中文聊天机器人（一）_用keras和seq2seq训练中文机器人

一、数据预处理

1.定义预处理参数，文件data_util.py

from keras.models import Model
from keras.layers import Input,LSTM,Dense
import numpy as np
import pandas as pd
 
num_samples = 100000
# 定义路径
question_path = 'question.txt'
answer_path = 'answer.txt'
 
 
max_encoder_seq_length = None
max_decoder_seq_length = None
num_encoder_tokens = None
num_decoder_tokens = None

2.获取训练数据X, Y

def get_xy_data():
    input_texts = []
    target_texts = []
    with open(question_path, 'r', encoding='utf-8') as f:
        input_texts = f.read().split('\n')
        input_texts = input_texts[:min(num_samples,len(input_texts)-1)]
    with open(answer_path, 'r', encoding='utf-8') as f:
        target_texts = ['\t' + line  + '\n' for line in f.read().split('\n')]
        target_texts = target_texts[:min(num_samples,len(input_texts)-1)]
    
    return input_texts, target_texts

3.需要将Input数据向量化，这里根据Input数据X, Y获取字符词典

def get_vocab_dict(X, Y):
    global max_encoder_seq_length, max_decoder_seq_length, num_encoder_tokens, num_decoder_tokens
    input_texts = X
    target_texts = Y
    input_characters = set()
    target_characters = set()
    for line in input_texts[:min(num_samples,len(input_texts)-1)]:
        for char in line:
            if char not in input_characters:
                input_characters.add(char)
    for line in target_texts[:min(num_samples,len(target_texts)-1)]:
        for char in line:
            if char not in target_characters:
                target_characters.add(char)
 
    input_characters = sorted(list(input_characters))
    target_characters = sorted(list(target_characters))
    num_encoder_tokens = len(input_characters)
    num_decoder_tokens = len(target_characters)
    max_encoder_seq_length = max([len(txt) for txt in input_texts])
    max_decoder_seq_length = max([len(txt) for txt in target_texts])
    
    print('Number of samples:', len(input_texts))
    print('Number of unique input tokens:', num_encoder_tokens)
    print('Number of unique output tokens:', num_decoder_tokens)
    print('Max sequence length for inputs:', max_encoder_seq_length)
    print('Max sequence length for outputs:', max_encoder_seq_length)
    
    input_token_index = dict(
            [(char,i) for i, char in enumerate(input_characters)])
    target_token_index = dict(
            [(char,i) for i, char in enumerate(target_characters)])
    
    return input_token_index, target_token_index

4.需要建立一个逆转词典，用于预测阶段将向量转化为可识别的字符

def get_rev_dict(input_token_index, target_token_index):
    reverse_input_char_index = dict(
        (i, char) for char, i in input_token_index.items())
    reverse_target_char_index  = dict(
        (i, char) for char, i in target_token_index.items())
    return reverse_input_char_index, reverse_target_char_index

二、训练

1.定义参数

from keras.models import Model
from keras.layers import Input,LSTM,Dense
import numpy as np
import pandas as pd
import data_util
from data_util import get_vocab_dict
from data_util import get_xy_data
 
# 定义超参数
batch_size = 32
epochs = 100
latent_dim = 256
 
input_texts = []
target_texts = []
input_token_index = []
target_token_index = []
encoder_input_data = None
decoder_input_data = None
decoder_target_data = None

2.调用预处理 data_util.py 得到训练数据和词典

def data_deal():
     global encoder_input_data,decoder_input_data,decoder_target_data
     global input_texts, target_texts, input_token_index,target_token_index
     input_texts, target_texts = get_xy_data()
     input_token_index, target_token_index = get_vocab_dict(input_texts, target_texts)
 
     # 每个input_text句子都是一个二维矩阵，
     # 那么input_texts是多个二维矩阵组合的三维矩阵
     encoder_input_data = np.zeros(
          (len(input_texts), data_util.max_encoder_seq_length, len(input_token_index)),dtype='float32')
     decoder_input_data = np.zeros(
          (len(input_texts), data_util.max_decoder_seq_length, len(target_token_index)),dtype='float32')
     decoder_target_data = np.zeros(
          (len(input_texts), data_util.max_decoder_seq_length, len(target_token_index)),dtype='float32')
 
     for i, (input_text, target_text) in enumerate(zip(input_texts, target_texts)):
          for t, char in enumerate(input_text):
               encoder_input_data[i, t, input_token_index[char]] = 1
          for t, char in enumerate(target_text):
               decoder_input_data[i, t, target_token_index[char]] = 1
               if t > 0:
                    decoder_target_data[i, t-1, target_token_index[char]] =1

