fix dnn and evaluation bug

supercoderhawk · supercoderhawk · commit d4efab666c09 · 2017-11-29T21:38:48.000+08:00
diff --git a/python/dnlp/core/dnn_crf.py b/python/dnlp/core/dnn_crf.py
@@ -30,6 +30,7 @@ def __init__(self, *, config: DnnCrfConfig = None, data_path: str = '', dtype: t
       self.seq_length = tf.placeholder(tf.int32, [self.batch_size])
     else:
       self.input = tf.placeholder(tf.int32, [None, self.windows_size])
+
     # &#26597;&#25214;&#34920;&#23618;
     self.embedding_layer = self.get_embedding_layer()
     # &#38544;&#34255;&#23618;
@@ -46,6 +47,9 @@ def __init__(self, *, config: DnnCrfConfig = None, data_path: str = '', dtype: t
 
     if mode == 'predict':
       self.output = tf.squeeze(tf.transpose(self.output), axis=2)
+      self.sess = tf.Session()
+      self.sess.run(tf.global_variables_initializer())
+      tf.train.Saver().restore(save_path=self.model_path, sess=self.sess)
     elif train == 'll':
       self.ll_loss, _ = tf.contrib.crf.crf_log_likelihood(self.output, self.real_indices, self.seq_length,
                                                           self.transition)
@@ -180,17 +184,15 @@ def generate_transition_update_index(self, correct_labels, current_labels):
   def predict(self, sentence: str, return_labels=False):
     if self.mode != 'predict':
       raise Exception('mode is not allowed to predict')
-    with tf.Session() as sess:
-      tf.global_variables_initializer().run()
-      tf.train.Saver().restore(save_path=self.model_path, sess=sess)
-      input = self.indices2input(self.sentence2indices(sentence))
-      runner = [self.output, self.transition, self.transition_init]
-      output, trans, trans_init = sess.run(runner, feed_dict={self.input: input})
-      labels = self.viterbi(output, trans, trans_init)
-      if not return_labels:
-        return self.tags2words(sentence, labels)
-      else:
-        return self.tags2words(sentence, labels), labels
+
+    input = self.indices2input(self.sentence2indices(sentence))
+    runner = [self.output, self.transition, self.transition_init]
+    output, trans, trans_init = self.sess.run(runner, feed_dict={self.input: input})
+    labels = self.viterbi(output, trans, trans_init)
+    if not return_labels:
+      return self.tags2words(sentence, labels)
+    else:
+      return self.tags2words(sentence, labels), self.tag2sequences(labels)
 
   def get_embedding_layer(self) -> tf.Tensor:
     embeddings = self.__get_variable([self.dict_size, self.embed_size], 'embeddings')
diff --git a/python/dnlp/utils/evaluation.py b/python/dnlp/utils/evaluation.py
@@ -1,6 +1,6 @@
 # -*- coding: UTF-8 -*-
 import pickle
-from dnlp.utils.constant import TAG_BEGIN, TAG_INSIDE, TAG_OTHER, TAG_END, TAG_SINGLE
+from dnlp.utils.constant import TAG_BEGIN, TAG_INSIDE, TAG_END, TAG_SINGLE
 
 
 def get_cws_statistics(correct_labels, predict_labels) -> (int, int, int):
@@ -33,7 +33,7 @@ def get_cws_statistics(correct_labels, predict_labels) -> (int, int, int):
       predicts[predict_start] = i
 
   for predict in predicts:
-    if corrects.get(predict) is not None and corrects[predict] == predicts[predict]:
+    if predict in corrects and corrects[predict] == predicts[predict]:
       true_positive_count += 1
 
   return true_positive_count, len(predicts), len(corrects)
@@ -72,22 +72,16 @@ def get_ner_statistics(correct_labels, predict_labels) -> (int, int, int):
 def evaluate_cws(model, data_path: str):
   with open(data_path, 'rb') as f:
     data = pickle.load(f)
-    dictionary = data['dictionary']
-    tags = data['tags']
-    reversed_map = dict(zip(tags.values(), tags.keys()))
     characters = data['characters']
     labels_true = data['labels']
     c_count = 0
     p_count = 0
-    r_count = -0
+    r_count = 0
     for sentence, label in enumerate(characters, labels_true):
       words, labels_predict = model.predict(sentence, return_labels=True)
-      seq = []
-      for l in zip(labels_predict):
-        seq.append(reversed_map[l])
-      c, p, r = get_cws_statistics(label, seq)
+      c, p, r = get_cws_statistics(label, labels_predict)
       c_count += c
       p_count += p
       r_count += r
-      print(c / p)
-      print(c / r)
+    print(c_count / p_count)
+    print(c_count / r_count)
