def run_epoch(self, sess, train, train_deps, dev, dev_deps, vocab_words,
vocab_tags, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
"""
self.config.istrain = True # set to train first, #batch normalization#
nbatches = (len(train_deps) + self.config.batch_size -
1) / self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words, poss, chunks, labels, btup_idx_list, btup_words_list,
btup_depwords_list, btup_deprels_list,
btup_depwords_length_list, upbt_idx_list, upbt_words_list,
upbt_depwords_list, upbt_deprels_list,
upbt_depwords_length_list, btup_formidx_list,
upbt_formidx_list) in enumerate(
minibatches(train, train_deps, self.config.batch_size)):
fd, sequence_lengths = self.get_feed_dict(
words, poss, chunks, labels, btup_idx_list, btup_words_list,
btup_depwords_list, btup_deprels_list,
btup_depwords_length_list, upbt_idx_list, upbt_words_list,
upbt_depwords_list, upbt_deprels_list,
upbt_depwords_length_list, btup_formidx_list,
upbt_formidx_list, self.config.lr, self.config.dropout)
_, train_loss, logits = sess.run(
[self.train_op, self.loss, self.logits], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
acc, recall, f1, test_acc = self.run_evaluate(sess, dev, dev_deps,
vocab_words, vocab_tags)
self.logger.info(
"- dev acc {:04.2f} - dev recall {:04.2f} - f1 {:04.2f} - test acc {:04.2f}"
.format(100 * acc, 100 * recall, 100 * f1, 100 * test_acc))
return acc, recall, f1, train_loss
def run_epoch(self, sess, train, dev, tags, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
Args:
sess: tensorflow session
train: dataset that yields tuple of sentences, tags
dev: dataset
tags: {tag: index} dictionary
epoch: (int) number of the epoch
"""
nbatches = (len(train) + self.config.batch_size -
1) // self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words,
labels) in enumerate(minibatches(train,
self.config.batch_size)):
fd, _ = self.get_feed_dict(words, labels, self.config.lr,
self.config.dropout)
_, train_loss, summary = sess.run(
[self.train_op, self.loss, self.merged], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
# tensorboard
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch * nbatches + i)
acc, f1 = self.run_evaluate(sess, dev, tags)
self.logger.info("- dev acc {:04.2f} - f1 {:04.2f}".format(
100 * acc, 100 * f1))
return acc, f1
def run_epoch(self, sess, train, dev, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
Args:
sess: tensorflow session
train: dataset that yields tuple of sentences, tags
dev: dataset
epoch: (int) number of the epoch
"""
nbatches = (len(train) + self.config.batch_size -
1) / self.config.batch_size
prog = Progbar(target=nbatches, verbose=False)
for i, (framelist,
phones) in enumerate(minibatches(train,
self.config.batch_size)):
fd, _ = self.get_feed_dict(framelist, phones, self.config.lr,
self.config.keep_prob)
_, train_loss, summary = sess.run(
[self.train_op, self.loss, self.merged], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
# tensorboard
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch * nbatches + i)
acc, per = self.run_evaluate(sess, dev)
self.logger.info(" - dev accuracy {:04.2f} - PER {:04.2f}".format(
100 * acc, 100 * per))
return acc, per
def run_epoch(self, sess, train, dev, tags, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
Args:
sess: tensorflow session
train: dataset that yields tuple of sentences, tags
dev: dataset
tags: {tag: index} dictionary
epoch: (int) number of the epoch
"""
nbatches = (len(train) + self.config.batch_size - 1) // self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words, labels) in enumerate(minibatches(train, self.config.batch_size)):
fd, _ = self.get_feed_dict(words, labels, self.config.lr, self.config.dropout)
_, train_loss, summary = sess.run([self.train_op, self.loss, self.merged], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
# tensorboard
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch*nbatches + i)
acc, f1 = self.run_evaluate(sess, dev, tags)
self.logger.info("- dev acc {:04.2f} - f1 {:04.2f}".format(100*acc, 100*f1))
return acc, f1
def run_epoch(self, session, train):
"""
Perform one complete pass over the training data and evaluate on dev
"""
nbatches = (len(train) + self.config.batch_size -
1) / self.config.batch_size
prog = Progbar(target=nbatches)
total_loss = 0
for i, (q_batch, c_batch, a_batch) in enumerate(
minibatches(train, self.config.batch_size)):
