|
| 1 | +# -*- coding: utf-8 -*- |
| 2 | + |
| 3 | +import sys |
| 4 | +import math |
| 5 | +import tflearn |
| 6 | +import tensorflow as tf |
| 7 | +from tensorflow.python.ops import rnn_cell |
| 8 | +from tensorflow.python.ops import rnn |
| 9 | +import chardet |
| 10 | +import numpy as np |
| 11 | +import struct |
| 12 | + |
| 13 | +seq = [] |
| 14 | + |
| 15 | +max_w = 50 |
| 16 | +float_size = 4 |
| 17 | +word_vector_dict = {} |
| 18 | +word_vec_dim = 200 |
| 19 | +max_seq_len = 16 |
| 20 | + |
| 21 | +def load_vectors(input): |
| 22 | + """从vectors.bin加载词向量，返回一个word_vector_dict的词典，key是词，value是200维的向量 |
| 23 | + """ |
| 24 | + print "begin load vectors" |
| 25 | + |
| 26 | + input_file = open(input, "rb") |
| 27 | + |
| 28 | + # 获取词表数目及向量维度 |
| 29 | + words_and_size = input_file.readline() |
| 30 | + words_and_size = words_and_size.strip() |
| 31 | + words = long(words_and_size.split(' ')[0]) |
| 32 | + size = long(words_and_size.split(' ')[1]) |
| 33 | + print "words =", words |
| 34 | + print "size =", size |
| 35 | + |
| 36 | + for b in range(0, words): |
| 37 | + a = 0 |
| 38 | + word = '' |
| 39 | + # 读取一个词 |
| 40 | + while True: |
| 41 | + c = input_file.read(1) |
| 42 | + word = word + c |
| 43 | + if False == c or c == ' ': |
| 44 | + break |
| 45 | + if a < max_w and c != '\n': |
| 46 | + a = a + 1 |
| 47 | + word = word.strip() |
| 48 | + |
| 49 | + vector = [] |
| 50 | + for index in range(0, size): |
| 51 | + m = input_file.read(float_size) |
| 52 | + (weight,) = struct.unpack('f', m) |
| 53 | + vector.append(float(weight)) |
| 54 | + |
| 55 | + # 将词及其对应的向量存到dict中 |
| 56 | + #word_vector_dict[word.decode('utf-8')] = vector |
| 57 | + word_vector_dict[word.decode('utf-8')] = vector[0:word_vec_dim] |
| 58 | + |
| 59 | + input_file.close() |
| 60 | + |
| 61 | + print "load vectors finish" |
| 62 | + |
| 63 | +def init_seq(): |
| 64 | + """读取切好词的文本文件，加载全部词序列 |
| 65 | + """ |
| 66 | + file_object = open('zhenhuanzhuan.segment', 'r') |
| 67 | + vocab_dict = {} |
| 68 | + while True: |
| 69 | + line = file_object.readline() |
| 70 | + if line: |
| 71 | + for word in line.decode('utf-8').split(' '): |
| 72 | + if word_vector_dict.has_key(word): |
| 73 | + seq.append(word_vector_dict[word]) |
| 74 | + else: |
| 75 | + break |
| 76 | + file_object.close() |
| 77 | + |
| 78 | +def vector_sqrtlen(vector): |
| 79 | + len = 0 |
| 80 | + for item in vector: |
| 81 | + len += item * item |
| 82 | + len = math.sqrt(len) |
| 83 | + return len |
| 84 | + |
| 85 | +def vector_cosine(v1, v2): |
| 86 | + if len(v1) != len(v2): |
| 87 | + sys.exit(1) |
| 88 | + sqrtlen1 = vector_sqrtlen(v1) |
| 89 | + sqrtlen2 = vector_sqrtlen(v2) |
| 90 | + value = 0 |
| 91 | + for item1, item2 in zip(v1, v2): |
| 92 | + value += item1 * item2 |
| 93 | + return value / (sqrtlen1*sqrtlen2) |
| 94 | + |
| 95 | + |
| 96 | +def vector2word(vector): |
| 97 | + max_cos = -10000 |
| 98 | + match_word = '' |
| 99 | + for word in word_vector_dict: |
| 100 | + v = word_vector_dict[word] |
| 101 | + cosine = vector_cosine(vector, v) |
| 102 | + if cosine > max_cos: |
| 103 | + max_cos = cosine |
| 104 | + match_word = word |
| 105 | + return (match_word, max_cos) |
| 106 | + |
| 107 | + |
| 108 | +class MySeq2Seq(object): |
| 109 | + """ |
| 110 | + 思路：输入输出序列一起作为input，然后通过slick和unpack切分 |
| 111 | + 完全按照论文说的编码器解码器来做 |
| 112 | + 输出的时候把解码器的输出按照词向量的200维展平，这样输出就是(?,seqlen*200) |
| 113 | + 这样就可以通过regression来做回归计算了，输入的y也展平，保持一致 |
| 114 | + """ |
| 115 | + def __init__(self, max_seq_len = 16, word_vec_dim = 200): |
| 116 | + self.max_seq_len = max_seq_len |
| 117 | + self.word_vec_dim = word_vec_dim |
| 118 | + |
| 119 | + def generate_trainig_data(self): |
| 120 | + load_vectors("./vectors.bin") |
| 121 | + init_seq() |
| 122 | + xy_data = [] |
| 123 | + y_data = [] |
| 124 | + for i in range(30,40,10): |
| 125 | + # 问句、答句都是16字，所以取32个 |
| 126 | + start = i*self.max_seq_len*2 |
| 127 | + middle = i*self.max_seq_len*2 + self.max_seq_len |
| 128 | + end = (i+1)*self.max_seq_len*2 |
| 129 | + sequence_xy = seq[start:end] |
