人工智能深度学习入门练习之（28）TensorFlow – 例子：循环神经网络(RNN)

循环神经网络(RNN)是一种用于处理序列数据的人工神经网络，序列数据是相互依赖的（有限或无限）数据流，比如时间序列数据、信息性的字符串、对话等。

长短时记忆网络(LSTM)是一类特殊的循环神经网络，具有学习长时依赖关系的能力，是目前最常用的循环神经网络。

注意: 关于循环神经网络的介绍，可参考我们的教程深度学习 – 循环神经网络(RNN)。

我们的例子是训练一个LSTM模型，训练时模型会学习一段短文，完成训练后，模型可以在输入一个短句后，预测接下来的单词。

数据集

数据集是一段短文，《伊索寓言》中的老鼠给猫挂铃铛的故事：

long ago , the mice had a general council to consider what measures they could take to outwit their common enemy , the cat . some said this , and some said that but at last a young mouse got up and said he had a proposal to make , which he thought would meet the case . you will all agree , said he , that our chief danger consists in the sly and treacherous manner in which the enemy approaches us . now , if we could receive some signal of her approach , we could easily escape from her . i venture , therefore , to propose that a small bell be procured , and attached by a ribbon round the neck of the cat . by this means we should always know when she was about , and could easily retire while she was in the neighborhood . this proposal met with general applause , until an old mouse got up and said that is all very well , but who is to bell the cat ? the mice looked at one another and nobody spoke . then the old mouse said it is easy to propose impossible remedies .

这篇短文有112个不重复的符号，单词和标点符号都被认为是符号。

训练

如果我们向LSTM输入3个正确排序的符号，和一个标签符号，模型最终将学会正确预测下一个符号(图1)。

图1. 具有三个输入和一个输出的LSTM单元。

技术上来说，LSTM只能理解数字。因此，需要对上面的短文作一些处理，每个不重复的单词和标点都用数字代替，总共有112个不重复的符号。

我们将创建2个字典，一个可以从单词/标点映射到数字，另一个可以从数字映射到单词/标点(反向字典)。例如，上面的文本中有112个独特的符号，第一个字典包含以下条目[ “,” : 0 ] [ “the” : 1 ], …, [ “council” : 37 ],…,[ “spoke” : 111 ]。同时构建一个反向字典，将用于解码LSTM的输出。

LSTM输入输出都是数字，LSTM应该输出一个符号数字代号，用来表示符号。例如，如果输出是37，表示单词”council”。

但是，LSTM输出的实际上是一个112个元素的向量，每个元素值表示对应符号的概率，概率最大的符号就是最终符号(图2)。

图2. 每个输入符号都被替换成数字代号。输出是一个表示本次输出各符号概率的向量，读取概率最大的符号代号，结合反向字典，最终查出符号。

实现

下面是实现代码：

from __future__ import print_function

import numpy as np
import tensorflow as tf
from tensorflow.contrib import rnn
import random
import collections
import time

start_time = time.time()
def elapsed(sec):
    if sec<60:
        return str(sec) + " sec"
    elif sec<(60*60):
        return str(sec/60) + " min"
    else:
        return str(sec/(60*60)) + " hr"


# 日志目录
logs_path = './train/rnn_words'
writer = tf.summary.FileWriter(logs_path)

# 训练用的短文
training_file = 'belling_the_cat.txt'

# 读取短文函数
def read_data(fname):
    with open(fname) as f:
        content = f.readlines()
    content = [x.strip() for x in content]
    content = [word for i in range(len(content)) for word in content[i].split()]
    content = np.array(content)
    return content

training_data = read_data(training_file)
print("Loaded training data...")

# 短文符号->数字字典，数字->短文符号反向字典
def build_dataset(words):
    count = collections.Counter(words).most_common()
    dictionary = dict()
    for word, _ in count:
        dictionary[word] = len(dictionary)
    reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
    return dictionary, reverse_dictionary

# 创建字典与反向字典
dictionary, reverse_dictionary = build_dataset(training_data)
# 符号数量
vocab_size = len(dictionary)

# 参数
learning_rate = 0.001
training_iters = 50000
display_step = 1000
n_input = 3

# RNN cell中的神经数量
n_hidden = 512

# tf Graph input
x = tf.placeholder("float", [None, n_input, 1])
y = tf.placeholder("float", [None, vocab_size])

# RNN 输出节点的 weights 与 biases
weights = {
    'out': tf.Variable(tf.random_normal([n_hidden, vocab_size]))
}
biases = {
    'out': tf.Variable(tf.random_normal([vocab_size]))
}

def RNN(x, weights, biases):

    # reshape 到 [-1, n_input]
    x = tf.reshape(x, [-1, n_input])

    # Generate a n_input-element sequence of inputs
    # (eg. [had] [a] [general] -> [20] [6] [33])
    x = tf.split(x,n_input,1)

    # 2-layer LSTM, each layer has n_hidden units.
    # Average Accuracy= 95.20% at 50k iter
    rnn_cell = rnn.MultiRNNCell([rnn.BasicLSTMCell(n_hidden),rnn.BasicLSTMCell(n_hidden)])

    # 1-layer LSTM with n_hidden units but with lower accuracy.
    # Average Accuracy= 90.60% 50k iter
    # Uncomment line below to test but comment out the 2-layer rnn.MultiRNNCell above
    # rnn_cell = rnn.BasicLSTMCell(n_hidden)