# at training time, dropout needs to be on.
input_feed = self.get_feed_dict(q_batch, c_batch, a_batch,
self.config.dropout_val)
_, train_loss = session.run([self.train_op, self.loss],
feed_dict=input_feed)
# tf.summary.scalar("train loss", train_loss)
prog.update(i + 1, [("train loss", train_loss)])
total_loss += train_loss
self.losses.append(total_loss)
print(">>>>>>>", str(total_loss))
def run_evaluate(self, test):
"""
在test数据集上进行验证,输出准确率和召回率
:param test:由(sequence, tags)组成的list
:return:
"""
nbatches = (len(test) + batch_size - 1) // batch_size # 训练batch的次数
prog = Progbar(target=nbatches)
accs = []
i = 0
cnt_eq = 0
for words1, words2, labels in minibatches(test, batch_size):
logits = self.predict_batch(words1, words2)
for logit, label in zip(logits[0], labels):
# print(logit, np.argmax(logit), label)
accs += [np.argmax(logit) == label]
if np.argmax(logit) == label:
cnt_eq += 1
pass
prog.update(i + 1, [("evaluate acc", 100 * np.mean(accs))])
i += 1
print('cnt_eq=', cnt_eq)
pass
# print('cnt_eq=', cnt_eq)
acc = np.mean(accs)
return {"acc": 100 * acc}
def run_epoch(self, sess, config, dataset, train_writer, merged):
prog = Progbar(target=1 + len(dataset.train_inputs[0]) / config.batch_size)
for i, (train_x, train_y) in enumerate(get_minibatches([dataset.train_inputs, dataset.train_targets],
config.batch_size, is_multi_feature_input=True)):
summary, loss = self.train_on_batch(sess, train_x, train_y, merged)
prog.update(i + 1, [("train loss", loss)])
# train_writer.add_summary(summary, global_step=i)
return summary, loss # Last batch
def run_epoch(self, sess, parser, train_examples, dev_set):
prog = Progbar(target=1 + len(train_examples) / self.config.batch_size)
for i, (train_x, train_y) in enumerate(minibatches(train_examples, self.config.batch_size)):
loss = self.train_on_batch(sess, train_x, train_y)
prog.update(i + 1, [("train loss", loss)])
print "Evaluating on dev set",
dev_UAS, _ = parser.parse(dev_set)
print "- dev UAS: {:.2f}".format(dev_UAS * 100.0)
return dev_UAS
def run_epoch(self, sess, train, dev, tags, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
Args:
sess: tensorflow session
train: dataset that yields tuple of sentences, tags
dev: dataset
tags: {tag: index} dictionary
epoch: (int) number of the epoch
"""
#trie setting
self.lis1 = []
self.lis2 = []
self.lis3 = []
self.lis4 = []
self.lis5 = []
trie.gazette(self.lis1, "data/dic/gazette.txt")
trie.gazette(self.lis2, "data/dic/thres3.txt")
trie.gazette_DTTI(self.lis3, "data/dic/DT_analysis.txt")
trie.gazette_DTTI(self.lis4, "data/dic/TI_analysis.txt")
trie.gazette(self.lis5, "data/dic/wiki_PS.txt")
nbatches = (len(train) + self.config.batch_size -
1) // self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words, fw_words, bw_words, labels, postags, sentences,
_) in enumerate(minibatches(train, self.config.batch_size)):
dict_labels = self.dict_trie(sentences)
fd, _ = self.get_feed_dict(words,
fw_words,
bw_words,
dict_labels,
labels,
self.config.lr,
self.config.dropout,
test_flag=0)
_, train_loss, summary = sess.run(
[self.train_op, self.loss, self.merged], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
# tensorboard
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch * nbatches + i)
acc, f1, p, r = self.run_evaluate(sess, dev, tags, test_flag=0)
self.logger.info(
"- dev acc {:04.2f} - f1 {:04.2f} - p {:04.2f} - r {:04.2f}".
format(100 * acc, 100 * f1, 100 * p, 100 * r))
return acc, f1
def run_epoch(session, config, model, data, eval_op, keep_prob, is_training):
n_samples = len(data[0])
print("Running %d samples:" % (n_samples))
minibatches = get_minibatches_idx(n_samples,
config.batch_size,
shuffle=False)
correct = 0.
total = 0
total_cost = 0
prog = Progbar(target=len(minibatches))
#dummynode_hidden_states_collector=np.array([[0]*config.hidden_size])
to_print_total = np.array([[0] * 2])
for i, inds in enumerate(minibatches):
x = data[0][inds]
if sys.argv[4] == 'cnn':
x = pad_sequences(x,
maxlen=700,
dtype='int32',
padding='post',
truncating='post',
value=0.)
else:
x = pad_sequences(x,
maxlen=None,
dtype='int32',
padding='post',
truncating='post',
value=0.)
y = data[1][inds]
mask = data[2][inds]
count, _, cost, to_print= \
session.run([model.accuracy, eval_op,model.cost, model.to_print],\
{model.input_data: x, model.labels: y, model.mask:mask, model.keep_prob:keep_prob})
if not is_training:
to_print_total = np.concatenate((to_print_total, to_print), axis=0)
correct += count
total += len(inds)
total_cost += cost
prog.update(i + 1, [("train loss", cost)])
#if not is_training:
# print(to_print_total[:, 0].tolist())
# print(data[1].tolist())
# print(data[2].tolist())
print("Total loss:")
print(total_cost)
accuracy = correct / total
return accuracy
def load_model(data):
embeddings = get_embedding_data()
nbatches = (len(data) + batch_size - 1) // batch_size # 训练batch的次数
prog = Progbar(target=nbatches)
accs = []
i = 0
with tf.Session() as sess:
saver = tf.train.import_meta_graph(
'save/SentenceRelationModel1/SentenceRelationModel1_nl_3_hsl_196__.cpkt.meta'
)
saver.restore(
sess, tf.train.latest_checkpoint("save/SentenceRelationModel1/"))
print(tf.get_collection('logit'))
y = tf.get_collection('logit')
graph = tf.get_default_graph()
seq1_word_embeddings = graph.get_operation_by_name(
'seq1_word_embeddings').outputs[0]
seq2_word_embeddings = graph.get_operation_by_name(
'seq2_word_embeddings').outputs[0]
sequence1_lengths1 = graph.get_operation_by_name(
'sequence1_lengths').outputs[0]
sequence2_lengths1 = graph.get_operation_by_name(
'sequence2_lengths').outputs[0]
for words1, words2, labels in minibatches(data, batch_size):
words1, sequence1_lengths = pad_sequences(words1)
words2, sequence2_lengths = pad_sequences(words2)
words1_embeddings = [[
embeddings[w1] if w1 in embeddings.keys() else embeddings[","]
for w1 in ws1
] if len(ws1) > 0 else [embeddings["。"]] for ws1 in words1]
words2_embeddings = [[
embeddings[w2] if w2 in embeddings.keys() else embeddings[","]
for w2 in ws2
] if len(ws2) > 0 else [embeddings["。"]] for ws2 in words2]
sess.run(y,
feed_dict={
seq1_word_embeddings: words1_embeddings,
seq2_word_embeddings: words2_embeddings,
sequence1_lengths: sequence1_lengths1,
sequence2_lengths: sequence2_lengths1
})
for logit, label in zip(y[0], labels):
accs += [np.argmax(logit) == label]
pass
prog.update(i + 1, [("evaluate acc", 100 * np.mean(accs))])
i += 1
pass
acc = np.mean(accs)
return {"acc": 100 * acc}
def run_epoch(self, session, train):
"""
Perform one complete pass over the training data and evaluate on dev
"""
nbatches = (len(train) + self.config.batch_size - 1) / self.config.batch_size
prog = Progbar(target=nbatches)
for i, (q_batch, c_batch, a_batch) in enumerate(minibatches(train, self.config.batch_size)):
