deep_cross_layer

关注

原创最新推荐文章于 2026-03-06 01:17:48 发布 167 阅读 AI 写同款· GEO 优化 ›

Deep Cross中的cross Layer详解(数学公式+代码+实现细节) 文章目录简介DCN中的Cross Layer操作流程数学原理Torch实现简介再推荐模型中，对于特征feat interaction的探索占据了相当一部分的工作。这也说明对于CTR问题来说，feature Interaction的探索是判断最终用户click的重要一环。除了FM和FFM等因子分解机模型之外，还有Cross Layer的出现也是非常具有代表性的特征交互模型。这里从DCN模型入手，然后单独将cross layer抽离出来，然后对于他的数学原理还有代码实现细节都进行比较深入的探索。 DC 阅读详情

在这里插入图片描述

#coding=utf-8
import tensorflow as tf
from tensorflow.keras.layers import Dense, ReLU, Layer

class DeepCrossLater(Layer):
    def __init__(self,dim_stack, name=None):
        """
        :param hidden_unit: A list. Neural network hidden units.
        :param dim_stack: A scalar. The dimension of inputs unit.
        """
        super(DeepCrossLater, self).__init__()
        # self.layer1 = Dense(units=hidden_unit, activation='relu')
        # self.layer2 = Dense(units=dim_stack, activation=None)
        self.relu = ReLU()
        self.cross_num = 3
        self.cross_weight = []
        self.bais_weight = []


        self.onehot_embedding = self.add_weight(shape=(10, 5),
                                                initializer=tf.initializers.glorot_normal())


        for i in range(3):
            self.cross_weight.append(
                self.add_weight(shape=(dim_stack, 1),
                                initializer=tf.initializers.glorot_normal(),
                                name="{}/cross_net_{}_weight".format(name,i)))


        for i in range(3):
            self.cross_weight.append(
                self.bais_weight.append(self.add_weight(shape=(dim_stack, 1),
                                                  initializer=tf.initializers.glorot_normal(),
                                                  name="{}/cross_net_{}_bias".format(name, i))))

    def call(self, inputs, **kwargs):

        x0 = tf.nn.embedding_lookup(self.onehot_embedding, inputs)

        ######################## 实现方法一
        xl = x0
        for i in range(3):
            xl = tf.matmul(x0, xl, transpose_a=True)  ## batch_size, (5, 1)*(1, 5)*(5, 1)
            xl = tf.matmul(xl, self.cross_weight[i])
            xl = xl + self.bais_weight[i] + tf.transpose(x0, [0, 2, 1])
            xl = tf.transpose(xl, [0, 2, 1])

        print("method1 xl:", xl)


        ######################## 实现方法二
        xl = tf.transpose(x0, [0, 2, 1])
        for i in range(3):
            xl = tf.matmul(x0, xl, transpose_a=True, transpose_b=True)  ## batch_size, (5, 1)*(1, 5)*(5, 1)
            xl = tf.matmul(xl, self.cross_weight[i])
            xl = xl + self.bais_weight[i] + tf.transpose(x0, [0, 2, 1])
        xl = tf.transpose(xl, [0, 2, 1])
        print("method3 xl:", xl)

        ######################## 实现方法三

        xl = x0
        x0 = tf.transpose(x0, [0, 2, 1])  ## (2, 5, 1)
        for i in range(3):
            # i = 0
            xl = tf.matmul(x0, xl)  ## batch_size, (2, 5, 1)*(2, 1, 5) = (2, 5, 5)
            xl = tf.matmul(xl, self.cross_weight[i])  ## (2, 5, 5)*(2, 5, 1) = (2, 5 ,1)
            xl = xl + x0 + self.bais_weight[i]
            xl = tf.transpose(xl, [0, 2, 1])

        # xl = tf.transpose(xl, [0, 2, 1])
        print("method2 xl:", xl)
        return xl


if __name__ == '__main__':

    input = tf.constant([[1],
                         [2]], dtype=tf.int32)

    print("input:", input.shape)

    ru_ins = DeepCrossLater(5, "CrossNet")


    res = ru_ins(input)

    print("res:", res)