def create_maps(sentences):
"""
建立字和标签的字典
"""
if config.pre_emb:
""" 首先利用训练集建立字典 """
dico_chars_train, _, _ = char_mapping(sentences, config.lower)
""" 预训练字向量中的字，如果不在上面的字典中，则加入 """
dico_chars, char_to_id, id_to_char = augment_with_pretrained(dico_chars_train.copy(),
config.emb_file)
else:
""" 只利用训练集建立字典 """
_, char_to_id, id_to_char = char_mapping(sentences, config.lower)
""" 利用训练集建立标签字典 """
_, tag_to_id, id_to_tag = tag_mapping(sentences)
with open(config.map_file, "wb") as f:
pickle.dump([char_to_id, id_to_char, tag_to_id, id_to_tag], f)
return char_to_id, id_to_char, tag_to_id, id_to_tag
def train():
#load data
train_sentences = load_sentences(FLAGS.train_file, FLAGS.lower, FLAGS.zeros)
dev_sentences = load_sentences(FLAGS.dev_file, FLAGS.lower, FLAGS.zeros)
test_sentences = load_sentences(FLAGS.test_file, FLAGS.lower, FLAGS.zeros)
if not os.path.isfile(FLAGS.map_file):
if FLAGS.pre_emb:
dico_chars_train = char_mapping(train_sentences, FLAGS.lower)[0]
dico_chars, char_to_id, id_to_char = augment_with_pretrained(
dico_chars_train.copy(),
FLAGS.emb_file,
list(itertools.chain.from_iterable(
[[w[0] for w in s] for s in test_sentences])
)
)
train_sentences = load_sentences(FLAGS.train_file, FLAGS.lower, FLAGS.zeros)
dev_sentences = load_sentences(FLAGS.dev_file, FLAGS.lower, FLAGS.zeros)
test_sentences = load_sentences(FLAGS.test_file, FLAGS.lower, FLAGS.zeros)
if not os.path.isfile(FLAGS.map_file):
if FLAGS.pre_emb:
dico_chars_train = char_mapping(train_sentences, FLAGS.lower)[0]
dico_chars, char_to_id, id_to_char = augment_with_pretrained(
dico_chars_train.copy(),
FLAGS.emb_file,
list(itertools.chain.from_iterable(
[[w[0] for w in s] for s in test_sentences])
)
)
else:
_c, char_to_id, id_to_char = char_mapping(train_sentences, FLAGS.lower)
# Create a dictionary and a mapping for tags
_t, tag_to_id, id_to_tag = tag_mapping(train_sentences)
with open(FLAGS.map_file, "wb") as f:
pickle.dump([char_to_id, id_to_char, tag_to_id, id_to_tag], f)
else:
with open(FLAGS.map_file, "rb") as f:
char_to_id, id_to_char, tag_to_id, id_to_tag = pickle.load(f)
train_data = prepare_dataset(
train_sentences, char_to_id, tag_to_id, FLAGS.lower, FLAGS.self_seg
)
dev_data = prepare_dataset(
dev_sentences, char_to_id, tag_to_id, FLAGS.lower, FLAGS.self_seg
)
def train():
# load data sets
train_sentences=load_sentences(FLAGS.train_file,FLAGS.zeros)
dev_sentences=load_sentences(FLAGS.dev_file,FLAGS.zeros)
test_sentences=load_sentences(FLAGS.test_file,FLAGS.zeros)
# appoint tagging scheme (IOB/IOBES)
train_sentences=update_tag_scheme(train_sentences,FLAGS.tag_schema)
dev_sentences=update_tag_scheme(dev_sentences,FLAGS.tag_schema)
test_sentences=update_tag_scheme(test_sentences,FLAGS.tag_schema)
#create maps if not exist
if not os.path.exists(FLAGS.map_file):
if FLAGS.pre_emb:
char_to_id,_=char_mapping(train_sentences)
char_to_id,id_to_char=augment_with_pretrained(char_to_id,'wiki_100.utf8')
else:
char_to_id, id_to_char=char_mapping(train_sentences)
tag_to_id, id_to_tag=tag_mapping(train_sentences)
with open(FLAGS.map_file,'wb') as f:
cPickle.dump([char_to_id,id_to_char,tag_to_id,id_to_tag],f,cPickle.HIGHEST_PROTOCOL)