# at training time, dropout needs to be on.
input_feed = self.get_feed_dict(q_batch, c_batch, a_batch, self.config.dropout_val)
_, train_loss = session.run([self.train_op, self.loss], feed_dict=input_feed)
prog.update(i + 1, [("train loss", train_loss)])
summary = session.run(self.merged, input_feed)
self.writer.add_summary(summary, i)
def run_epoch(self, sess, train, dev, tags, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
Args:
sess: tensorflow session
train: dataset that yields tuple of sentences, tags
dev: dataset
tags: {tag: index} dictionary
epoch: (int) number of the epoch
"""
nbatches = (len(train) + self.config.batch_size - 1) // self.config.batch_size
prog = Progbar(target=nbatches)
total_loss = 0.0
count = 0
for i, (words, labels, iob, mention_type, mentions, _) in enumerate(minibatches(train, self.config.batch_size)):
mention_sizes = []
for m in mentions:
mention_sizes.append(len(m))
fd, _ = self.get_feed_dict(words, self.config.lr, self.config.dropout, iob, mention_type, mentions, mention_sizes)
logits, _, a, b, train_loss= sess.run([self.boundry_logits, self.train_op, self.loss_a, self.loss_b, self.loss], feed_dict=fd)
total_loss += train_loss
count += 1
print total_loss/count
acc, f1 = self.run_evaluate(sess, dev, tags)
self.logger.info("- dev acc {:04.2f} - f1 {:04.2f}".format(100*acc, 100*f1))
return acc, f1
def run_epoch(self, session, train):
"""
Perform one complete pass over the training data and evaluate on dev
"""
nbatches = (len(train) + config.BATCH_SIZE - 1) / config.BATCH_SIZE
prog = Progbar(target=nbatches)
for i, q_batch in enumerate(minibatches(train, config.BATCH_SIZE)):
# at training time, dropout needs to be on.
input_feed = self.get_feed_dict(q_batch, config.DROPOUT_VAL)
_, train_loss = session.run([self.train_op, self.loss],
feed_dict=input_feed)
prog.update(i + 1, [("train loss", train_loss)])
def train(self, train, dev):
best_score = 0
nepoch_no_imprv = 0 # for early stopping
self.add_summary() # tensorboard
for epoch in range(self.config.nepochs):
self.logger.info("Epoch {:} out of {:}".format(epoch + 1,
self.config.nepochs))
batch_size = self.config.batch_size
nbatches = (len(train) + batch_size - 1) // batch_size
prog = Progbar(target=nbatches)
#self.config.lr *= self.config.lr_decay
for i, (words, labels, intent, all_tags) in enumerate(minibatches(train, batch_size)):
fd, _ = self.get_feed_dict(words, all_tags, labels, intent, self.config.lr,\
self.config.dropout)
_, train_loss, summary, intent_loss, slot_loss= self.sess.run(
[self.train_op, self.loss, self.merged, self.intent_loss, self.slot_loss], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss), \
("intent_loss", intent_loss), ("slot_loss", slot_loss)])
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch*nbatches + i)
metrics = self.run_evaluate(dev)
msg = " - ".join(["{} {:04.2f}".format(k, v)
for k, v in metrics.items()])
self.logger.info(msg)
score = metrics["f1"] + metrics["intent_acc"]
self.config.lr *= self.config.lr_decay
if score >= best_score:
nepoch_no_imprv = 0
self.save_session()
best_score = score
self.logger.info("- new best score!")
else:
nepoch_no_imprv += 1
if nepoch_no_imprv >= self.config.nepoch_no_imprv:
if not self.embedding_trainable:
self.logger.info("fine tuning word embedding")
for i in range(10):
self.logger.info("######################")
self.set_word_embeddings_trainable()
self.config.lr = 0.001
nepoch_no_imprv = 0
else:
self.logger.info("- early stopping {} epochs without "\
"improvement".format(nepoch_no_imprv))
break
def run_epoch(session,
config,
model,
data,
eval_op,
keep_prob,
is_training,
saver=None):
n_samples = len(data[0])
minibatches = get_minibatches_idx(n_samples,
config.batch_size,
shuffle=config.shuffle)
correct = 0.
