Implementation of Hierarchical Attention Networks for Document Classification的讲解与Tensorflow实现

放完暑假回到实验室，tensorflow的代码已经忘得都差不多了。因此急需要对以前的代码进行复习，并且对未来一段时间的方向进行安排。在github上有一篇使用了最近几年深度学习常用模型，来做文本分类的例子。是作者brightmart参加知乎“看山杯”数据竞赛的各种模型的baseline的代码，因此拿来做以后这段时间学习的规划。下面是这个开源项目所用模型：

专栏里实现过使用char-CNN和char-RNN来做文本分类，TextCNN和TextRNN实际上和基于字符的文本分类过程差不多，数据处理阶段比基于字符的还要简单，所以不再单独实现。fastText模型更是极致的简单，这里也不再单独实现。模型4和模型5是下一阶段的主要实现目标。这篇文章主要讲加入attention模块的层次RNN，想法来自于论文Implementation of Hierarchical Attention Networks for Document Classification。代码已经上传至github。
Attention机制来自于论文 NEURAL MACHINE TRANSLATION BY JOINTLY LEARNING TO ALIGN AND TRANSLATE，最早提出是在机器翻译领域，这里引用该论文的插图来简要介绍attention机制的思想。

在上图中， xT 代表的是源语言的一个词， yt 代表的是目标语言的一个词。在得到翻译目标词 yt 的过程中，翻译结果取决于 yt 的上一个词 yt−1 和源语言的词 x1−−xT 的表示。比如，源语言为”I love you” 翻译为“我/爱/你”，在把翻译结果为”你”的过程中，起重要作用的是单词“you”而与”I love”无关，所以attention机制就是给源语言的词 x1−−xT 表示赋予权重 α ，以代表在翻译过程中，每个单词起的不同作用。
具体到Hierarchical Attention Networks文本分类的项目，框架可以分为四层，从下到上依次为：word encoder，word attention， sentence encoder， sentence attention。

在拿到一个新的论文以后，主要应该考虑两个问题：1、数据怎么处理？2、模型怎么训练？那我们首先解决第一个问题。

数据处理

首先说明数据集，yelp的review数据。大规模的数据可以在官网上下载，压缩数据大概2GB左右，还有一个较小规模的数据集，大概200M+，适用于我们进行练手。这里采用第二个数据，将yelp_academic_dataset_review.json文件解压，读取其中的’text’项作为文本，代表点评网站yelp上客人的评价；’starts’项作为分类结果，代表客人的评分。论文中提到的一些细节：

• 构建词库时，将在数据集中没有超过5次的词，统一作为UNK处理。
• 采用单个神经元采用GRU，要比LSTM得到的效果好。
• 采用通过word2vec方式预训练好的词向量。在本次试验中没有另外预训练，直接采用随机初始化嵌入层的矩阵，这样会导致小数据集的训练结果不好，但是比较方便。
• 词向量维度为2000维

有了这些参数的设置，再通过下面的注释，处理数据集的代码data_helper.py应该就能很好看懂了。在数据中，每篇文档的句子个数不一样，每个句子的单词数也不一样，本例中采用实验室同学的想法，将每篇文档都处理为30*30的矩阵，将每篇文章中多余30的句子截断，将每个句子中多于30的单词进行截断，当然这样的处理是很粗糙的，正确的想法应该是google的nmt采用bucket将句子个数相近的，单词个数相近的文章排到一起，然后取当前batch中最大的长度进行截断，但是我自己在写代码的过程中还是出现问题，所以留作待处理问题。

#coding=utf-8
import os
import json
import pickle
import nltk
import numpy as np
from nltk.tokenize import WordPunctTokenizer
from collections import defaultdict