3.建立seq2seq模型

def build_model():
     global input_token_index,target_token_index
     encoder_inputs = Input(shape=(None, len(input_token_index)))
     encoder = LSTM(latent_dim, return_state=True)
     encoder_outputs, state_h, state_c = encoder(encoder_inputs)
     encoder_states = [state_h, state_c]
     decoder_inputs = Input(shape=(None, len(target_token_index)))
     decoder_lstm = LSTM(latent_dim, return_sequences=True,return_state=True)
     decoder_outputs, _, _ = decoder_lstm(decoder_inputs,
                                        initial_state=encoder_states)
     decoder_dense = Dense(len(target_token_index), activation='softmax')
     decoder_outputs = decoder_dense(decoder_outputs)
     model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
     # 新序列预测时需要的encoder
     encoder_model = Model(encoder_inputs, encoder_states)
     # 新序列预测时需要的decoder
     decoder_state_input_h = Input(shape=(latent_dim,))
     decoder_state_input_c = Input(shape=(latent_dim,))
     decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
     decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs, initial_state=decoder_states_inputs)
     decoder_states = [state_h, state_c]
     decoder_outputs = decoder_dense(decoder_outputs)
     decoder_model = Model([decoder_inputs] + decoder_states_inputs, [decoder_outputs] + decoder_states)
 
     return model, encoder_model, decoder_model

4.训练模型并保存

# 训练并保存
if __name__ == "__main__":
     data_deal()
     model,encoder_model,decoder_model = build_model()
     model.compile(optimizer='rmsprop',loss='categorical_crossentropy')
     model.fit([encoder_input_data,decoder_input_data],decoder_target_data,
          batch_size=batch_size,
          epochs=epochs,
          validation_split=0.2)
     model.save('model.h5')
     encoder_model.save('encoder_model.h5')
     decoder_model.save('decoder_model.h5')

三、预测

1.定义参数

from keras.models import Model,load_model
from keras.layers import Input,LSTM,Dense
import numpy as np
import pandas as pd
 
from data_util import get_vocab_dict
from data_util import get_xy_data
from data_util import get_rev_dict
import data_util
 
latent_dim = 256
 
# 语料向量化
input_texts = []
target_texts = []
input_token_index = []
target_token_index = []

2.开始预测

# 开始inference
def decoder_sequence(input_seq):
     # Encode the input as state vectors
     states_value = encoder_model.predict(input_seq)
 
     target_seq = np.zeros((1,1,data_util.num_decoder_tokens))
     # '\t' is starting character
     target_seq[0,0,target_token_index['\t']] = 1
 
     # Sampling loop for a batch of sequences
     stop_condition = False
     decoded_sentence= ''
     while not stop_condition:
          output_tokens, h, c = decoder_model.predict(
               [target_seq] + states_value)
          sampled_token_index = np.argmax(output_tokens[0,-1,:])
          sampled_char = reverse_target_char_index[sampled_token_index]
          decoded_sentence += sampled_char
          if(sampled_char == '\n' or len(decoded_sentence) > data_util.max_decoder_seq_length):
               stop_condition = True
 
          # Update the target sequenco to predict next token
          target_seq = np.zeros((1,1,data_util.num_decoder_tokens))
          target_seq[0,0,sampled_token_index] = 1
 
          # Update state
          states_value = [h, c]
 
     return decoded_sentence
 
def predict_ans(question):
     input_seq = np.zeros((1, data_util.max_encoder_seq_length, data_util.num_encoder_tokens),dtype='float16')
     for t, char in list(enumerate(question)):
          input_seq[0,t,input_token_index[char]] = 1
     decoded_sentence = decoder_sequence(input_seq)
     return decoded_sentence
 
if __name__ == "__main__":
     input_texts, target_texts = get_xy_data()
     input_token_index, target_token_index = get_vocab_dict(input_texts, target_texts)
     reverse_input_char_index, reverse_target_char_index = get_rev_dict(input_token_index, target_token_index)
     encoder_model = load_model('encoder_model.h5')
     decoder_model = load_model('decoder_model.h5')
     print('Decoded sentence:', predict_ans('这是个傻子'))

github：项目链接