total = 0
total_cost = 0
prog = Progbar(target=len(minibatches))
for i, idx in enumerate(minibatches):
x = data[0][idx]
chars = data[3][idx]
x, mask = pad_sequences(x, 0, max_seq_len=config.max_seq_len)
if [] in chars:
continue
chars, char_mask, max_word_len = pad_sequences(
chars,
pad_tok=0,
nlevels=2,
max_seq_len=config.max_seq_len,
max_word_len=config.max_word_len)
y = data[1][idx]
max_word_len = max_word_len if max_word_len < config.max_word_len else config.max_word_len
global_step, count, _, cost = session.run([model.global_step, model.accuracy, eval_op, model.cost],\
{model.input_data: x, model.labels: y, model.mask: mask, model.keep_prob: keep_prob, model.chars: chars, model.char_mask: char_mask, model.max_word_len : max_word_len})
correct += count
total += len(idx)
total_cost += cost
if is_training:
prog.update(i + 1, [("train loss", cost), ("step", global_step)])
if global_step % config.save_step == 0 and saver:
saver.save(session, config.save_path, global_step=global_step)
del cost
accuracy = correct / total
return accuracy, total_cost,
def run_epoch(self, sess, train, dev, epoch):
nbatches = (len(train) + self.config.batch_size -
1) // self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words, imp_labels) in enumerate(
minibatches(train, self.config.batch_size)):
if self.config.model == "lstm_crf":
imp_labels = list(map(self.config.digitize_labels, imp_labels))
fd, _ = self.get_feed_dict(words, imp_labels, self.config.lr,
self.config.dropout)
_, train_loss = sess.run([self.optimize_, self.loss], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
result = self.run_evaluate(sess, dev)
self.logger.info(
"- dev acc {:04.4f} - f {:04.4f} - rms {:04.4f}".format(
100 * result['accuracy'], 100 * result['f-score'],
-1 * result['rms']))
return result
def run_epoch(self, train, dev, train_eval, epoch):
"""Performs one complete pass over the train set and evaluate on dev
Args:
train: dataset that yields tuple of sentences, tags
dev: dataset
epoch: (int) index of the current epoch
Returns:
f1: (python float), score to select model on, higher is better
"""
# progbar stuff for logging
batch_size = self.config.batch_size
nbatches = (len(train) + batch_size - 1) // batch_size
prog = Progbar(target=nbatches)
# iterate over dataset
for i, (words, labels) in enumerate(minibatches(train, batch_size)):
fd, _ = self.get_feed_dict(True, words, labels, lr=self.config.lr)
_, train_loss = self.sess.run([self.train_op, self.loss],
feed_dict=fd)
prog.update(i + 1, values=[("train loss", train_loss)])
acc_train = self.evaluate(train_eval)
acc_test = self.evaluate(dev)
prog.update(i + 1,
epoch, [("train loss", train_loss)],
exact=[("dev acc", acc_test), ("train acc", acc_train),
("lr", self.config.lr)])
return acc_train, acc_test, train_loss
def run_epoch(self, train, dev, epoch):
"""
在训练集上和测试集上完整地跑一回
:param train:(sentences, tags)，其中sentences尚未对齐
:param dev:类似于train的测试数据集
:param epoch:当前epoch的编号
:return f1:准确率
"""
nbatches = (len(train) + batch_size - 1) // batch_size # 训练batch的次数
prog = Progbar(target=nbatches)
# 首先遍历整个训练集
for i, (words1, words2,
labels) in enumerate(minibatches(train, batch_size)):
fd, _, _ = self.get_feed_dict(words1, words2, labels, lr, dropout)
_, train_loss = self.sess.run([self.train_op, self.loss],
feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
pass
# 接下来是测试了
metrics = self.run_evaluate(dev)
# print('第{}次的准确率为:\t{}\t召回率为:\t{}'.format(epoch, metrics['acc'], metrics['rec']))
return metrics["acc"] # 返回准确率
def run_epoch(self, train, dev, epoch):
"""Performs one complete pass over the train set and evaluate on dev
Args:
train: dataset that yields tuple of sentences, tags
dev: dataset
epoch: (int) index of the current epoch
Returns:
f1: (python float), score to select model on, higher is better
"""
# progbar stuff for logging
batch_size = self.config.batch_size
nbatches = (len(train) + batch_size - 1) // batch_size
prog = Progbar(target=nbatches)
# iterate over dataset
for i, (words, labels, pred_flags) in enumerate(minibatches(train, batch_size)):
fd, _ = self.get_feed_dict(words, labels, pred_flags, self.config.lr,
self.config.dropout)
_, train_loss, train_summary = self.sess.run(
[self.train_op, self.loss, self.train_merged], feed_dict=fd)
prog.update(i + 1, [("{}_train loss".format(self.config.name), train_loss)])
# tensorboard
if i % 10 == 0:
self.train_file_writer.add_summary(train_summary, epoch*nbatches + i)
yield None
metrics = self.run_evaluate(dev, log_step=epoch*nbatches+i)
msg = " - ".join(["{} {:04.2f}".format(k, v)
for k, v in metrics.items()])
self.logger.info(msg)
# Score for early stopping
return self.config.early_stop_metric_sign * metrics[self.config.early_stopping_metric]
def run_epoch(self, sess, train, dev, tags, epoch):
nbatches = (len(train) + self.config.batch_size -
1) // self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words,
labels) in enumerate(minibatches(train,
self.config.batch_size)):
fd, _ = self.get_feed_dict(words, labels, self.config.lr,
self.config.dropout)
_, train_loss, summary = sess.run(
[self.train_op, self.loss, self.merged], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
# tensorboard
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch * nbatches + i)
acc, f1 = self.evaluate(sess, dev, tags)
self.logger.info("- dev acc {:04.2f} - f1 {:04.2f}".format(
100 * acc, 100 * f1))
return acc, f1
def run_epoch(self, train, dev, epoch):
nbatches = (len(train) + self.batch_size - 1) // self.batch_size
prog = Progbar(target=nbatches) #进度条
batches = batch_yield(train, self.batch_size)
for i, (sentence, label) in enumerate(batches):
fd, _ = self.get_feed_dict(sentence, self.weight_dropout_list,
label, self.lr, self.dropout)
_, train_loss, summary = self.sess.run(
[self.train_op, self.loss, self.merged], feed_dict=fd)
prog.update(i + 1, [("train loss", train_loss)])
if i % 10 == 0:
self.file_writer.add_summary(summary, epoch * nbatches + i)
acc, p, r, f1 = self.run_evaluate(dev)
self.logger.info(
"- dev acc {:04.2f} - p {:04.2f}- r {:04.2f}- f1 {:04.2f}".format(
100 * acc, 100 * p, 100 * r, 100 * f1))
return acc, p, r, f1
def run_epoch(self, sess, train, dev, epoch):
'''
Performs one complete pass over the train set and evaluates on dev
Args:
sess: tensorflow session
train: large BIODataSentence list (training set)
dev: large BIODataSentence list (dev set)
epoch: (int) number of the epoch
'''
nbatches = len(train) // self.config.batch_size
if len(train) % self.config.batch_size != 0:
nbatches += 1