else:
with open(FLAGS.map_file,'rb') as f:
char_to_id, id_to_char, tag_to_id, id_to_tag=cPickle.load(f)
# prepare data, get a collection of list containing index
train_data=prepare_dataset(train_sentences,char_to_id,tag_to_id,True)
dev_data=prepare_dataset(dev_sentences,char_to_id,tag_to_id,True)
test_data=prepare_dataset(test_sentences,char_to_id,tag_to_id,True)
print "%i %i %i sentences in train / dev / test." % (len(train_data),len(dev_data),len(test_data))
if not FLAGS.pre_emb:
pre_emb=None
else:
pre_emb=load_word2vec(FLAGS.pre_emb_file,char_to_id,FLAGS.char_dim)
print "init embedding shape: (%d,%d)" %(pre_emb.shape[0],pre_emb.shape[1])
train_manager=BatchManager(train_data,FLAGS.batch_size,True)
dev_manager=BatchManager(dev_data,FLAGS.batch_size,False)
test_manager=BatchManager(test_data,FLAGS.batch_size,False)
config=BasicModelConfig(FLAGS,len(char_to_id),len(tag_to_id),4)
tfConfig = tf.ConfigProto()
tfConfig.gpu_options.per_process_gpu_memory_fraction = FLAGS.memory_usage
with tf.Session(config=tfConfig) as sess:
print "Train started!"
model=BasicModel(config,pre_emb)
saver=tf.train.Saver()
# tensorboard
if not os.path.exists(FLAGS.summaries_dir):
os.mkdir(FLAGS.summaries_dir)
merged=tf.summary.merge_all()
train_writer=tf.summary.FileWriter(os.path.join(FLAGS.summaries_dir,FLAGS.model_name,"train"),sess.graph)
test_writer=tf.summary.FileWriter(os.path.join(FLAGS.summaries_dir,FLAGS.model_name,"test"),sess.graph)
# load previous trained model or create a new model
if not os.path.exists(FLAGS.checkpoints):
os.mkdir(FLAGS.checkpoints)
model_name=os.path.join(FLAGS.checkpoints,FLAGS.model_name)
ckpt=tf.train.get_checkpoint_state(FLAGS.checkpoints)
if ckpt and ckpt.model_checkpoint_path:
print "restore from previous traied model: %s" % FLAGS.model_name
saver.restore(sess,ckpt.model_checkpoint_path)
else:
sess.run(tf.global_variables_initializer())
def evaluate(sess,model,manager):
strings=[]
predicts=[]
goldens=[]
bar = ProgressBar(max_value=manager.num_batch)
for batch in bar(manager.iter_batch()):
batch_string,batch_predict,batch_golden=model.evaluate_step(sess,batch)
strings.extend(batch_string)
predicts.extend(batch_predict)
goldens.extend(batch_golden)
return strings,predicts,goldens
best_eval_f1=0
noimpro_num=0
for i in range(FLAGS.max_epoch):
#train
train_loss=[]
bar = ProgressBar(max_value=train_manager.num_batch)
for step,batch in bar(enumerate(train_manager.iter_batch())):
batch.append(merged)
summary,global_step,batch_loss=model.train_step(sess,batch,FLAGS.dropout_keep)
#add summary to tensorboard
train_writer.add_summary(summary,global_step)
train_loss.append(batch_loss)
print "Epoch %d Train loss is %.4f" % (i+1,np.mean(train_loss))
#dev
strings,predicts,goldens=evaluate(sess,model,dev_manager)
eval_f1=report_results(strings,predicts,goldens,id_to_char,id_to_tag,'outputs/dev')
if eval_f1>best_eval_f1:
best_eval_f1=eval_f1
noimpro_num=0
saver.save(sess,model_name)
else:
noimpro_num+=1
print "Epoch %d Best eval f1:%.6f" % (i+1,best_eval_f1)
#test
strings,predicts,goldens=evaluate(sess,model,test_manager)
