def evaluate(self,
data_sentences,
data_chars,
data_tags,
data_labels,
ignore_label=None,
batch_size=64,
simple_compute=True):
"""
Args:
data_sentences, data_tags, data_labels: np.array
ignore_label: int, 负例的编号，或者None
simple_compute: bool, 是否画出性能详细指标表格
Return:
p, r, f1
"""
pre_labels = []
if int(len(data_labels) / batch_size) == (len(data_labels) * 1.0 /
batch_size):
nb_dev = int(len(data_labels) / batch_size)
else:
nb_dev = int(len(data_labels) / batch_size) + 1
for i in tqdm(range(nb_dev)):
sentences_feed = data_sentences[i * batch_size:(i + 1) *
batch_size]
tags_feed = data_tags[i * batch_size:(i + 1) * batch_size]
labels_feed = data_labels[i * batch_size:(i + 1) * batch_size]
feed_dict = {
self.input_sentence_ph: sentences_feed,
self.input_tag_ph: tags_feed,
self.label_ph: labels_feed,
self.keep_prob_ph: 1.0,
self.word_keep_prob_ph: 1.0,
self.tag_keep_prob_ph: 1.0
}
if config.use_chars:
char_ids = data_chars[i * batch_size:(i + 1) * batch_size]
char_feed, word_lengths_feed = cnn_bilstm_load_data.pad_sequences(
char_ids, config.MAX_LEN, pad_tok=0, nlevels=2)
feed_dict[self.input_char_ph] = char_feed
feed_dict[self.word_lengths] = word_lengths_feed
pre_temp = self.sess.run(self.pre_op, feed_dict=feed_dict)
pre_labels += list(pre_temp)
right_labels = data_labels[:len(pre_labels)]
pre, rec, f = sim_compute(pre_labels,
right_labels,
ignore_label=ignore_label)
return pre, rec, f
def predict(self, data_sentences, data_chars, data_tags, batch_size=50):
"""
Args:
data_sentences, data_tags: np.array
batch_size: int
Return:
pre_labels: list
"""
pre_labels = []
pre_proba = []
nb_test = 0
if int(len(data_sentences) / batch_size) == (len(data_sentences) *
1.0 / batch_size):
nb_test = int(len(data_sentences) / batch_size)
else:
nb_test = int(len(data_sentences) / batch_size) + 1
for i in range(nb_test):
sentences_feed = data_sentences[i * batch_size:(i + 1) *
batch_size]
tags_feed = data_tags[i * batch_size:(i + 1) * batch_size]
feed_dict = {
self.input_sentence_ph: sentences_feed,
self.input_tag_ph: tags_feed,
self.keep_prob_ph: 1.0,
self.word_keep_prob_ph: 1.0,
self.tag_keep_prob_ph: 1.0
}
if config.use_chars:
char_ids = data_chars[i * batch_size:(i + 1) * batch_size]
char_feed, word_lengths_feed = cnn_bilstm_load_data.pad_sequences(
char_ids, config.MAX_LEN, pad_tok=0, nlevels=2)
feed_dict[self.input_char_ph] = char_feed
feed_dict[self.word_lengths] = word_lengths_feed
#pre_temp,pre_out = self.sess.run([self.pre_op, self.pre_ouput_op],feed_dict=feed_dict)
pre_proba_tmp, pre_temp = self.sess.run(
[self.proba_op, self.pre_op], feed_dict=feed_dict)
pre_labels += list(pre_temp)
pre_proba += list(pre_proba_tmp)
return pre_labels, pre_proba
def fit_all(self, sentences_train, chars_train, tags_train, labels_train,
batch_size=64, nb_epoch=40, keep_prob=1.0, word_keep_prob=1.0,
tag_keep_prob=1.0, seed=137):
"""
fit model
Args:
sentences_train, tags_train, labels_train: 训练数据
sentences_dev, tags_dev, labels_dev: 开发数据
batch_size: int, batch size
nb_epoch: int, 迭代次数
keep_prob: float between [0, 1], 全连接层前的dropout
word_keep_prob: float between [0, 1], 词向量层dropout
tag_keep_prob: float between [0, 1], 标记向量层dropout
"""
self.saver = tf.train.Saver()
self.nb_epoch_scores = [] # 存放nb_epoch次迭代的f值
n_total = len(labels_train)
nb_train = 0
if int( n_total / batch_size) == (n_total *1.0 / batch_size):
nb_train = int(n_total/ batch_size)
else :
nb_train = int(n_total / batch_size) + 1
for step in range(nb_epoch):
print('Epoch %d / %d:' % (step+1, nb_epoch))
# train
total_loss = 0.
for i in tqdm( range(nb_train) ):
# for i in range(nb_train):
#start = i * batch_size % n_total
if (i+1)*batch_size >= n_total:
sentences_feed = sentences_train[i * batch_size:] + sentences_train[0:(i+1) * batch_size-n_total]
tags_feed = tags_train[i * batch_size:] + tags_train[0:(i+1) * batch_size-n_total]
labels_feed = labels_train[i * batch_size:] + labels_train[0:(i+1) * batch_size-n_total]
char_ids = chars_train[i * batch_size:] + chars_train[0:(i+1) * batch_size-n_total]
else :
sentences_feed = sentences_train[i*batch_size:(i+1)*batch_size]
tags_feed = tags_train[i*batch_size:(i+1)*batch_size]
labels_feed = labels_train[i*batch_size:(i+1)*batch_size]
char_ids = chars_train[i*batch_size:(i+1)*batch_size]
feed_dict = {
self.input_sentence_ph: sentences_feed,
self.input_tag_ph: tags_feed,
self.label_ph: labels_feed,
self.keep_prob_ph: keep_prob,
self.word_keep_prob_ph: word_keep_prob,
self.tag_keep_prob_ph: tag_keep_prob,
}
if config.use_chars:
#char_ids = chars_train[i*batch_size:(i+1)*batch_size]
char_feed, word_lengths_feed = cnn_bilstm_load_data.pad_sequences(char_ids, config.MAX_LEN, pad_tok=0,nlevels=2)
feed_dict[self.input_char_ph] = char_feed
feed_dict[self.word_lengths] = word_lengths_feed
_, loss_value = self.sess.run(
[self.train_op, self.loss], feed_dict=feed_dict)
total_loss += loss_value
total_loss /= float(nb_train)
# 计算在训练集、开发集、测试集上的性能
p_train, r_train, f_train = self.evaluate(sentences_train,chars_train, tags_train, labels_train,batch_size=batch_size)
#pre_labels = self.predict(sentences_test, tags_test)
#with codecs.open('./Data/result/epoch_%d.csv' % (step+1), 'w', encoding='utf-8') as file_w:
#for num, label in enumerate(pre_labels):
#file_w.write('%d,%s\n' % (num+1, self._label_voc_rev[label]))
print('\tloss=%f, p_train=%f, r_train=%f, f_train=%f' % (total_loss,p_train, r_train, f_train))
self.nb_epoch_scores.append([total_loss,p_train, f_train])
self.saver.save(self.sess, config.TRAIN_ALL_MODEL)
print('model has saved to %s' % config.TRAIN_ALL_MODEL)
return self.nb_epoch_scores