prog = Progbar(target=nbatches)
for i, sent_batch in enumerate(
minibatches(train, self.config.batch_size)):
fd, _ = self.prepare_feed_dict_optimized( \
bio_data_sentence_batch=sent_batch, \
dropout_keep_prob=self.config.dropout_keep_prob, \
learning_rate=self.config.learning_rate)
#_, train_loss, summary = sess.run([self.train_op, self.loss, self.merged], feed_dict=fd)
_, train_loss = sess.run([self.train_op, self.loss], feed_dict=fd)
prog.update(i + 1, [('train loss', train_loss)])
# tensorboard
#if i % 10 == 0:
#self.file_writer.add_summary(summary, epoch*nbatches + i)
acc, f1, mod_p = self.run_evaluate(sess, dev)
self.logger.info(
'- dev acc {:04.2f} - f1 {:04.2f} - mod prec {:04.2f}'.format(
100 * acc, 100 * f1, 100 * mod_p))
return acc, f1
def run_epoch(self, sess, train_data, dev_data, test_data, epoch):
"""
:param train_data: contains concatenated sentence(user and system list type) and ground_labels(O, T, X)
:return: accuracy and f1 scroe
"""
num_batches = (len(train_data) + self.config.batch_size -
1) // self.config.batch_size
prog = Progbar(target=num_batches)
for i, (concat_utter_list, ground_label) in enumerate(
minibatches(train_data + dev_data + test_data[:200],
self.config.batch_size)):
input_features = []
for each_utter_list in concat_utter_list:
user_sentence = each_utter_list[0]
system_sentence = each_utter_list[1]
user_embedding = self.utter_embed.embed_utterance(
user_sentence)
system_embedding = self.utter_embed.embed_utterance(
system_sentence)
input_feature = np.concatenate(
(user_embedding, system_embedding), axis=0)
input_features.append(input_feature)
input_features = np.array([input_features])
ground_label_list = []
for label in ground_label:
# label.strip().encode('utf-8')
ground_label_list.append(self.cate_mapping_dict[label.strip()])
ground_label_list = np.array([ground_label_list])
dropout_keep_prob = 0.8
feed_dict = {
self.input_features: input_features,
self.ground_label: ground_label_list,
self.dropout_keep_prob: dropout_keep_prob
}
# self.merged = tf.summary.merge_all()
self.file_writer = tf.summary.FileWriter(self.config.output_path,
sess.graph)
_, train_loss = sess.run([self.train_step, self.loss],
feed_dict=feed_dict)
prog.update(i + 1, [("train loss", train_loss)])
js_divergence_value, accuracy, precision_X, recall_X, f1_score_X, precision_B_T, recall_B_T, f1_score_B_T = self.run_evaluate(
sess, test_data[200:])
self.logger.info("JS_divergence : {:f}".format(js_divergence_value))
self.logger.info("accuracy : {:f}".format(accuracy))
self.logger.info("precision_X : {:f}".format(precision_X))
self.logger.info("recall_X : {:f}".format(recall_X))
self.logger.info("f1_score_X : {:f}".format(f1_score_X))
self.logger.info("precision X + T : {:f}".format(precision_B_T))
self.logger.info("recall X + T : {:f}".format(recall_B_T))
self.logger.info("f1_score X + T : {:f}".format(f1_score_B_T))
return accuracy, f1_score_X
def do_train(model, config, train_data, dev_data, test_data):
learning_rate_init=0.001 # initial
learning_rate_final=0.0001 # final
learning_rate=learning_rate_init
intermid_epoch = 20 # after this epoch, change learning rate
maximum = 0
session_conf = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)
sess = tf.Session(config=session_conf)
with sess.as_default():
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
if config.restore is not None:
saver.restore(sess, config.restore)
print('model restored')
# summary setting
loss_summary = tf.summary.scalar('loss', model.loss)
acc_summary = tf.summary.scalar('accuracy', model.accuracy)
train_summary_op = tf.summary.merge([loss_summary, acc_summary])
train_summary_dir = os.path.join(config.summary_dir, 'summaries', 'train')
train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph)
dev_summary_op = tf.summary.merge([loss_summary, acc_summary])
dev_summary_dir = os.path.join(config.summary_dir, 'summaries', 'dev')
dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph)
# training steps
for e in range(config.epoch):
# run epoch
idx = 0
nbatches = (len(train_data.sentence_word_ids) + config.batch_size - 1) // config.batch_size
prog = Progbar(target=nbatches)
for ptr in range(0, len(train_data.sentence_word_ids), config.batch_size):
feed_dict={model.input_data_word_ids: train_data.sentence_word_ids[ptr:ptr + config.batch_size],
model.input_data_wordchr_ids: train_data.sentence_wordchr_ids[ptr:ptr + config.batch_size],
model.input_data_pos_ids: train_data.sentence_pos_ids[ptr:ptr + config.batch_size],
model.input_data_etc: train_data.sentence_etc[ptr:ptr + config.batch_size],
model.output_data: train_data.sentence_tag[ptr:ptr + config.batch_size],
model.learning_rate:learning_rate}
step, train_summaries, _, train_loss, train_accuracy = \
sess.run([model.global_step, train_summary_op, model.train_op, model.loss, model.accuracy], feed_dict=feed_dict)
prog.update(idx + 1, [('train loss', train_loss), ('train accuracy', train_accuracy)])
train_summary_writer.add_summary(train_summaries, step)
idx += 1
# evaluate on dev data
feed_dict={model.input_data_word_ids: dev_data.sentence_word_ids,
model.input_data_wordchr_ids: dev_data.sentence_wordchr_ids,
model.input_data_pos_ids: dev_data.sentence_pos_ids,
model.input_data_etc: dev_data.sentence_etc,
model.output_data: dev_data.sentence_tag}
step, dev_summaries, logits, logits_indices, trans_params, output_data_indices, length, dev_loss, dev_accuracy = \
sess.run([model.global_step, dev_summary_op, model.logits, model.logits_indices, model.trans_params, model.output_data_indices, model.length, model.loss, model.accuracy], feed_dict=feed_dict)
print('epoch: %d / %d, step: %d, dev loss: %s, dev accuracy: %s' % (e, config.epoch, step, dev_loss, dev_accuracy))
dev_summary_writer.add_summary(dev_summaries, step)
print('dev precision, recall, f1(token): ')
token_f1 = TokenEval.compute_f1(config.class_size, logits, dev_data.sentence_tag, length)
if config.use_crf:
viterbi_sequences = viterbi_decode(logits, trans_params, length)
tag_preds = dev_data.logits_indices_to_tags_seq(viterbi_sequences, length)
else:
tag_preds = dev_data.logits_indices_to_tags_seq(logits_indices, length)
tag_corrects = dev_data.logits_indices_to_tags_seq(output_data_indices, length)
dev_prec, dev_rec, dev_f1 = ChunkEval.compute_f1(tag_preds, tag_corrects)
print('dev precision, recall, f1(chunk): ', dev_prec, dev_rec, dev_f1)
chunk_f1 = dev_f1
# save best model
'''
m = chunk_f1 # slightly lower than token-based f1 for test
'''
m = token_f1
if m > maximum:
print('new best f1 score!')
