _cross_net_v2(self, tensor, num_cross_layers):
+    x0 = tensor
+    x = tensor
+    params = {}  # picked from proto, env acts only as fallback
+    cross_tower = self._model_config.cross_tower
+    if getattr(cross_tower, 'projection_dim', 0):
+      params['projection_dim'] = int(cross_tower.projection_dim)
+    preact = getattr(cross_tower, 'preactivation', '')
+    if preact:
+      params['preactivation'] = preact
+    diag = getattr(cross_tower, 'diag_scale', 0.0)
+    if diag:
+      params['diag_scale'] = float(diag)
+    if hasattr(cross_tower, 'use_bias'):
+      params['use_bias'] = bool(cross_tower.use_bias)
+    # env fallback
+    proj_env = os.getenv('EASYREC_DCN_V2_PROJ', None)
+    if 'projection_dim' not in params and proj_env:
+      try:
+        params['projection_dim'] = int(proj_env)
+      except Exception:
+        pass
+    preact_env = os.getenv('EASYREC_DCN_V2_PREACT', None)
+    if 'preactivation' not in params and preact_env:
+      params['preactivation'] = preact_env
+    diag_env = os.getenv('EASYREC_DCN_V2_DIAG', None)
+    if 'diag_scale' not in params and diag_env:
+      try:
+        params['diag_scale'] = float(diag_env)
+      except Exception:
+        pass
+    use_bias_env = os.getenv('EASYREC_DCN_V2_USE_BIAS', None)
+    if 'use_bias' not in params and use_bias_env:
+      params['use_bias'] = use_bias_env.lower() in ('1', 'true', 'yes')
+    layer_params = Parameter(params, True)
+    for i in range(num_cross_layers):
+      x = Cross(layer_params, name='cross_v2_%d' % i)([x0, x])
+    return x
+
   def build_predict_graph(self):
     tower_fea_arr = []
     # deep tower
@@ -56,7 +98,21 @@ def build_predict_graph(self):
     # cross tower
     cross_tower_config = self._model_config.cross_tower
     num_cross_layers = cross_tower_config.cross_num
-    cross_tensor = self._cross_net(self._features, num_cross_layers)
+    use_v2 = getattr(cross_tower_config, 'version', 1) == 2
+    if not use_v2:
+      # auto-enable v2 if v2-specific params are set
+      if getattr(cross_tower_config, 'projection_dim', 0) or \
+         getattr(cross_tower_config, 'preactivation', '') or \
+         getattr(cross_tower_config, 'diag_scale', 0.0) != 0.0:
+        use_v2 = True
+    # env fallback switch
+    if not use_v2:
+      if os.getenv('EASYREC_DCN_V2', '0').lower() in ('1', 'true', 'yes'):
+        use_v2 = True
+    if use_v2:
+      cross_tensor = self._cross_net_v2(self._features, num_cross_layers)
+    else:
+      cross_tensor = self._cross_net(self._features, num_cross_layers)
     tower_fea_arr.append(cross_tensor)
     # final tower
     all_fea = tf.concat(tower_fea_arr, axis=1)

easy_rec/python/protos/dcn.proto

@@ -8,6 +8,17 @@ message CrossTower {
     required string input = 1;
     // The number of cross layers
     required uint32 cross_num = 2 [default = 3];
+    // DCN version: 1 for vector cross, 2 for matrix/low-rank cross
+    optional uint32 version = 3 [default = 1];
+    // Parameters for DCN v2 (matrix/low-rank)
+    // If projection_dim is set, use low-rank U*V; otherwise use full-rank.
+    optional uint32 projection_dim = 4;
+    // Add diag scale to improve stability: W + diag_scale * I
+    optional float diag_scale = 5 [default = 0.0];
+    // Preactivation applied to Dense output before element-wise product with x0
+    optional string preactivation = 6;
+    // Whether to include bias terms in cross layers
+    optional bool use_bias = 7 [default = true];
 };
 
 message DCN {