# 使用nltk分词分句器
sent_tokenizer = nltk.data.load('tokenizers/punkt/english.pickle')
word_tokenizer = WordPunctTokenizer()


def build_vocab(vocab_path, yelp_json_path):

    if os.path.exists(vocab_path):
        vocab_file = open(vocab_path, 'rb')
        vocab = pickle.load(vocab_file)
        print "load focab finish!"
    else:
        # 记录每个单词及其出现的频率
        word_freq = defaultdict(int)
        # 读取数据集，并进行分词，统计每个单词出现次数，保存在word freq中
        with open(yelp_json_path, 'rb') as f:
            for line in f:
                review = json.loads(line)
                words = word_tokenizer.tokenize(review['text'])
                for word in words:
                    word_freq[word] += 1
            print "load finished"

        # 构建vocablary，并将出现次数小于5的单词全部去除，视为UNKNOW
        vocab = {}
        i = 1
        vocab['UNKNOW_TOKEN'] = 0
        for word, freq in word_freq.items():
            if freq > 5:
                vocab[word] = i
                i += 1

        # 将词汇表保存下来
        with open(vocab_path, 'wb') as g:
            pickle.dump(vocab, g)
            print len(vocab)  # 159654
            print "vocab save finished"

    return vocab

def load_dataset(yelp_json_path, max_sent_in_doc, max_word_in_sent):
    yelp_data_path = yelp_json_path[0:-5] + "_data.pickle"
    vocab_path = yelp_json_path[0:-5] + "_vocab.pickle"
    doc_num = 229907 #数据个数
    if not os.path.exists(yelp_data_path):

        vocab = build_vocab(vocab_path, yelp_json_path)
        num_classes = 5
        UNKNOWN = 0
        # data_x的shape: [doc_num,max_sent_in_doc,max_word_in_sent]
        # 而model中input_x的shape: [batch,max_sent_in_doc,max_word_in_sent]
        data_x = np.zeros([doc_num,max_sent_in_doc,max_word_in_sent])
        data_y = []

        #将所有的评论文件都转化为30*30的索引矩阵，也就是每篇都有30个句子，每个句子有30个单词
        # 不够的补零，多余的删除，并保存到最终的数据集文件之中
        with open(yelp_json_path, 'rb') as f:
            for line_index, line in enumerate(f):

                review = json.loads(line)
                sents = sent_tokenizer.tokenize(review['text'])
                doc = np.zeros([max_sent_in_doc, max_word_in_sent])

                for i, sent in enumerate(sents):
                    if i < max_sent_in_doc:
                        word_to_index = np.zeros([max_word_in_sent],dtype=int)
                        for j, word in enumerate(word_tokenizer.tokenize(sent)):
                            if j < max_word_in_sent:
                                    word_to_index[j] = vocab.get(word, UNKNOWN)
                        doc[i] = word_to_index

                data_x[line_index] = doc
                label = int(review['stars'])
                labels = [0] * num_classes
                labels[label-1] = 1
                data_y.append(labels)
                print line_index
            pickle.dump((data_x, data_y), open(yelp_data_path, 'wb'))
            print len(data_x) #229907
    else:
        data_file = open(yelp_data_path, 'rb')
        data_x, data_y = pickle.load(data_file)

    length = len(data_x)
    train_x, dev_x = data_x[:int(length*0.9)], data_x[int(length*0.9)+1 :]
    train_y, dev_y = data_y[:int(length*0.9)], data_y[int(length*0.9)+1 :]

    return train_x, train_y, dev_x, dev_y

if __name__ == '__main__':
    load_dataset("data/yelp_academic_dataset_review.json", 30, 30)

模型构建

这篇文章比较复杂的部分是模型构建。GRU的公式在Implementation of Hierarchical Attention Networks for Document Classification 论文中写的很清楚。在brightmart的例子的代码中，具体的写出了GRU的公式，当时就给我看跪了，我还傻傻的GRU和LSTM记不住呢！对于大多数初学者来说，可以使用tensorflow内置的RNN函数（在本例中为BidirectionalGRUEncoder）。对于想要实现这篇论文的同学，最好的办法还是clone github的代码，在模型中加断点观察shape，来与模型架构层次进行比对。