maximum = m
save_path = saver.save(sess, config.checkpoint_dir + '/' + 'model_max.ckpt')
print('max model saved in file: %s' % save_path)
feed_dict={model.input_data_word_ids: test_data.sentence_word_ids,
model.input_data_wordchr_ids: test_data.sentence_wordchr_ids,
model.input_data_pos_ids: test_data.sentence_pos_ids,
model.input_data_etc: test_data.sentence_etc,
model.output_data: test_data.sentence_tag}
step, logits, logits_indices, trans_params, output_data_indices, length, test_loss, test_accuracy = \
sess.run([model.global_step, model.logits, model.logits_indices, model.trans_params, model.output_data_indices, model.length, model.loss, model.accuracy], feed_dict=feed_dict)
print('epoch: %d / %d, step: %d, test loss: %s, test accuracy: %s' % (e, config.epoch, step, test_loss, test_accuracy))
print('test precision, recall, f1(token): ')
TokenEval.compute_f1(config.class_size, logits, test_data.sentence_tag, length)
if config.use_crf:
viterbi_sequences = viterbi_decode(logits, trans_params, length)
tag_preds = test_data.logits_indices_to_tags_seq(viterbi_sequences, length)
else:
tag_preds = test_data.logits_indices_to_tags_seq(logits_indices, length)
tag_corrects = test_data.logits_indices_to_tags_seq(output_data_indices, length)
test_prec, test_rec, test_f1 = ChunkEval.compute_f1(tag_preds, tag_corrects)
print('test precision, recall, f1(chunk): ', test_prec, test_rec, test_f1)
# learning rate change
if e > intermid_epoch: learning_rate=learning_rate_final
def run_epoch(session, config, model, data, eval_op, keep_prob, is_training):
n_samples = len(data[0])
print("Running %d samples:" % (n_samples))
minibatches = get_minibatches_idx(n_samples,
config.batch_size,
shuffle=False)
correct = 0.
target_correct = 0.0
total = 0
total_cost = 0
prog = Progbar(target=len(minibatches))
# dummynode_hidden_states_collector=np.array([[0]*config.hidden_size])
# to_print_total = np.zeros([2, config.num_label])
corr = data[5]
for i, inds in enumerate(minibatches):
# print("i:",i,"inds:",inds)
x = [data[0][j] for j in inds]
y = np.array([data[1][j] for j in inds])
multi = [data[2][j] for j in inds]
length = [data[3][j] for j in inds]
rank = [data[4][j] for j in inds]
target = [data[6][j] for j in inds]
ranksignal, prosignal = signalMatrix(multi, rank, length,
config.num_label)
# 计算文本实际长度
text_mask = [len(s) for s in x]
# 对文本做padding
text_padding = data_helper.padding_text(x, text_mask, padding_word=[0])
x = []
for j in range(len(inds)):
x = x + text_padding[j]
config.sentence_size = len(x) / len(inds)
# 计算每个句子长度
sentence_mask = [len(s) for s in x]
# 对句子做padding
sentence_padding = data_helper.padding_text(x,
sentence_mask,
padding_word=0)
x = np.array([np.array(s) for s in sentence_padding])
count, _, cost, to_print, prediction, text_alphas = session.run(
[
model.accuracy, eval_op, model.cost, model.to_print,
model.prediction, model.text_alphas
], {
model.input_data: x,
model.labels: y,
model.sentence_mask: sentence_mask,
model.text_mask: text_mask,
model.keep_prob: keep_prob,
model.signal: prosignal,
model.length: length,
model.corr: corr
})
# to_print = session.run([model.to_print], {model.input_data: x, model.labels: y, model.sentence_mask:sentece_mask.astype(int), model.text_mask:text_mask.astype(int), model.keep_prob: keep_prob})
# target predict
target_predict = np.argmax(text_alphas)
target_count = np.equal(target_predict, target)
target_correct = target_correct + target_count
# if not is_training:
# to_print_total = np.concatenate((to_print_total, to_print), axis=0)
correct += count
total += len(inds)
total_cost += cost
prog.update(i + 1, [("train loss", cost)])
# print(correct / total)
print("Total loss:")
print(total_cost)
accuracy = correct / total
target_accuracy = target_correct / total
return accuracy, target_accuracy
def train():
print("Training")
# tf Graph input
x = tf.placeholder(dtype=tf.float32, shape=[None, config.input_window_size - 1, config.input_size], name="input_sequence")
y = tf.placeholder(dtype=tf.float32, shape=[None, config.output_window_size, config.input_size], name="raw_labels")
dec_in = tf.placeholder(dtype=tf.float32, shape=[None, config.output_window_size, config.input_size], name="decoder_input")
labels = tf.transpose(y, [1, 0, 2])
labels = tf.reshape(labels, [-1, config.input_size])
labels = tf.split(labels, config.output_window_size, axis=0, name='labels')
# Define model
prediction = models.seq2seq(x, dec_in, config, True)
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.6
sess_config.allow_soft_placement = True
sess_config.log_device_placement = False
sess = tf.Session(config=sess_config)
# Define cost function
loss = eval('loss_functions.' + config.loss + '_loss(prediction, labels, config)')
# Add a summary for the loss
train_loss = tf.summary.scalar('train_loss', loss)
valid_loss = tf.summary.scalar('valid_loss', loss)
# Defining training parameters
optimizer = tf.train.AdamOptimizer(config.learning_rate)
