def create_instance():
tag_to_id = FLAGS.tag_to_id
id_to_tag = {v: k for k, v in tag_to_id.items()}
# 字典生成
print "dict building......"
if not isExists(FLAGS.dict_file):
print "build dict starting..."
train_file = read_conll_file(FLAGS.train_file)
word_to_id, _ = word_mapping(train_file, FLAGS.min_freq)
write_file(word_to_id, FLAGS.dict_file)
else:
print "build dict from pickle..."
word_to_id = load_dict(FLAGS.dict_file)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.3)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
with sess.as_default():
print "begin for create model..."
model = create_model(sess, word_to_id, id_to_tag) # just struct
# load model
model.logger.info("testing ner")
ckpt = tf.train.get_checkpoint_state(FLAGS.model_path)
model.logger.info("Reading model parameters from %s" % ckpt.model_checkpoint_path)
model.saver.restore(sess, ckpt.model_checkpoint_path)
return word_to_id, tag_to_id, id_to_tag, sess, model
def train():
# 加载数据集
train_sentences = dl.load_sentences(FLAGS.train_file)
dev_sentences = dl.load_sentences(FLAGS.dev_file)
test_sentences = dl.load_sentences(FLAGS.test_file)
# 转换编码 bio转bioes
dl.update_tag_scheme(train_sentences, FLAGS.tag_schema)
dl.update_tag_scheme(test_sentences, FLAGS.tag_schema)
dl.update_tag_scheme(dev_sentences, FLAGS.tag_schema)
# 创建单词映射及标签映射
if not os.path.isfile(FLAGS.map_file):
_, word_to_id, id_to_word = dl.word_mapping(train_sentences)
_, tag_to_id, id_to_tag = dl.tag_mapping(train_sentences)
with open(FLAGS.map_file, 'wb') as f:
pickle.dump([word_to_id, id_to_word, tag_to_id, id_to_tag], f)
else:
with open(FLAGS.map_file, 'rb') as f:
unpickler = pickle.Unpickler(f)
scores = unpickler.load()
word_to_id, id_to_word, tag_to_id, id_to_tag = scores
train_data = dl.prapare_dataset(train_sentences, word_to_id, tag_to_id)
dev_data = dl.prapare_dataset(train_sentences, word_to_id, tag_to_id)
test_data = dl.prapare_dataset(train_sentences, word_to_id, tag_to_id)
print('train_data %i, dev_data_num %i, test_data_num %i' %
(len(train_data), len(dev_data), len(test_data)))
mu.make_path(FLAGS)
if os.path.isfile(FLAGS.config_file):
config = mu.load_config(FLAGS.config_file)
else:
config = mu.config_model(FLAGS, word_to_id, tag_to_id)
mu.save_config(config, FLAGS.config_file)
log_path = os.path.join('log', FLAGS.log_file)
logger = mu.get_log(log_path)
mu.print_config(config, logger)
print('aa')
def train():
# 1、加载数据集
train_sentences = data_loader.load_sentences(FLAGS.train_file)
dev_sentences = data_loader.load_sentences(FLAGS.dev_file)
test_sentences = data_loader.load_sentences(FLAGS.test_file)
# 2、转换编码 BIO->BIOES
data_loader.update_tag_scheme(train_sentences, FLAGS.tag_schema)
data_loader.update_tag_scheme(dev_sentences, FLAGS.tag_schema)
data_loader.update_tag_scheme(test_sentences, FLAGS.tag_schema)
# 3、创建单词映射与标签映射
if not os.path.isfile(FLAGS.map_file):
_, word_to_id, id_to_word = data_loader.word_mapping(train_sentences)
_, tag_to_id, id_to_tag = data_loader.tag_mapping(train_sentences)
with open(FLAGS.map_file, "wb") as f:
# 序列化pickle.dump(obj, file, [,protocol]),,序列化对象，将对象obj保存到文件file中去。
pickle.dump([word_to_id, id_to_word, tag_to_id, id_to_tag], f)
else:
# 反序列化对象，将文件中的数据解析为一个python对象。file中有read()接口和readline()接口
with open(FLAGS.map_file, "rb") as f:
word_to_id, id_to_word, tag_to_id, id_to_tag = pickle.load(f)
# 4、数据预处理
train_data = data_loader.prepare_dataset(train_sentences, word_to_id,
tag_to_id)
dev_data = data_loader.prepare_dataset(dev_sentences, word_to_id,
tag_to_id)
test_data = data_loader.prepare_dataset(test_sentences, word_to_id,
tag_to_id)
model_utils.make_path(FLAGS)
config = model_utils.config_model(FLAGS, word_to_id, tag_to_id)
def train():
# 加载数据集
train_sentences = data_loader.load_sentences(FLAGS.train_file)
dev_sentences = data_loader.load_sentences(FLAGS.dev_file)
test_sentences = data_loader.load_sentences(FLAGS.test_file)
# 转换编码
data_loader.update_tag_scheme(train_sentences, FLAGS.tag_schema)
data_loader.update_tag_scheme(dev_sentences, FLAGS.tag_schema)
data_loader.update_tag_scheme(test_sentences, FLAGS.tag_schema)
# 创建单词和词典映射
if not os.path.isfile(FLAGS.map_file):
if FLAGS.pre_emb:
dico_words_train = data_loader.word_mapping(train_sentences)[0]
dico_word, word_to_id, id_to_word = data_utils.augment_with_pretrained(
dico_words_train.copy(), FLAGS.emb_file,
list(
itertools.chain.from_iterable([[w[0] for w in s]
for s in test_sentences])))
else:
_, word_to_id, id_to_word = data_loader.word_mapping(
train_sentences)
_, tag_to_id, id_to_tag = data_loader.tag_mapping(train_sentences)
with open(FLAGS.map_file, 'wb') as f:
pickle.dump([word_to_id, id_to_word, tag_to_id, id_to_tag], f)
else:
with open(FLAGS.map_file, 'rb') as f:
word_to_id, id_to_word, tag_to_id, id_to_tag = pickle.load(f)
# 准备数据
train_data = data_loader.prepare_dataset(train_sentences, word_to_id,
tag_to_id)
dev_data = data_loader.prepare_dataset(dev_sentences, word_to_id,
tag_to_id)
test_data = data_loader.prepare_dataset(test_sentences, word_to_id,
tag_to_id)
# 将数据分批处理
train_manager = data_utils.BatchManager(train_data, FLAGS.batch_size)
dev_manager = data_utils.BatchManager(dev_data, FLAGS.batch_size)
test_manager = data_utils.BatchManager(test_data, FLAGS.batch_size)
# 创建不存在的文件夹
model_utils.make_path(FLAGS)
# 判断配置文件
if os.path.isfile(FLAGS.config_file):
config = model_utils.load_config(FLAGS.config_file)
else:
config = model_utils.config_model(FLAGS, word_to_id, tag_to_id)
model_utils.save_config(config, FLAGS.config_file)
# 配置印logger
log_path = os.path.join('log', FLAGS.log_file)
logger = model_utils.get_logger(log_path)
model_utils.print_config(config, logger)
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
step_per_epoch = train_manager.len_data
with tf.Session(config=tf_config) as sess:
model = model_utils.create(sess, Model, FLAGS.ckpt_path, load_word2vec,
config, id_to_word, logger)
logger.info('开始训练')
loss = []
start = time.time()
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.setps_chech == 0:
iteration = step // step_per_epoch + 1
logger.info(
"iteration{}: step{}/{}, NER loss:{:>9.6f}".format(
iteration, step % step_per_epoch, step_per_epoch,
np.mean(loss)))
loss = []
best = evaluate(sess, model, 'dev', dev_manager, id_to_tag, logger)
if best:
model_utils.save_model(sess, model, FLAGS.ckpt_path, logger)
evaluate(sess, model, 'test', test_manager, id_to_tag, logger)
t = time.time() - start
