机器学习写唐诗

首先前提是安装了TensorFlow，按照官方文档进行就可以，有时候网络可能比较慢，有条件的话最好挂代理。注意Ubuntu本机一般是python2.7，最好使用virtualenv建立python3+的环境，否则如果导致本地python环境异常，得不偿失。在安装好tensorflow之后，就可以进行以下试验了。

唐诗语料后续补充下载链接，包含4万首唐诗，本文是基于腾讯云的一篇教程做的改进和记录，后面会放出详细的链接。教程里包括唐诗，验证码识别，聊天机器人等多个实验，都是可以简单修改跑通的例子，非常有价值，有兴趣可以了解下。

如下代码分为几个部分，1.语料整理，规范化。2.训练：两层RRN，使用LSTM模型训练。3.执行训练，默认的40000步。4.用于生成古诗。
先贴下代码。文字处理首要的是合适的训练资源外加处理，这部分其实会占很大的工作量。

本地是python3.5做了简单修改。

generate_poetry.py

#-*- coding:utf-8 -*-
import numpy as np
from io import open
import sys
import collections
import imp
imp.reload(sys)
#reload(sys)
#sys.setdefaultencoding('utf8')  //本地环境是python3.5做适配

class Poetry:
    def __init__(self):
        self.filename = "poetry"
        self.poetrys = self.get_poetrys()
        self.poetry_vectors,self.word_to_id,self.id_to_word = self.gen_poetry_vectors()
        self.poetry_vectors_size = len(self.poetry_vectors)
        self._index_in_epoch = 0

    def get_poetrys(self):
        poetrys = list()
        f = open(self.filename,"r", encoding='utf-8')
        for line in f.readlines():
            _,content = line.strip('\n').strip().split(':')
            content = content.replace(' ','')
            #过滤含有特殊符号的唐诗
            if(not content or '_' in content or '(' in content or '（' in content or "□" in content
                   or '《' in content or '[' in content or ':' in content or '：'in content):
                continue
            #过滤较长或较短的唐诗
            if len(content) < 5 or len(content) > 79:
                continue
            content_list = content.replace('，', '|').replace('。', '|').split('|')
            flag = True
            #过滤即非五言也非七验的唐诗
            for sentence in content_list:
                slen = len(sentence)
                if 0 == slen:
                    continue
                if 5 != slen and 7 != slen:
                    flag = False
                    break
            if flag:
                #每首古诗以'['开头、']'结尾
                poetrys.append('[' + content + ']')
        return poetrys

    def gen_poetry_vectors(self):
        words = sorted(set(''.join(self.poetrys) + ' '))
        #数字ID到每个字的映射
        id_to_word = {i: word for i, word in enumerate(words)}
        #每个字到数字ID的映射
        word_to_id = {v: k for k, v in id_to_word.items()}
        to_id = lambda word: word_to_id.get(word)
        #唐诗向量化
        poetry_vectors = [list(map(to_id, poetry)) for poetry in self.poetrys]
        return poetry_vectors,word_to_id,id_to_word

    def next_batch(self,batch_size):
        assert batch_size < self.poetry_vectors_size
        start = self._index_in_epoch
        self._index_in_epoch += batch_size
        #取完一轮数据，打乱唐诗集合，重新取数据
        if self._index_in_epoch > self.poetry_vectors_size:
            np.random.shuffle(self.poetry_vectors)
            start = 0
            self._index_in_epoch = batch_size
        end = self._index_in_epoch
        batches = self.poetry_vectors[start:end]
        x_batch = np.full((batch_size, max(map(len, batches))), self.word_to_id[' '], np.int32)
        for row in range(batch_size):
            x_batch[row,:len(batches[row])] = batches[row]
        y_batch = np.copy(x_batch)
        y_batch[:,:-1] = x_batch[:,1:]
        y_batch[:,-1] = x_batch[:, 0]

        return x_batch,y_batch

poetry_model.py

#-*- coding:utf-8 -*-
import tensorflow as tf

class poetryModel:
    #定义权重和偏置项
    def rnn_variable(self,rnn_size,words_size):
        with tf.variable_scope('variable'):
            w = tf.get_variable("w", [rnn_size, words_size])
            b = tf.get_variable("b", [words_size])
        return w,b

    #损失函数
    def loss_model(self,words_size,targets,logits):
        targets = tf.reshape(targets,[-1])
        loss = tf.contrib.legacy_seq2seq.sequence_loss_by_example([logits], [targets], [tf.ones_like(targets, dtype=tf.float32)],words_size)
        loss = tf.reduce_mean(loss)
        return loss

    #优化算子
    def optimizer_model(self,loss,learning_rate):
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(loss, tvars), 5)
        train_op = tf.train.AdamOptimizer(learning_rate)
        optimizer = train_op.apply_gradients(zip(grads, tvars))
        return optimizer

    #每个字向量化
    def embedding_variable(self,inputs,rnn_size,words_size):
        with tf.variable_scope('embedding'):
            with tf.device("/cpu:0"):
                embedding = tf.get_variable('embedding', [words_size, rnn_size])
                input_data = tf.nn.embedding_lookup(embedding,inputs)
        return input_data