global_step = tf.Variable(0, name='global_step', trainable=False)
# Gradient Clipping
grads = tf.gradients(loss, tf.trainable_variables())
grads, _ = tf.clip_by_global_norm(grads, config.max_grad_norm)
optimizer.apply_gradients(zip(grads, tf.trainable_variables()))
train_op = optimizer.minimize(loss, global_step=global_step)
saver = tf.train.Saver(max_to_keep=10)
train_writer = tf.summary.FileWriter("./log", sess.graph)
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
# Obtain total training parameters
total_parameters = 0
for variable in tf.trainable_variables():
# shape is an array of tf.Dimension
shape = variable.get_shape()
variable_parameters = 1
for dim in shape:
variable_parameters *= dim.value
total_parameters += variable_parameters
print('Total training parameters: ' + str(total_parameters))
if not(os.path.exists(checkpoint_dir)):
os.makedirs(checkpoint_dir)
saved_epoch = 0
train_size = config.training_size
valid_size = config.validation_size
best_val_loss = float('inf')
if config.restore & os.path.exists(checkpoint_dir+'checkpoint'):
with open(checkpoint_dir + 'checkpoint') as f:
content = f.readlines()
saved_epoch = int(re.search(r'\d+', content[0]).group())
model_name = checkpoint_dir + "Epoch_" + str(saved_epoch)
saver.restore(sess, model_name)
v_loss_mean = 0.0
for i in range(valid_size):
batch_x, batch_dec_in, batch_y = data_utils.get_batch(config, test_set)
v_loss, valid_summary = sess.run([loss, valid_loss], feed_dict={x: batch_x, y: batch_y, dec_in: batch_dec_in})
v_loss_mean = v_loss_mean*i/(i+1) + v_loss/(i+1)
best_val_loss = v_loss_mean
print("Restored session from Epoch ", str(saved_epoch))
print("Best Validation Loss: ", best_val_loss, "\n")
print("________________________________________________________________")
best_val_epoch = saved_epoch
for j in range(saved_epoch, config.max_epoch):
print("Epoch ", j+1)
prog = Progbar(target=train_size)
prog_valid = Progbar(target=valid_size)
for i in range(train_size):
batch_x, batch_dec_in, batch_y = data_utils.get_batch(config, train_set)
current_cost, train_summary, _ = sess.run([loss, train_loss, train_op], feed_dict={x: batch_x, y: batch_y, dec_in: batch_dec_in})
train_writer.add_summary(train_summary, j*train_size+i)
prog.update(i+1, [("Training Loss", current_cost)])
v_loss_mean = 0.0
for i in range(valid_size):
batch_x, batch_dec_in, batch_y = data_utils.get_batch(config, test_set)
v_loss, valid_summary = sess.run([loss, valid_loss], feed_dict={x: batch_x, y: batch_y, dec_in: batch_dec_in})
v_loss_mean = v_loss_mean*i/(i+1) + v_loss/(i+1)
prog_valid.update(i + 1, [("Validation Loss", v_loss)])
train_writer.add_summary(valid_summary, j*valid_size+i)
if v_loss_mean < best_val_loss:
model_name = checkpoint_dir + "Epoch_" + str(j+1)
best_val_loss = v_loss_mean
best_val_epoch = j+1
saver.save(sess, model_name)
print("Current Best Epoch: ", best_val_epoch, ", Best Validation Loss: ", best_val_loss, "\n")
if j+1 - best_val_epoch > config.early_stop:
break
def run_epoch(self, sess, train, dev, vocab_words, vocab_poss,
vocab_chunks, vocab_aspect_tags, vocab_polarity_tags,
vocab_joint_tags, epoch):
"""
Performs one complete pass over the train set and evaluate on dev
Args:
sess: tensorflow session
train: dataset that yields tuple of sentences, tags
dev: dataset
vocab_aspect_tags: {tag: index} dictionary
epoch: (int) number of the epoch
"""
self.config.istrain = True # set to train first, #batch normalization#
losses = []
nbatches = (len(train) + self.config.batch_size -
1) / self.config.batch_size
prog = Progbar(target=nbatches)
for i, (words, poss, chunks, labels_aspect, labels_polarity,
labels_joint) in enumerate(
minibatches_for_sequence(train, self.config.batch_size)):
fd, sequence_lengths = self.get_feed_dict(
words,
poss,
chunks,
labels_aspect,
labels_polarity,
labels_joint,
self.config.lr,
self.config.dropout,
vocab_aspect_tags=vocab_aspect_tags)
_, lr, train_loss, summary = sess.run(
[self.train_op, self.learning_rate, self.loss, self.merged],
feed_dict=fd)
losses.append(train_loss)
if self.config.show_process_logs:
print_mess = [("train loss", train_loss)]
if self.config.use_labels_length:
label_length_loss = sess.run(self.label_length_loss,
feed_dict=fd)
print_mess.append(("label_length_loss", label_length_loss))
if self.config.use_mpqa:
mpqa_loss = sess.run(self.mpqa_loss, feed_dict=fd)
print_mess.append(("mpqa_loss", mpqa_loss))
print_mess.append(("lr", lr))
prog.update(i + 1, print_mess)
# tensorboard
if i % 2 == 0:
self.file_writer.add_summary(summary, epoch * nbatches + i)
if self.config.data_sets.startswith("twitter"):
aspect_p, aspect_r, aspect_f1, aspect_test_acc, \
polarity_p, polarity_r, polarity_f1, polarity_test_acc, dev_loss = 0, 0, 0, 0, 0, 0, 0, 0, 0
self.logger.info("Ignore validating without corresponding data~")
else:
aspect_p, aspect_r, aspect_f1, aspect_test_acc, \
polarity_p, polarity_r, polarity_f1, polarity_test_acc, dev_loss = self.run_evaluate(
