DSSM文本相似度

1.文本编码(类似准备mnist图像数据矩阵)
2.构造网络，损失函数是关于query-doc的cos相似度的
3.跑session，数据按batch送入网络

实验数据

搜索query \t 被点击doc-title \t label=1
搜索query \t 被点击条目的预测相似query \t label=0

支	支付宝官网	1
支	支气管炎的症状	0
支	支付宝	0
支	支原体感染	0
支	支付宝客服电话多少	0
东风风行x3	东风风行X3	1
东风风行x3	东风风行x3多少钱	0
东风风行x3	东风风行x3报价及图片	0

构造dssm网络图及run.sess训练

class Graph:
    def __init__(self):
        self.p = tf.placeholder(dtype=tf.int32, shape=(None, args.seq_length), name='p')
        self.h = tf.placeholder(dtype=tf.int32, shape=(None, args.seq_length), name='h')
        self.y = tf.placeholder(dtype=tf.int32, shape=None, name='y')
        self.keep_prob = tf.placeholder(dtype=tf.float32, name='drop_rate')
        self.embedding = tf.get_variable(dtype=tf.float32, shape=(args.vocab_size, args.char_embedding_size), name='embedding')   # 完整的嵌入张量
        self.forward()

    def dropout(self, x):
        return tf.nn.dropout(x, keep_prob=self.keep_prob)

    def fully_connect(self, x):
        x = tf.layers.dense(x, 128, activation='tanh')   # 全连接层
        x = self.dropout(x)
        x = tf.layers.dense(x, 256, activation='tanh')
        x = self.dropout(x)
        x = tf.layers.dense(x, 128, activation='tanh')
        x = self.dropout(x)
        x = tf.reshape(x, shape=(-1, x.shape[1] * x.shape[2]))   #
        return x

    @staticmethod
    def cosine(p, h):
        p_norm = tf.norm(p, axis=1, keepdims=True)
        h_norm = tf.norm(h, axis=1, keepdims=True)
        cosine = tf.reduce_sum(tf.multiply(p, h), axis=1, keepdims=True) / (p_norm * h_norm)
        return cosine

    def forward(self):
        p_embedding = tf.nn.embedding_lookup(self.embedding, self.p)
        h_embedding = tf.nn.embedding_lookup(self.embedding, self.h)
        p_context = self.fully_connect(p_embedding)
        h_context = self.fully_connect(h_embedding)
        pos_result = self.cosine(p_context, h_context)
        neg_result = 1 - pos_result
        logits = tf.concat([neg_result, pos_result], axis=1)
        self.train(logits)

    def train(self, logits):
        y = tf.one_hot(self.y, args.class_size)   # onehot矩阵维度：y的第一个维度*2列
        loss = tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=logits)
        self.loss = tf.reduce_mean(loss)
        self.train_op = tf.train.AdamOptimizer(args.learning_rate).minimize(self.loss)
        prediction = tf.argmax(logits, axis=1)   # neg类 预测0
        correct_prediction = tf.equal(tf.cast(prediction, tf.int32), self.y)
        self.acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

import os
import sys

sys.path.append(os.path.join(os.path.dirname(__file__), '../'))

from dssm.graph import Graph
import tensorflow as tf
from utils.load_data import load_char_data
from dssm import args

p, h, y = load_char_data('input/train_4neg_20180929.csv', data_size=None)    # 做成文本字粒度的index编码，整个文本编码完的长度 做了padding
p_eval, h_eval, y_eval = load_char_data('input/vali_4neg_20180929.csv', data_size=None)

p_holder = tf.placeholder(dtype=tf.int32, shape=(None, args.seq_length), name='p')
h_holder = tf.placeholder(dtype=tf.int32, shape=(None, args.seq_length), name='h')
y_holder = tf.placeholder(dtype=tf.int32, shape=None, name='y')

dataset = tf.data.Dataset.from_tensor_slices((p_holder, h_holder, y_holder))
dataset = dataset.batch(args.batch_size).repeat(args.epochs)
iterator = dataset.make_initializable_iterator()
next_element = iterator.get_next()

model = Graph()
saver = tf.train.Saver()

config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.9

with tf.Session(config=config)as sess:
    sess.run(tf.global_variables_initializer())
    sess.run(iterator.initializer, feed_dict={p_holder: p, h_holder: h, y_holder: y})   # iterator.initializer 先对迭代器初始化
    steps = int(len(y) / args.batch_size)   # 每个epoch多少个batch
    for epoch in range(args.epochs):
        for step in range(steps):
            p_batch, h_batch, y_batch = sess.run(next_element)
            _, loss, acc = sess.run([model.train_op, model.loss, model.acc],
                                    feed_dict={model.p: p_batch, model.h: h_batch, model.y: y_batch, model.keep_prob: args.keep_prob})
            if step % 50 == 0:
                print('epoch:', epoch, ' step:', step, ' loss: ', loss, ' acc:', acc)

        loss_eval, acc_eval = sess.run([model.loss, model.acc],
                                       feed_dict={model.p: p_eval, model.h: h_eval, model.y: y_eval, model.keep_prob: 1})
        print('loss_eval: ', loss_eval, ' acc_eval:', acc_eval)
        print('\n')
        saver.save(sess, f'../output/dssm/dssm_{epoch}.ckpt')

实验代码参考