66
77import constant
88from transform_data_lstm import TransformDataLSTM
9+ from seg_base import SegBase
910
1011
11- class SegLSTM :
12+ class SegLSTM ( SegBase ) :
1213 def __init__ (self ):
14+ SegBase .__init__ (self )
1315 self .dtype = tf .float32
1416 self .skip_window_left = constant .LSTM_SKIP_WINDOW_LEFT
1517 self .skip_window_right = constant .LSTM_SKIP_WINDOW_RIGHT
@@ -21,6 +23,7 @@ def __init__(self):
2123 trans = TransformDataLSTM ()
2224 self .words_batch = trans .words_batch
2325 self .tags_batch = trans .labels_batch
26+ self .dictionary = trans .dictionary
2427 self .vocab_size = constant .VOCAB_SIZE
2528 self .alpha = 0.02
2629 self .lam = 0.001
@@ -68,15 +71,16 @@ def model(self, embeds):
6871 return path
6972
7073 def train_exe (self ):
74+ saver = tf .train .Saver ([self .embeddings , self .A , self .init_A ].extend (self .params ), max_to_keep = 100 )
7175 self .sess .graph .finalize ()
7276 last_time = time .time ()
73- saver = tf .train .Saver ([self .embeddings , self .A ,self .init_A ].extend (self .params ), max_to_keep = 100 )
7477 for sentence_index , (sentence , tags ) in enumerate (zip (self .words_batch , self .tags_batch )):
7578 self .train_sentence (sentence , tags , len (tags ))
7679 if sentence_index % 500 == 0 :
80+ print (sentence_index )
7781 print (time .time () - last_time )
7882 last_time = time .time ()
79- 83+ saver . save ( self . sess , 'tmp/lstm-model%d.ckpt' % 0 )
8084
8185 def train_sentence (self , sentence , tags , length ):
8286 sentence_embeds = self .sess .run (self .lookup_op , feed_dict = {self .sentence_holder : sentence }).reshape (
@@ -109,7 +113,7 @@ def train_sentence(self, sentence, tags, length):
109113 # 更新词向量
110114 embed_index = sentence [update_index ]
111115 for i in range (update_length ):
112- embed = np .expand_dims (np .expand_dims (update_embed [:, i ], 0 ),0 )
116+ embed = np .expand_dims (np .expand_dims (update_embed [:, i ], 0 ),0 )
113117 grad = self .sess .run (self .grad_embed , feed_dict = {self .x_plus : embed ,
114118 self .map_matrix : np .expand_dims (sentence_matrix [:, i ], 1 )})[0 ]
115119
@@ -147,41 +151,6 @@ def gen_update_A(correct_tags, current_tags):
147151
148152 return A_update , init_A_update , update_init
149153
150- def viterbi (self , emission , A , init_A , return_score = False ):
151- """
152- 维特比算法的实现,所有输入和返回参数均为numpy数组对象
153- :param emission: 发射概率矩阵,对应于本模型中的分数矩阵,4*length
154- :param A: 转移概率矩阵,4*4
155- :param init_A: 初始转移概率矩阵,4
156- :param return_score: 是否返回最优路径的分值,默认为False
157- :return: 最优路径,若return_score为True,返回最优路径及其对应分值
158- """
159- 160- length = emission .shape [1 ]
161- path = np .ones ([4 , length ], dtype = np .int32 ) * - 1
162- corr_path = np .zeros ([length ], dtype = np .int32 )
163- path_score = np .ones ([4 , length ], dtype = np .float64 ) * (np .finfo ('f' ).min / 2 )
164- path_score [:, 0 ] = init_A + emission [:, 0 ]
165- 166- for pos in range (1 , length ):
167- for t in range (4 ):
168- for prev in range (4 ):
169- temp = path_score [prev ][pos - 1 ] + A [prev ][t ] + emission [t ][pos ]
170- if temp >= path_score [t ][pos ]:
171- path [t ][pos ] = prev
172- path_score [t ][pos ] = temp
173- 174- max_index = np .argmax (path_score [:, - 1 ])
175- corr_path [length - 1 ] = max_index
176- for i in range (length - 1 , 0 , - 1 ):
177- max_index = path [max_index ][i ]
178- corr_path [i - 1 ] = max_index
179- if return_score :
180- return corr_path , path_score [max_index , :]
181- else :
182- return corr_path
183- 184- 185154if __name__ == '__main__' :
186155 seg = SegLSTM ()
187156 seg .train_exe ()
0 commit comments