Python实现FFM【tensorflow2.0】

糖逗

论文：https://www.csie.ntu.edu.tw/~cjlin/papers/ffm.pdf

参考：https://zhuanlan.zhihu.com/p/170607706

说明：待验证

import tensorflow as tf
from tensorflow.keras import layers, optimizers
from tensorflow import keras
from sklearn.model_selection import train_test_split
import numpy as np

#将数据划分为测试集和训练集
def preprocess(x, y):
    x = tf.cast(x, dtype = tf.float64)
    y = tf.cast(y, dtype = tf.int64)
    return x, y

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
data = load_breast_cancer()
x_train, x_test, y_train, y_test = train_test_split(data.data, data.target, test_size=0.2,
                                                        random_state = 11, stratify = data.target)
print(x_train.shape, y_train.shape, x_test.shape, y_test.shape)

train_db = tf.data.Dataset.from_tensor_slices((np.array(x_train), y_train))
train_db = train_db.shuffle(123).map(preprocess).batch(20)

test_db = tf.data.Dataset.from_tensor_slices((np.array(x_test), y_test))
test_db = test_db.map(preprocess).batch(20)

sample = next(iter(train_db))
print('sample:', sample[0].shape, sample[1].shape,
      tf.reduce_min(sample[0]), tf.reduce_max(sample[0]))

class FFM(keras.Model):
    def __init__(self, field_num, feature_field_dict, dim_num, k = 2):
        super(FFM, self).__init__()
        self.field_num = field_num
        self.k = k
        self.feature_field_dict = feature_field_dict
        self.dim_num = dim_num

    def build(self, input_shape):
        self.fc = tf.keras.layers.Dense(units = 1,
                                  bias_regularizer = tf.keras.regularizers.l2(0.01),
                                  kernel_regularizer = tf.keras.regularizers.l1(0.02))
        self.w = self.add_weight(shape = (input_shape[-1], self.field_num, self.k),
                                      initializer = 'glorot_uniform',
                                      trainable = True)
        super(FFM, self).build(input_shape)
        
    def call(self, x, training):
        linear = self.fc(x)
        temp = tf.cast(0, tf.float32)
        temp = tf.expand_dims(temp, axis = 0)
        for j1 in range(self.dim_num):
            for j2 in range(j1 + 1, self.dim_num):
                f1 = self.feature_field_dict[j2]
                f2 = self.feature_field_dict[j1]
                #[, , k] * [, , k] = [, , k] -> [1, k]
                ww = tf.expand_dims(tf.multiply(self.w[j1, f2, :], self.w[j2, f1, :]), axis = 0)
                #[x, ] * [x, ] = [x, ] -> [x, 1]
                xx = tf.expand_dims(tf.multiply(x[:, j1], x[:, j2]),axis = 1)
                #[x, 1] @ [1, k] = [x, k]
                store = tf.matmul(xx, ww)
                #[x, k] -> [x]
                temp += tf.reduce_mean(store, keepdims = True, axis = 1)
        out = layers.Add()([linear, temp])
        return tf.sigmoid(out)

store = {}   
for i in range(30):
        store[i] = int(i / 15)
model = FFM(field_num = 2, feature_field_dict = store, dim_num = 30)
model.build((None, 30))
model.summary()

def main():
    store = {}
    for i in range(30):
        store[i] = int(i / 15) #实际要根据数据字段含义定义，这里只是做一个随意的分组
    model = FFM(field_num = 2, feature_field_dict = store, dim_num = 30)
    optimizer = optimizers.Adam(lr = 1e-2)
    for epoch in range(50):
        for step, (x,y) in enumerate(train_db):
            with tf.GradientTape() as tape:
                logits = model(x,training=True)
                loss = tf.reduce_mean(tf.losses.binary_crossentropy(y, logits))
                loss_regularization = []
                for i in model.trainable_variables:
                    loss_regularization.append(tf.nn.l2_loss(i))
                loss_regularization = tf.reduce_sum(tf.stack(loss_regularization))
                loss = 0.001 * loss_regularization + loss
            grads = tape.gradient(loss, model.trainable_variables)
            optimizer.apply_gradients(zip(grads, model.trainable_variables))
            print(epoch, step, 'loss:', float(loss))
            
        total_num = 0
        total_correct = 0
        for x,y in test_db:
            pred = model(x, training=False)
            pred = tf.squeeze(pred)
            pred = pred > 0.5
            pred = tf.cast(pred, dtype = tf.int64)
            correct = tf.cast(tf.equal(pred, y), tf.int64)
            correct = tf.reduce_sum(correct)
            total_num += x.shape[0]
            total_correct += int(correct)
        acc = total_correct / total_num
        print(epoch, 'acc:', acc)
        print("-"*25)
        
if __name__ == '__main__':
    main()