#coding=utf8
import tensorflow as tf
from tensorflow.contrib import rnn
from tensorflow.contrib import layers

def length(sequences):
    used = tf.sign(tf.reduce_max(tf.abs(sequences), reduction_indices=2))
    seq_len = tf.reduce_sum(used, reduction_indices=1)
    return tf.cast(seq_len, tf.int32)

class HAN():

    def __init__(self, vocab_size, num_classes, embedding_size=200, hidden_size=50):

        self.vocab_size = vocab_size
        self.num_classes = num_classes
        self.embedding_size = embedding_size
        self.hidden_size = hidden_size

        with tf.name_scope('placeholder'):
            self.max_sentence_num = tf.placeholder(tf.int32, name='max_sentence_num')
            self.max_sentence_length = tf.placeholder(tf.int32, name='max_sentence_length')
            self.batch_size = tf.placeholder(tf.int32, name='batch_size')
            #x的shape为[batch_size, 句子数， 句子长度(单词个数)]，但是每个样本的数据都不一样，，所以这里指定为空
            #y的shape为[batch_size, num_classes]
            self.input_x = tf.placeholder(tf.int32, [None, None, None], name='input_x')
            self.input_y = tf.placeholder(tf.float32, [None, num_classes], name='input_y')

        #构建模型
        word_embedded = self.word2vec()
        sent_vec = self.sent2vec(word_embedded)
        doc_vec = self.doc2vec(sent_vec)
        out = self.classifer(doc_vec)

        self.out = out


    def word2vec(self):
        with tf.name_scope("embedding"):
            embedding_mat = tf.Variable(tf.truncated_normal((self.vocab_size, self.embedding_size)))
            #shape为[batch_size, sent_in_doc, word_in_sent, embedding_size]
            word_embedded = tf.nn.embedding_lookup(embedding_mat, self.input_x)
        return word_embedded

    def sent2vec(self, word_embedded):
        with tf.name_scope("sent2vec"):
            #GRU的输入tensor是[batch_size, max_time, ...].在构造句子向量时max_time应该是每个句子的长度，所以这里将
            #batch_size * sent_in_doc当做是batch_size.这样一来，每个GRU的cell处理的都是一个单词的词向量
            #并最终将一句话中的所有单词的词向量融合（Attention）在一起形成句子向量

            #shape为[batch_size*sent_in_doc, word_in_sent, embedding_size]
            word_embedded = tf.reshape(word_embedded, [-1, self.max_sentence_length, self.embedding_size])
            #shape为[batch_size*sent_in_doce, word_in_sent, hidden_size*2]
            word_encoded = self.BidirectionalGRUEncoder(word_embedded, name='word_encoder')
            #shape为[batch_size*sent_in_doc, hidden_size*2]
            sent_vec = self.AttentionLayer(word_encoded, name='word_attention')
            return sent_vec

    def doc2vec(self, sent_vec):
        with tf.name_scope("doc2vec"):
            sent_vec = tf.reshape(sent_vec, [-1, self.max_sentence_num, self.hidden_size*2])
            #shape为[batch_size, sent_in_doc, hidden_size*2]
            doc_encoded = self.BidirectionalGRUEncoder(sent_vec, name='sent_encoder')
            #shape为[batch_szie, hidden_szie*2]
            doc_vec = self.AttentionLayer(doc_encoded, name='sent_attention')
            return doc_vec

    def classifer(self, doc_vec):
        with tf.name_scope('doc_classification'):
            out = layers.fully_connected(inputs=doc_vec, num_outputs=self.num_classes, activation_fn=None)
            return out

