今天周六，准备在家休息，顺便思考了一下最近编程的内容，想把程序运用到古代诗词上，这样就可以用
深度学习的技术写诗了，最后根据读入成功的写出了诗词，感觉挺好玩的！
使用神经网络训练写诗
上次已经把Keras使用神经网络的所有内容都介绍完了，这次我们就自己搭建一个网络来试试让程序写诗吧！
首先我们要导入可能需要用到的包：
from __future__ import print_function
from keras.models import Sequential
from keras.layers import Dense, Activation
from keras.layers import Dropout
from keras.optimizers import RMSprop
from keras.utils import to_categorical
from keras.layers import LSTM,GRU,Bidirectional
from keras.layers.convolutional import Conv1D,Conv2D
from keras.layers.pooling import MaxPooling1D,AveragePooling1D,GlobalMaxPooling1D
from keras.callbacks import EarlyStopping
from keras.utils import plot_model
from keras.models import Model
from keras.layers import Input, Dense, Permute, Reshape, Lambda, RepeatVector,merge,Embedding
import keras.backend as K
import keras
import read_utils
import os
import pickle
import numpy as np
import sys
import json
import dense_new
然后通过这里下载所需要的诗词，我为了简单测试，只使用了一个文件，格式是json的，因此需要进行 数据读取：
def readJsonPoetry(fn,pl):
'''
fn代表读取json文件的路径，pl代表读取的诗词的长度
'''
with open(fn,encoding='UTF-8') as f:
content = f.read()
JsonResult = json.loads(content)
allX = []
allY = []
allData = []
for dic in JsonResult:
singleP = dic['paragraphs']
for line in singleP:
if len(line) == pl*2+2 and line[pl]=='，' and line[pl*2+1]=='。':
allX.append(line)
temline = line[1:] + 'e'
allY.append(temline)
for word in line:
allData.append(word)
allData.append('e')
return allX,allY,allData
实现用于预测的类：
class UserRNNReader():
def __init__(self,modelFile,converter,start,end,steps):
self.model = keras.models.load_model(modelFile)
self.converter = converter
self.startToken = start
self.endToken = end
self.steps = steps
def prediction_next_n(self, prime, vocab_size, next_n=3,**k): #prime：开始的几个基本单元；vocab_size：一共有多少个类型的基本单元+1(未知数据编码)
index = len(prime)-1 #检索的位置
prime = [prime]
x = keras.preprocessing.sequence.pad_sequences(prime, maxlen=self.steps, dtype='int32',padding='post', value=self.converter.word_to_int_table[self.endToken])
# x = to_categorical(x, self.converter.vocab_size)
# x = x.reshape(-1,x.shape[0],x.shape[1])
y = self.model.predict(x)
preds = y[0,index,:]
p = np.squeeze(preds) #squeeze函数从数组的形状中删除单维度条目，即把shape中为1的维度去掉
# 将next_n个最大的概率的位置得到
next_n_num = np.argsort(p,kind = 'mergesort')[-next_n:] #argsort函数可以按照给定方法排序
#返回的应该是标号和对应的概率值
s_p_d = []
for i in next_n_num:
s_p_d.append((i,p[i]))
return s_p_d
下面就是进行网络构建：
if __name__ == '__main__':
batch_size = 100 #每批训练的大小
poetry_line_length = 12 #输入是固定的序列长，包括一个逗号和一个句号
#数据读取
JsonFileName = 'p.json'
allX,allY,allData = readJsonPoetry(JsonFileName,int(poetry_line_length/2-1))
converter = read_utils.TextConverter(allData)
x_train,y_train = getNumberData(allX[:1000],allY[:1000],converter)
x_test,y_test = getNumberData(allX[1001:],allY[1001:],converter)
inp = Input(shape=(poetry_line_length,),name='Input')
emb = Embedding(input_dim=converter.vocab_size,
output_dim=64,
input_length=poetry_line_length)(inp)
# 下面是普通RNN网络
rnn1 = GRU(128, input_shape=(poetry_line_length, converter.vocab_size),return_sequences=True,dropout=0.5,name='RNN')(emb)
