DeepFM模型tensorflow代码

以下是 DeepFM 模型的 TensorFlow 2.0 实现代码：

import tensorflow as tf
from tensorflow.keras.layers import Input, Dense, Dropout, Concatenate
from tensorflow.keras.regularizers import l2
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam

class DeepFM:
    def __init__(self, feature_dim, embedding_dim=8, hidden_units=(32, 32), dropout_rate=0.5, l2_reg=0.01, learning_rate=0.001):
        self.feature_dim = feature_dim
        self.embedding_dim = embedding_dim
        self.hidden_units = hidden_units
        self.dropout_rate = dropout_rate
        self.l2_reg = l2_reg
        self.learning_rate = learning_rate

    def build(self):
        input_layer = Input((self.feature_dim,))
        embedding_layer = tf.keras.layers.Embedding(self.feature_dim, self.embedding_dim)(input_layer)
        embedding_layer = tf.keras.layers.Flatten()(embedding_layer)

        linear_layer = tf.keras.layers.Dense(units=1, activation=None)(input_layer)

        fm_layer = tf.keras.layers.GlobalAveragePooling1D()(embedding_layer)

        deep_layer = tf.keras.layers.Dense(units=self.hidden_units[0], activation='relu')(embedding_layer)
        deep_layer = tf.keras.layers.Dropout(self.dropout_rate)(deep_layer)
        for units in self.hidden_units[1:]:
            deep_layer = tf.keras.layers.Dense(units=units, activation='relu')(deep_layer)
            deep_layer = tf.keras.layers.Dropout(self.dropout_rate)(deep_layer)

        deep_fm_layer = Concatenate()([fm_layer, deep_layer])
        output_layer = tf.keras.layers.Dense(units=1, activation='sigmoid', kernel_regularizer=l2(self.l2_reg))(deep_fm_layer)

        model = Model(inputs=input_layer, outputs=output_layer)

        optimizer = Adam(lr=self.learning_rate)
        model.compile(optimizer=optimizer, loss='binary_crossentropy', metrics=['accuracy'])

        return model

其中，`feature_dim` 表示特征的维度，`embedding_dim` 表示嵌入层的维度，`hidden_units` 表示深度部分的神经元数量，`dropout_rate` 表示 dropout 比例，`l2_reg` 表示 L2 正则项系数，`learning_rate` 表示学习率。在 `build()` 方法里，我们首先定义了输入层和嵌入层，然后分别计算了线性部分、FM 部分和深度部分的结果，最后将 FM 和深度部分的结果拼接起来，并通过一个全连接层输出最终结果。这里使用了 Adam 优化器和二分类交叉熵损失函数，并编译了模型。