    def BidirectionalGRUEncoder(self, inputs, name):
        #输入inputs的shape是[batch_size, max_time, voc_size]
        with tf.variable_scope(name):
            GRU_cell_fw = rnn.GRUCell(self.hidden_size)
            GRU_cell_bw = rnn.GRUCell(self.hidden_size)
            #fw_outputs和bw_outputs的size都是[batch_size, max_time, hidden_size]
            ((fw_outputs, bw_outputs), (_, _)) = tf.nn.bidirectional_dynamic_rnn(cell_fw=GRU_cell_fw,
                                                                                 cell_bw=GRU_cell_bw,
                                                                                 inputs=inputs,
                                                                                 sequence_length=length(inputs),
                                                                                 dtype=tf.float32)
            #outputs的size是[batch_size, max_time, hidden_size*2]
            outputs = tf.concat((fw_outputs, bw_outputs), 2)
            return outputs

    def AttentionLayer(self, inputs, name):
        #inputs是GRU的输出，size是[batch_size, max_time, encoder_size(hidden_size * 2)]
        with tf.variable_scope(name):
            # u_context是上下文的重要性向量，用于区分不同单词/句子对于句子/文档的重要程度,
            # 因为使用双向GRU，所以其长度为2×hidden_szie
            u_context = tf.Variable(tf.truncated_normal([self.hidden_size * 2]), name='u_context')
            #使用一个全连接层编码GRU的输出的到期隐层表示,输出u的size是[batch_size, max_time, hidden_size * 2]
            h = layers.fully_connected(inputs, self.hidden_size * 2, activation_fn=tf.nn.tanh)
            #shape为[batch_size, max_time, 1]
            alpha = tf.nn.softmax(tf.reduce_sum(tf.multiply(h, u_context), axis=2, keep_dims=True), dim=1)
            #reduce_sum之前shape为[batch_szie, max_time, hidden_szie*2]，之后shape为[batch_size, hidden_size*2]
            atten_output = tf.reduce_sum(tf.multiply(inputs, alpha), axis=1)
            return atten_output

模型训练

模型训练部分的代码如下，都是一些常用的训练过程。模型的训练效果不太好，经过一万多步的训练，准确率只能达到40%左右。原因主要有两个，首先数据集较小，采用随机初始化的方式做嵌入层，词向量的表达就不太准确。其次，用30*30的矩阵代表一篇文章，并不能很好的表示文章，所以应该首先进行排序，将相似大小的文章放在一起然后再做截断或者填充。我会在实现了brightmart所有模型以后，再来处理这些细节的问题，从整体上看问题有助于理解文本分类的各种处理方法。

#coding=utf-8
import tensorflow as tf
import time
import os
from data_helper import load_dataset
from HAN_model import HAN


# Data loading params
tf.flags.DEFINE_string("yelp_json_path", 'data/yelp_academic_dataset_review.json', "data directory")
tf.flags.DEFINE_integer("vocab_size", 46960, "vocabulary size")
tf.flags.DEFINE_integer("num_classes", 5, "number of classes")
tf.flags.DEFINE_integer("embedding_size", 200, "Dimensionality of character embedding (default: 200)")
tf.flags.DEFINE_integer("hidden_size", 50, "Dimensionality of GRU hidden layer (default: 50)")
tf.flags.DEFINE_integer("batch_size", 32, "Batch Size (default: 64)")
tf.flags.DEFINE_integer("num_epochs", 10, "Number of training epochs (default: 50)")
tf.flags.DEFINE_integer("checkpoint_every", 100, "Save model after this many steps (default: 100)")
tf.flags.DEFINE_integer("num_checkpoints", 5, "Number of checkpoints to store (default: 5)")
tf.flags.DEFINE_integer("max_sent_in_doc", 30, "Number of checkpoints to store (default: 5)")
tf.flags.DEFINE_integer("max_word_in_sent", 30, "Number of checkpoints to store (default: 5)")
tf.flags.DEFINE_integer("evaluate_every", 100, "evaluate every this many batches")
tf.flags.DEFINE_float("learning_rate", 0.01, "learning rate")
tf.flags.DEFINE_float("grad_clip", 5, "grad clip to prevent gradient explode")