    # generate prediction
    outputs, states = rnn.static_rnn(rnn_cell, x, dtype=tf.float32)

    # there are n_input outputs but
    # we only want the last output
    return tf.matmul(outputs[-1], weights['out']) + biases['out']

pred = RNN(x, weights, biases)

# 计算损失
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
# 优化
optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate).minimize(cost)

# 模型评估
correct_pred = tf.equal(tf.argmax(pred,1), tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

# 初始化变量
init = tf.global_variables_initializer()

# 执行图
with tf.Session() as session:
    session.run(init)
    step = 0
    offset = random.randint(0,n_input+1)
    end_offset = n_input + 1
    acc_total = 0
    loss_total = 0

    writer.add_graph(session.graph)

    while step < training_iters:
        # Generate a minibatch. Add some randomness on selection process.
        if offset > (len(training_data)-end_offset):
            offset = random.randint(0, n_input+1)

        # 准备样本数据和标签
        symbols_in_keys = [ [dictionary[ str(training_data[i])]] for i in range(offset, offset+n_input) ]
        symbols_in_keys = np.reshape(np.array(symbols_in_keys), [-1, n_input, 1])

        symbols_out_onehot = np.zeros([vocab_size], dtype=float)
        symbols_out_onehot[dictionary[str(training_data[offset+n_input])]] = 1.0
        symbols_out_onehot = np.reshape(symbols_out_onehot,[1,-1])

        # 执行optimizer, accuracy, cost, pred
        _, acc, loss, onehot_pred = session.run([optimizer, accuracy, cost, pred], \
                                                feed_dict={x: symbols_in_keys, y: symbols_out_onehot})
        loss_total += loss
        acc_total += acc

        # 每过一定步数(display_step)，打印信息
        if (step+1) % display_step == 0:
            print("Iter= " + str(step+1) + ", Average Loss= " + \
                  "{:.6f}".format(loss_total/display_step) + ", Average Accuracy= " + \
                  "{:.2f}%".format(100*acc_total/display_step))
            acc_total = 0
            loss_total = 0
            symbols_in = [training_data[i] for i in range(offset, offset + n_input)]
            symbols_out = training_data[offset + n_input]
            symbols_out_pred = reverse_dictionary[int(tf.argmax(onehot_pred, 1).eval())]
            print("%s - [%s] vs [%s]" % (symbols_in,symbols_out,symbols_out_pred))

        # 递增step, offset
        step += 1
        offset += (n_input+1)

    # 训练完成，打印输出
    print("Optimization Finished!")
    print("Elapsed time: ", elapsed(time.time() - start_time))
    print("Run on command line.")
    print("\ttensorboard --logdir=%s" % (logs_path))
    print("Point your web browser to: http://localhost:6006/")

    # 测试：接受用户输入，生成输出
    while True:
        prompt = "%s words: " % n_input
        sentence = input(prompt)
        sentence = sentence.strip()
        words = sentence.split(' ')
        if len(words) != n_input:
            continue
        try:
            symbols_in_keys = [dictionary[str(words[i])] for i in range(len(words))]

            # 连续进行32次
            for i in range(32):
                keys = np.reshape(np.array(symbols_in_keys), [-1, n_input, 1])
                onehot_pred = session.run(pred, feed_dict={x: keys})
                onehot_pred_index = int(tf.argmax(onehot_pred, 1).eval())
                sentence = "%s %s" % (sentence,reverse_dictionary[onehot_pred_index])
                symbols_in_keys = symbols_in_keys[1:]
                symbols_in_keys.append(onehot_pred_index)
            print(sentence)
        except:
            print("Word not in dictionary")

输出

Iter= 1000, Average Loss= 4.428141, Average Accuracy= 5.10%
['nobody', 'spoke', '.'] - [then] vs [then]
Iter= 2000, Average Loss= 2.937925, Average Accuracy= 17.60%
['?', 'the', 'mice'] - [looked] vs [looked]
Iter= 3000, Average Loss= 2.401870, Average Accuracy= 31.00%
['an', 'old', 'mouse'] - [got] vs [got]
Iter= 4000, Average Loss= 2.079050, Average Accuracy= 46.10%
['when', 'she', 'was'] - [about] vs [about]
Iter= 5000, Average Loss= 1.756826, Average Accuracy= 52.40%
['a', 'small', 'bell'] - [be] vs [be]
Iter= 6000, Average Loss= 1.620517, Average Accuracy= 57.80%
['from', 'her', '.'] - [i] vs [cat]
Iter= 7000, Average Loss= 1.410994, Average Accuracy= 61.50%
['some', 'signal', 'of'] - [her] vs [be]
Iter= 8000, Average Loss= 1.340336, Average Accuracy= 65.60%
['enemy', 'approaches', 'us'] - [.] vs [,]

...

Iter= 48000, Average Loss= 0.447481, Average Accuracy= 90.60%
['general', 'council', 'to'] - [consider] vs [consider]
Iter= 49000, Average Loss= 0.527762, Average Accuracy= 89.30%
['ago', ',', 'the'] - [mice] vs [mice]
Iter= 50000, Average Loss= 0.375872, Average Accuracy= 91.50%
['spoke', '.', 'then'] - [the] vs [the]
Optimization Finished!
Elapsed time:  31.482173871994018 min
Run on command line.
        tensorboard --logdir=./train/rnn_words
Point your web browser to: http://localhost:6006/
3 words: the mice had
the mice had a general council to said that is all a consider what measures they enemy , the cat . by this means we should always know when when when when when when when