rnn2 = GRU(128,return_sequences=True,dropout=0.5,name='2RNN')(rnn1)
#下面是LSTM网络
lstm1 = LSTM(128, input_shape=(poetry_line_length, converter.vocab_size),return_sequences=True,dropout=0.5,name = 'LSTM')(emb)
lstm2 = LSTM(128,return_sequences=True,dropout=0.5,name = '2LSTM')(lstm1)
#下面attention
# attention_mul = attention_3d_block(lstm1,n_steps,True)
#下面是最后全连接输出
# dense1 = Dense(128,name='Dense_hind')(emb)
op_dense = Dense(converter.vocab_size,name='Dense')(rnn2) #这里修改输出的从哪里开始
#attention应该在这里！
attention_mul = attention_3d_block(op_dense,poetry_line_length,True)
op = Activation('softmax',name='Softmax_Activate')(attention_mul)
model = Model(inputs=inp,outputs=op)
#模型的所有参数和层进行总结
model.summary()
# read_activations.get_activations(model,)
#简单绘制一下网络图的结构
plot_model(model, to_file='model.png')
# sys.exit() #这里测试，先退出
有了网络后就可以进行损失函数，相关条目操作还有训练和预测了：
#设置优化器和损失函数
optimizer = RMSprop(lr=0.01) #该优化器通常是面对递归神经网络时的一个良好选择
# optimizer = keras.optimizers.Adam()
model.compile(loss='categorical_crossentropy', optimizer=optimizer,metrics=['accuracy'])
#设置回调函数相关的操作
# early_stopping_monitor = EarlyStopping(patience=5) #早停
tensor_board = keras.callbacks.TensorBoard() #数据输出到TensorBoard
#下面设置模型保存点,一步检测一次，只保存最好的模型
check_point = keras.callbacks.ModelCheckpoint(
filepath='logs\weights-{epoch:03d}-{val_acc:.4f}.hdf5',
monitor='val_acc',
period=1,
save_best_only=True)
# histories = Histories() #记录一些数据
hsty = model.fit(x_train, y_train, batch_size=batch_size, epochs=100,validation_data=(x_test,y_test),
callbacks=[tensor_board,check_point]) # 训练
# callbacks=[early_stopping_monitor,tensor_board,histories]) # 训练
model.save('my_model.h5')
这样就完成了模型的训练程序，进行训练后可以在本地保存最后一次训练完成后的模型，之后 我们就可以读取模型并进行写诗：
import os
#import model
import read_utils
import pickle
import tensorflow as tf
#from treelib import Tree
from graphviz import Digraph
from dense_new import Dense_New
import gen_poetry
import keras
import random
if __name__ == '__main__':
startToken = 'start'
endToken = 'e'
poetry_line_length = 12
n_steps = 100
#数据读取
JsonFileName = 'p.json'
allX,allY,allData = gen_poetry.readJsonPoetry(JsonFileName,int(poetry_line_length/2-1))
converter = read_utils.TextConverter(allData)
from gen_poetry import UserRNNReader
with keras.utils.CustomObjectScope({'Dense_New': Dense_New}):
model = UserRNNReader('my_model.h5',converter,startToken,endToken,poetry_line_length)
next_n = 3 #设置预测步数
before_list = [converter.word_to_int_table['爱']]
pred = model.prediction_next_n(before_list,converter.vocab_size,next_n)
for i in range(4):
num = pred[random.randint(0,2)][0]
before_list.append(num)
pred = model.prediction_next_n(before_list,converter.vocab_size,next_n)
before_list.append(converter.word_to_int_table['，'])
for i in range(5):
num = pred[random.randint(0,2)][0]
before_list.append(num)
pred = model.prediction_next_n(before_list,converter.vocab_size,next_n)
before_list.append(converter.word_to_int_table['。'])
w = converter.arr_to_text(before_list)
print(w)
上面的代码就是开头是“爱”字，写出的一首诗。
最后，本篇的全部代码在下面这个网页可以下载：
https://github.com/Dongzhixiao/Python_Exercise/tree/master/LSTM