FLAGS = tf.flags.FLAGS

train_x, train_y, dev_x, dev_y = load_dataset(FLAGS.yelp_json_path, FLAGS.max_sent_in_doc, FLAGS.max_word_in_sent)
print "data load finished"

with tf.Session() as sess:
    han = HAN(vocab_size=FLAGS.vocab_size,
                    num_classes=FLAGS.num_classes,
                    embedding_size=FLAGS.embedding_size,
                    hidden_size=FLAGS.hidden_size)

    with tf.name_scope('loss'):
        loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=han.input_y,
                                                                      logits=han.out,
                                                                      name='loss'))
    with tf.name_scope('accuracy'):
        predict = tf.argmax(han.out, axis=1, name='predict')
        label = tf.argmax(han.input_y, axis=1, name='label')
        acc = tf.reduce_mean(tf.cast(tf.equal(predict, label), tf.float32))

    timestamp = str(int(time.time()))
    out_dir = os.path.abspath(os.path.join(os.path.curdir, "runs", timestamp))
    print("Writing to {}\n".format(out_dir))

    global_step = tf.Variable(0, trainable=False)
    optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
    # RNN中常用的梯度截断，防止出现梯度过大难以求导的现象
    tvars = tf.trainable_variables()
    grads, _ = tf.clip_by_global_norm(tf.gradients(loss, tvars), FLAGS.grad_clip)
    grads_and_vars = tuple(zip(grads, tvars))
    train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)

    # Keep track of gradient values and sparsity (optional)
    grad_summaries = []
    for g, v in grads_and_vars:
        if g is not None:
            grad_hist_summary = tf.summary.histogram("{}/grad/hist".format(v.name), g)
            grad_summaries.append(grad_hist_summary)

    grad_summaries_merged = tf.summary.merge(grad_summaries)

    loss_summary = tf.summary.scalar('loss', loss)
    acc_summary = tf.summary.scalar('accuracy', acc)


    train_summary_op = tf.summary.merge([loss_summary, acc_summary, grad_summaries_merged])
    train_summary_dir = os.path.join(out_dir, "summaries", "train")
    train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph)

    dev_summary_op = tf.summary.merge([loss_summary, acc_summary])
    dev_summary_dir = os.path.join(out_dir, "summaries", "dev")
    dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph)

    checkpoint_dir = os.path.abspath(os.path.join(out_dir, "checkpoints"))
    checkpoint_prefix = os.path.join(checkpoint_dir, "model")
    if not os.path.exists(checkpoint_dir):
        os.makedirs(checkpoint_dir)
    saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints)

    sess.run(tf.global_variables_initializer())

    def train_step(x_batch, y_batch):
        feed_dict = {
            han.input_x: x_batch,
            han.input_y: y_batch,
            han.max_sentence_num: 30,
            han.max_sentence_length: 30,
            han.batch_size: 64
        }
        _, step, summaries, cost, accuracy = sess.run([train_op, global_step, train_summary_op, loss, acc], feed_dict)

        time_str = str(int(time.time()))
        print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, cost, accuracy))
        train_summary_writer.add_summary(summaries, step)

        return step

    def dev_step(x_batch, y_batch, writer=None):
        feed_dict = {
            han.input_x: x_batch,
            han.input_y: y_batch,
            han.max_sentence_num: 30,
            han.max_sentence_length: 30,
            han.batch_size: 64
        }
        step, summaries, cost, accuracy = sess.run([global_step, dev_summary_op, loss, acc], feed_dict)
        time_str = str(int(time.time()))
        print("++++++++++++++++++dev++++++++++++++{}: step {}, loss {:g}, acc {:g}".format(time_str, step, cost, accuracy))
        if writer:
            writer.add_summary(summaries, step)

    for epoch in range(FLAGS.num_epochs):
        print('current epoch %s' % (epoch + 1))
        for i in range(0, 200000, FLAGS.batch_size):

            x = train_x[i:i + FLAGS.batch_size]
            y = train_y[i:i + FLAGS.batch_size]
            step = train_step(x, y)
            if step % FLAGS.evaluate_every == 0:
                dev_step(dev_x, dev_y, dev_summary_writer)