sess, dev, vocab_aspect_tags, vocab_polarity_tags, vocab_joint_tags, vocab_words, is_dev=True)
self.logger.info(
"- aspect_dev precision {:04.2f} - aspect_dev recall {:04.2f} - "
"aspect_dev f1 {:04.2f} - aspect_dev acc {:04.2f}".format(
100 * aspect_p, 100 * aspect_r, 100 * aspect_f1,
100 * aspect_test_acc))
self.logger.info(
"- polarity_dev precision {:04.2f} - polarity_dev recall {:04.2f} - "
"polarity_dev f1 {:04.2f} - polarity_dev acc {:04.2f}".format(
100 * polarity_p, 100 * polarity_r, 100 * polarity_f1,
100 * polarity_test_acc))
return aspect_p, aspect_r, aspect_f1, sum(losses) / len(
losses), polarity_p, polarity_r, polarity_f1, dev_loss
def train_stepwise(self, train, dev, train_eval):
"""Performs training with early stopping and lr exponential decay
Args:
train: dataset that yields tuple of (sentences, tags)
dev: dataset
"""
best_score = 0
nepoch_no_imprv = 0
updates = 0
epoch_train_loss = 0
prog = Progbar(target=self.config.updates_per_epoch)
while 1:
for words, labels in minibatches(train, self.config.batch_size):
fd, _ = self.get_feed_dict(True,
words,
labels,
lr=self.config.lr)
_, train_loss = self.sess.run([self.train_op, self.loss],
feed_dict=fd)
prog.update((updates % self.config.updates_per_epoch) + 1,
values=[("train loss", train_loss)])
epoch_train_loss += train_loss
updates += 1
if updates % self.config.updates_per_epoch == 0:
acc_train = self.evaluate(train_eval)
acc_test = self.evaluate(dev)
epoch = updates / self.config.updates_per_epoch
prog.update(self.config.updates_per_epoch,
epoch, [("train loss", train_loss)],
exact=[("dev acc", acc_test),
("train acc", acc_train),
("lr", self.config.lr)])
self.write_epoch_results(
epoch, acc_train, acc_test,
epoch_train_loss / self.config.updates_per_epoch)
epoch_train_loss = 0
# early stopping and saving best parameters
if acc_test >= best_score:
nepoch_no_imprv = 0
self.save_session()
best_score = acc_test
else:
nepoch_no_imprv += 1
if nepoch_no_imprv >= self.config.nepoch_no_imprv:
return best_score
# apply decay
if self.config.lr_decay_strategy == "on-no-improvement":
if acc_test < best_score:
self.config.lr *= self.config.lr_decay
elif self.config.lr_decay_strategy == "exponential":
self.config.lr *= self.config.lr_decay
elif self.config.lr_decay_strategy == "step":
self.config.lr = self.config.step_decay_init_lr * math.pow(
self.config.step_decay_drop,
math.floor(
epoch / self.config.step_decay_epochs_drop))
elif self.config.lr_decay_strategy is None:
pass
else:
raise ValueError("Invalid 'decay_strategy' setting: " +
self.config.lr_decay_strategy)
if updates < self.config.max_updates:
prog = Progbar(target=self.config.updates_per_epoch)
if updates >= self.config.max_updates:
return best_score
def train_epochwise(self, train, dev, train_eval):
"""Performs training with early stopping and lr decay"""
updates, epoch, best_score, nepoch_no_imprv = 0, 0, 0, 0
batch_size = self.config.batch_size
max_epochs = self.config.max_epochs
nbatches = (len(train) + batch_size - 1) // batch_size
while epoch < max_epochs:
# Run one epoch
epoch_time = time()
train_time = time()
epoch_train_loss = 0
iter = 0
prog = Progbar(target=nbatches)
for feed_dict in self.iter_prebuilt_feed_dict_batches(
train, batch_size):
fd, _ = self.get_final_feed_dict(True,
feed_dict,
lr=self.config.lr)
_, train_loss = self.sess.run([self.train_op, self.loss],
feed_dict=fd)
epoch_train_loss += train_loss
updates += 1
if updates % self.config.lr_decay_step == 0:
# apply decay
if self.config.lr_decay_strategy == "on-no-improvement":
if acc_test < best_score:
self.config.lr *= self.config.lr_decay
elif self.config.lr_decay_strategy == "exponential":
self.config.lr *= self.config.lr_decay
elif self.config.lr_decay_strategy == "step":
self.config.lr = self.config.step_decay_init_lr * \
math.pow(self.config.step_decay_drop, math.floor(
(epoch) / self.config.step_decay_epochs_drop))
elif self.config.lr_decay_strategy is None:
pass
else:
raise ValueError("Invalid 'decay_strategy' setting: " +
self.config.lr_decay_strategy)
prog.update(iter + 1, values=[("train loss", train_loss)])
iter += 1
train_time = time() - train_time
# evaluate epoch
acc_train = self.evaluate(train_eval)
eval_time = time()
acc_test = self.evaluate(dev)
eval_time = time() - eval_time
epoch_time = time() - epoch_time
# log epoch
prog.update(iter + 1,
epoch, [("train loss", train_loss)],
exact=[("dev acc", acc_test), ("train acc", acc_train),
("lr", self.config.lr)])
self.write_epoch_results(epoch,
acc_train,
acc_test,
epoch_train_loss / iter,
nbatches,
epoch_time=epoch_time,
train_time=train_time,
eval_time=eval_time)
# early stopping and saving checkpoint
if acc_test >= best_score:
nepoch_no_imprv = 0
self.save_session()
best_score = acc_test
else:
nepoch_no_imprv += 1
if nepoch_no_imprv >= self.config.nepoch_no_imprv:
self.logger.info(
"- early stopping {} epochs without improvement".
format(nepoch_no_imprv))
break
epoch += 1
return best_score