test_f1=report_results(strings,predicts,goldens,id_to_char,id_to_tag,'outputs/test',True)
#early_stop
if noimpro_num>=3:
print "Early stop! Final F1 scores on test data is :%.6f" % test_f1
break
print
def train():
# 加载数据集
train_sentences = load_sentences(FLAGS.train_file, FLAGS.lower,
FLAGS.zeros)
dev_sentences = load_sentences(FLAGS.dev_file, FLAGS.lower, FLAGS.zeros)
test_sentences = load_sentences(FLAGS.test_file, FLAGS.lower, FLAGS.zeros)
# 选择tag形式 (IOB / IOBES) 默认使用IOBES
update_tag_scheme(train_sentences, FLAGS.tag_schema)
update_tag_scheme(test_sentences, FLAGS.tag_schema)
if not os.path.isfile(FLAGS.map_file):
if FLAGS.pre_emb:
dico_chars_train = char_mapping(train_sentences, FLAGS.lower)[0]
dico_chars, char_to_id, id_to_char = augment_with_pretrained(
dico_chars_train.copy(), FLAGS.emb_file,
list(
itertools.chain.from_iterable([[w[0] for w in s]
for s in test_sentences])))
else:
_c, char_to_id, id_to_char = char_mapping(train_sentences,
FLAGS.lower)
# Create a dictionary and a mapping for tags
_t, tag_to_id, id_to_tag = tag_mapping(train_sentences)
with open(FLAGS.map_file, "wb") as f:
pickle.dump([char_to_id, id_to_char, tag_to_id, id_to_tag], f)
else:
with open(FLAGS.map_file, "rb") as f:
# {'S-LOC': 10, 'E-LOC': 3, 'B-ORG': 4, 'S-PER': 11, 'S-ORG': 12, 'O': 0,
# 'E-ORG': 5, 'I-LOC': 6, 'I-PER': 7, 'I-ORG': 1, 'B-PER': 8, 'B-LOC': 2, 'E-PER': 9}
char_to_id, id_to_char, tag_to_id, id_to_tag = pickle.load(f)
# 转化成数字化的数据
train_data = prepare_dataset(train_sentences, char_to_id, tag_to_id,
FLAGS.lower)
dev_data = prepare_dataset(dev_sentences, char_to_id, tag_to_id,
FLAGS.lower)
test_data = prepare_dataset(test_sentences, char_to_id, tag_to_id,
FLAGS.lower)
print("%i / %i / %i sentences in train / dev / test." %
(len(train_data), len(dev_data), len(test_data)))
#长度不足补0
train_manager = BatchManager(train_data, FLAGS.batch_size)
dev_manager = BatchManager(dev_data, 100)
test_manager = BatchManager(test_data, 100)
make_path(FLAGS)
if os.path.isfile(FLAGS.config_file):
config = load_config(FLAGS.config_file)
else:
config = config_model(char_to_id, tag_to_id)
save_config(config, FLAGS.config_file)
make_path(FLAGS)
log_path = os.path.join("log", FLAGS.log_file)
logger = get_logger(log_path)
print_config(config, logger)
# GPU设置
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
steps_per_epoch = train_manager.len_data
with tf.Session(config=tf_config) as sess:
model = create_model(sess, Model, FLAGS.ckpt_path, load_word2vec,
config, id_to_char, logger)
logger.info("start training")
loss = []
for i in range(100):
for batch in train_manager.iter_batch(shuffle=True):
step, batch_loss = model.run_step(sess, True, batch)
loss.append(batch_loss)
if step % FLAGS.steps_check == 0:
iteration = step // steps_per_epoch + 1
logger.info("iteration:{} step:{}/{}, "
"NER loss:{:>9.6f}".format(
iteration, step % steps_per_epoch,
steps_per_epoch, np.mean(loss)))
# 每100次算一次平均loss
loss = []
best = evaluate(sess, model, "dev", dev_manager, id_to_tag, logger)
if best:
save_model(sess, model, FLAGS.ckpt_path, logger)
evaluate(sess, model, "test", test_manager, id_to_tag, logger)