    #构建LSTM模型
    def create_model(self,inputs,batch_size,rnn_size,words_size,num_layers,is_training,keep_prob):
        lstm = tf.contrib.rnn.BasicLSTMCell(num_units=rnn_size,state_is_tuple=True)
        input_data = self.embedding_variable(inputs,rnn_size,words_size)
        if is_training:
            lstm = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=keep_prob)
            input_data = tf.nn.dropout(input_data,keep_prob)
        cell = tf.contrib.rnn.MultiRNNCell([lstm] * num_layers,state_is_tuple=True)
        initial_state = cell.zero_state(batch_size, tf.float32)
        outputs,last_state = tf.nn.dynamic_rnn(cell,input_data,initial_state=initial_state)
        outputs = tf.reshape(outputs,[-1, rnn_size])
        w,b = self.rnn_variable(rnn_size,words_size)
        logits = tf.matmul(outputs,w) + b
        probs = tf.nn.softmax(logits)
        return logits,probs,initial_state,last_state

train_poetry.py

#-*- coding:utf-8 -*-
from generate_poetry import Poetry
from poetry_model import poetryModel
import tensorflow as tf
import numpy as np

if __name__ == '__main__':
    batch_size = 50
    epoch = 20
    rnn_size = 128
    num_layers = 2
    poetrys = Poetry()
    words_size = len(poetrys.word_to_id)
    inputs = tf.placeholder(tf.int32, [batch_size, None])
    targets = tf.placeholder(tf.int32, [batch_size, None])
    keep_prob = tf.placeholder(tf.float32, name='keep_prob')
    model = poetryModel()
    logits,probs,initial_state,last_state = model.create_model(inputs,batch_size,
                                                               rnn_size,words_size,num_layers,True,keep_prob)
    loss = model.loss_model(words_size,targets,logits)
    learning_rate = tf.Variable(0.0, trainable=False)
    optimizer = model.optimizer_model(loss,learning_rate)
    saver = tf.train.Saver()
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        sess.run(tf.assign(learning_rate, 0.002 * 0.97 ))
        next_state = sess.run(initial_state)
        step = 0
        while True:
            x_batch,y_batch = poetrys.next_batch(batch_size)
            feed = {inputs:x_batch,targets:y_batch,initial_state:next_state,keep_prob:0.5}
            train_loss, _ ,next_state = sess.run([loss,optimizer,last_state], feed_dict=feed)
            print("step:%d loss:%f" % (step,train_loss))
            if step > 40000:
                break
            if step%1000 == 0:
                n = step/1000
                sess.run(tf.assign(learning_rate, 0.002 * (0.97 ** n)))
            step += 1
        saver.save(sess,"poetry_model.ckpt")

predicty_poetry.py

#-*- coding:utf-8 -*-
from generate_poetry import Poetry
from poetry_model import poetryModel
from operator import itemgetter
import tensorflow as tf
import numpy as np
import random


if __name__ == '__main__':
    batch_size = 1
    rnn_size = 128
    num_layers = 2
    poetrys = Poetry()
    words_size = len(poetrys.word_to_id)

    def to_word(prob):
        prob = prob[0]
        indexs, _ = zip(*sorted(enumerate(prob), key=itemgetter(1)))
        rand_num = int(np.random.rand(1)*10);
        index_sum = len(indexs)
        max_rate = prob[indexs[(index_sum-1)]]
        if max_rate > 0.9 :
            sample = indexs[(index_sum-1)]
        else:
            sample = indexs[(index_sum-1-rand_num)]
        return poetrys.id_to_word[sample]

    inputs = tf.placeholder(tf.int32, [batch_size, None])
    keep_prob = tf.placeholder(tf.float32, name='keep_prob')
    model = poetryModel()
    logits,probs,initial_state,last_state = model.create_model(inputs,batch_size,
                                                               rnn_size,words_size,num_layers,False,keep_prob)
    saver = tf.train.Saver()
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        saver.restore(sess,"poetry_model.ckpt")
        next_state = sess.run(initial_state)

        x = np.zeros((1, 1))
        x[0,0] = poetrys.word_to_id['[']
        feed = {inputs: x, initial_state: next_state, keep_prob: 1}
        predict, next_state = sess.run([probs, last_state], feed_dict=feed)
        word = to_word(predict)
        poem = ''
        while word != ']':
            poem += word
            x = np.zeros((1, 1))
            x[0, 0] = poetrys.word_to_id[word]
            feed = {inputs: x, initial_state: next_state, keep_prob: 1}
            predict, next_state = sess.run([probs, last_state], feed_dict=feed)
            word = to_word(predict)
        print(poem)

最终的效果：
龙门不可见山人，日夕无情有故人。莫向西南不曾见，更应春雨在山风。

白雪新风月未同，山花一月一人春。风流白日春秋月，月色青松白玉衣。

poetry

实验原始链接：https://cloud.tencent.com/developer/labs/lab/10295
本地训练代码和数据：https://iss.igosh.com/share/201903/rrn_poem-me.tar.gz