| 130 | + sequence_y = seq[middle:end] |
| 131 | + print "right answer" |
| 132 | + for w in sequence_y: |
| 133 | + (match_word, max_cos) = vector2word(w) |
| 134 | + print match_word |
| 135 | + sequence_y = [np.ones(self.word_vec_dim)] + sequence_y |
| 136 | + xy_data.append(sequence_xy) |
| 137 | + y_data.append(sequence_y) |
| 138 | + |
| 139 | + return np.array(xy_data), np.array(y_data) |
| 140 | + |
| 141 | + |
| 142 | + def model(self, feed_previous=False): |
| 143 | + # 通过输入的XY生成encoder_inputs和带GO头的decoder_inputs |
| 144 | + input_data = tflearn.input_data(shape=[None, self.max_seq_len*2, self.word_vec_dim], dtype=tf.float32, name = "XY") |
| 145 | + encoder_inputs = tf.slice(input_data, [0, 0, 0], [-1, self.max_seq_len, self.word_vec_dim], name="enc_in") |
| 146 | + decoder_inputs_tmp = tf.slice(input_data, [0, self.max_seq_len, 0], [-1, self.max_seq_len-1, self.word_vec_dim], name="dec_in_tmp") |
| 147 | + go_inputs = tf.ones_like(decoder_inputs_tmp) |
| 148 | + go_inputs = tf.slice(go_inputs, [0, 0, 0], [-1, 1, self.word_vec_dim]) |
| 149 | + decoder_inputs = tf.concat(1, [go_inputs, decoder_inputs_tmp], name="dec_in") |
| 150 | + |
| 151 | + # 编码器 |
| 152 | + # 把encoder_inputs交给编码器，返回一个输出(预测序列的第一个值)和一个状态(传给解码器) |
| 153 | + (encoder_output_tensor, states) = tflearn.lstm(encoder_inputs, self.word_vec_dim, return_state=True, scope='encoder_lstm') |
| 154 | + encoder_output_sequence = tf.pack([encoder_output_tensor], axis=1) |
| 155 | + |
| 156 | + # 解码器 |
| 157 | + # 预测过程用前一个时间序的输出作为下一个时间序的输入 |
| 158 | + # 先用编码器的最后一个输出作为第一个输入 |
| 159 | + if feed_previous: |
| 160 | + first_dec_input = go_inputs |
| 161 | + else: |
| 162 | + first_dec_input = tf.slice(decoder_inputs, [0, 0, 0], [-1, 1, self.word_vec_dim]) |
| 163 | + decoder_output_tensor = tflearn.lstm(first_dec_input, self.word_vec_dim, initial_state=states, return_seq=False, reuse=False, scope='decoder_lstm') |
| 164 | + decoder_output_sequence_single = tf.pack([decoder_output_tensor], axis=1) |
| 165 | + decoder_output_sequence_list = [decoder_output_tensor] |
| 166 | + # 再用解码器的输出作为下一个时序的输入 |
| 167 | + for i in range(self.max_seq_len-1): |
| 168 | + if feed_previous: |
| 169 | + next_dec_input = decoder_output_sequence_single |
| 170 | + else: |
| 171 | + next_dec_input = tf.slice(decoder_inputs, [0, i+1, 0], [-1, 1, self.word_vec_dim]) |
| 172 | + decoder_output_tensor = tflearn.lstm(next_dec_input, self.word_vec_dim, return_seq=False, reuse=True, scope='decoder_lstm') |
| 173 | + decoder_output_sequence_single = tf.pack([decoder_output_tensor], axis=1) |
| 174 | + decoder_output_sequence_list.append(decoder_output_tensor) |
| 175 | + |
| 176 | + decoder_output_sequence = tf.pack(decoder_output_sequence_list, axis=1) |
| 177 | + real_output_sequence = tf.concat(1, [encoder_output_sequence, decoder_output_sequence]) |
| 178 | + |
| 179 | + net = tflearn.regression(real_output_sequence, optimizer='sgd', learning_rate=0.1, loss='mean_square') |
| 180 | + model = tflearn.DNN(net) |
| 181 | + return model |
| 182 | + |
| 183 | + def train(self): |
| 184 | + trainXY, trainY = self.generate_trainig_data() |
| 185 | + model = self.model(feed_previous=False) |
| 186 | + model.fit(trainXY, trainY, n_epoch=1000, snapshot_epoch=False) |
| 187 | + model.save('./model/model') |
| 188 | + return model |
| 189 | + |
| 190 | + def load(self): |
| 191 | + model = self.model(feed_previous=True) |
| 192 | + model.load('./model/model') |
| 193 | + return model |
| 194 | + |
| 195 | +if __name__ == '__main__': |
| 196 | + phrase = sys.argv[1] |
| 197 | + my_seq2seq = MySeq2Seq(word_vec_dim=word_vec_dim, max_seq_len=max_seq_len) |
| 198 | + if phrase == 'train': |
| 199 | + my_seq2seq.train() |
| 200 | + else: |
| 201 | + model = my_seq2seq.load() |
| 202 | + trainXY, trainY = my_seq2seq.generate_trainig_data() |
| 203 | + predict = model.predict(trainXY) |
| 204 | + for sample in predict: |
| 205 | + print "predict answer" |
| 206 | + for w in sample[1:]: |
| 207 | + (match_word, max_cos) = vector2word(w) |
| 208 | + print match_word, max_cos |
0 commit comments