logger.info('cost time: %f' % t)
data_loader.update_tag_scheme(train_sentences, FLAGS.tag_schema) data_loader.update_tag_scheme(test_sentences, FLAGS.tag_schema) data_loader.update_tag_scheme(dev_sentences, FLAGS.tag_schema) # 创建单词映射及标签映射 if not os.path.isfile(FLAGS.map_file): if FLAGS.pre_emb: dico_words_train = data_loader.word_mapping(train_sentences)[0] dico_word, word_to_id, id_to_word = data_utils.augment_with_pretrained( dico_words_train.copy(), FLAGS.emb_file, list( itertools.chain.from_iterable( [[w[0] for w in s] for s in test_sentences] )
def train():
# 加载数据集
train_sentences = data_loader.load_sentences(FLAGS.train_file)
dev_sentences = data_loader.load_sentences(FLAGS.dev_file)
test_sentences = data_loader.load_sentences(FLAGS.test_file)
# 转换编码 bio转bioes
data_loader.update_tag_scheme(train_sentences, FLAGS.tag_schema)
data_loader.update_tag_scheme(test_sentences, FLAGS.tag_schema)
data_loader.update_tag_scheme(dev_sentences, FLAGS.tag_schema)
# 创建单词映射及标签映射
if not os.path.isfile(FLAGS.map_file):
if FLAGS.pre_emb:
dico_words_train = data_loader.word_mapping(train_sentences)[0]
dico_word, word_to_id, id_to_word = data_utils.augment_with_pretrained(
dico_words_train.copy(),
FLAGS.emb_file,
list(
itertools.chain.from_iterable(
[[w[0] for w in s] for s in test_sentences]
)
)
)
else:
_, word_to_id, id_to_word = data_loader.word_mapping(train_sentences)
_, tag_to_id, id_to_tag = data_loader.tag_mapping(train_sentences)
with open(FLAGS.map_file, "wb") as f:
pickle.dump([word_to_id, id_to_word, tag_to_id, id_to_tag], f)
else:
with open(FLAGS.map_file, 'rb') as f:
word_to_id, id_to_word, tag_to_id, id_to_tag = pickle.load(f)
train_data = data_loader.prepare_dataset(
train_sentences, word_to_id, tag_to_id
)
dev_data = data_loader.prepare_dataset(
dev_sentences, word_to_id, tag_to_id
)
test_data = data_loader.prepare_dataset(
test_sentences, word_to_id, tag_to_id
)
train_manager = data_utils.BatchManager(train_data, FLAGS.batch_size)
dev_manager = data_utils.BatchManager(dev_data, FLAGS.batch_size)
test_manager = data_utils.BatchManager(test_data, FLAGS.batch_size)
print('train_data_num %i, dev_data_num %i, test_data_num %i' % (len(train_data), len(dev_data), len(test_data)))
model_utils.make_path(FLAGS)
if os.path.isfile(FLAGS.config_file):
config = model_utils.load_config(FLAGS.config_file)
else:
config = model_utils.config_model(FLAGS, word_to_id, tag_to_id)
model_utils.save_config(config, FLAGS.config_file)
log_path = os.path.join("log", FLAGS.log_file)
logger = model_utils.get_logger(log_path)
model_utils.print_config(config, logger)
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 = model_utils.create(sess, Model, FLAGS.ckpt_path, load_word2vec, config, id_to_word, logger)
logger.info("开始训练")
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.setps_chech== 0:
iterstion = step // steps_per_epoch + 1
logger.info("iteration:{} step{}/{},NER loss:{:>9.6f}".format(iterstion, step%steps_per_epoch, steps_per_epoch, np.mean(loss)))
loss = []
best = evaluate(sess,model,"dev", dev_manager, id_to_tag, logger)
if best:
model_utils.save_model(sess, model, FLAGS.ckpt_path, logger)
evaluate(sess, model, "test", test_manager, id_to_tag, logger)