def train():
train_iter, val_iter, test_iter = get_data_iter()
model = SequenceLabelingModel(args, logger).cuda()
optimizer, scheduler = get_optimizer_scheduler(model)
early_stopping_criterion = EarlyStoppingCriterion(patience=args.early_stopping_patience)
logger.info('Start training')
for epoch in range(args.max_epochs):
cur_lr = get_lr(optimizer)
logger.info('Epoch %d, lr %.6f' % (epoch, cur_lr))
model.train()
train_score = []
batch_num = len(train_iter)
cur_num = 0
train_iter.init_epoch()
progress = tqdm(train_iter, mininterval=2, leave=False, file=sys.stdout)
for i, batch in enumerate(progress):
optimizer.zero_grad()
batch_score = model.forward(batch)
train_score.append(batch_score.item())
cur_num += batch.batch_size
progress.set_description(desc='%d/%d, train loss %.4f' % (i, batch_num, sum(train_score) / cur_num))
batch_score.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5.0)
optimizer.step()
val_score = evaluate(model, val_iter, 'val')
test_score = evaluate(model, test_iter, 'test')
if not early_stopping_criterion.step(val_score):
break
scheduler.step(val_score)
def main():
# 加载配置文件
with open('./config.yml') as file_config:
config = yaml.load(file_config)
feature_names = config['model_params']['feature_names']
# 初始化embedding shape, dropouts, 预训练的embedding也在这里初始化)
feature_weight_shape_dict, feature_weight_dropout_dict, \
feature_init_weight_dict = dict(), dict(), dict()
for feature_name in feature_names:
feature_weight_shape_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['shape']
feature_weight_dropout_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['dropout_rate']
path_pre_train = config['model_params']['embed_params'][feature_name]['path']
if path_pre_train:
with open(path_pre_train, 'rb') as file_r:
feature_init_weight_dict[feature_name] = pickle.load(file_r)
# 加载数据
# 加载vocs
path_vocs = []
for feature_name in feature_names:
path_vocs.append(config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
# 加载训练数据
sep_str = config['data_params']['sep']
assert sep_str in ['table', 'space']
sep = '\t' if sep_str == 'table' else ' '
data_dict = init_data(
path=config['data_params']['path_train'], feature_names=feature_names, sep=sep,
vocs=vocs, max_len=config['model_params']['sequence_length'], model='train')
# 训练模型
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'],
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'], feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
clip=config['model_params']['clip'],
path_model=config['model_params']['path_model'])
model.fit(
data_dict=data_dict, dev_size=config['model_params']['dev_size'])
def __init__(self):
self.vocab_path = FLAGS.vocab_path
self.checkpoint_path = FLAGS.checkpoint_path
self.freeze_graph_path = FLAGS.freeze_graph_path
self.saved_model_path = FLAGS.saved_model_path
self.use_crf = FLAGS.use_crf
self.num_steps = FLAGS.num_steps
self.default_label = FLAGS.default_label
self.default_score = FLAGS.default_predict_score
self.data_utils = DataUtils()
self.tensorflow_utils = TensorflowUtils()
self.num_classes = self.data_utils.get_vocabulary_size(os.path.join(FLAGS.vocab_path, 'labels_vocab.txt'))
self.sequence_labeling_model = SequenceLabelingModel()
self.init_predict_graph()
def __init__(self):
self.tfrecords_path = FLAGS.tfrecords_path
self.checkpoint_path = FLAGS.checkpoint_path
self.tensorboard_path = FLAGS.tensorboard_path
self.use_crf = FLAGS.use_crf
self.learning_rate = FLAGS.learning_rate
self.learning_rate_decay_factor = FLAGS.learning_rate_decay_factor
self.decay_steps = FLAGS.decay_steps
self.clip_norm = FLAGS.clip_norm
self.max_training_step = FLAGS.max_training_step
self.train_tfrecords_filename = os.path.join(self.tfrecords_path,
'train.tfrecords')
self.test_tfrecords_filename = os.path.join(self.tfrecords_path,
'test.tfrecords')
self.data_utils = DataUtils()
self.num_classes = self.data_utils.get_vocabulary_size(
os.path.join(FLAGS.vocab_path, 'labels_vocab.txt'))
self.tensorflow_utils = TensorflowUtils()
self.sequence_labeling_model = SequenceLabelingModel()
def main():
# 加载配置文件
with open('./config.yml') as file_config:
config = yaml.load(file_config)
feature_names = config['model_params']['feature_names']
use_char_feature = config['model_params']['use_char_feature']
# 初始化embedding shape, dropouts, 预训练的embedding也在这里初始化)
feature_weight_shape_dict, feature_weight_dropout_dict, \
feature_init_weight_dict = dict(), dict(), dict()
for feature_name in feature_names:
feature_weight_shape_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['shape']
feature_weight_dropout_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['dropout_rate']
path_pre_train = config['model_params']['embed_params'][feature_name]['path']
if path_pre_train:
with open(path_pre_train, 'rb') as file_r:
feature_init_weight_dict[feature_name] = pickle.load(file_r)
# char embedding shape
if use_char_feature:
feature_weight_shape_dict['char'] = \
config['model_params']['embed_params']['char']['shape']
conv_filter_len_list = config['model_params']['conv_filter_len_list']
conv_filter_size_list = config['model_params']['conv_filter_size_list']
else:
conv_filter_len_list = None
conv_filter_size_list = None
# 加载数据
# 加载vocs
path_vocs = []
if use_char_feature:
path_vocs.append(config['data_params']['voc_params']['char']['path'])
for feature_name in feature_names:
path_vocs.append(config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
# 加载数据
sep_str = config['data_params']['sep']
assert sep_str in ['table', 'space']
sep = '\t' if sep_str == 'table' else ' '
max_len = config['model_params']['sequence_length']
word_len = config['model_params']['word_length']
data_dict = init_data(
path=config['data_params']['path_test'], feature_names=feature_names, sep=sep,
vocs=vocs, max_len=max_len, model='test', use_char_feature=use_char_feature,
word_len=word_len)
# 加载模型
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'],
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
num_layers=config['model_params']['bilstm_params']['num_layers'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'], feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
use_char_feature=use_char_feature,
conv_filter_size_list=conv_filter_size_list,
conv_filter_len_list=conv_filter_len_list,
word_length=word_len,
path_model=config['model_params']['path_model'])
saver = tf.train.Saver()
saver.restore(model.sess, config['model_params']['path_model'])
# 标记
viterbi_sequences = model.predict(data_dict)
# 写入文件
label_voc = dict()
for key in vocs[-1]:
label_voc[vocs[-1][key]] = key
with codecs.open(config['data_params']['path_test'], 'r', encoding='utf-8') as file_r:
sentences = file_r.read().strip().split('\n\n')
file_result = codecs.open(
config['data_params']['path_result'], 'w', encoding='utf-8')
for i, sentence in enumerate(sentences):
for j, item in enumerate(sentence.split('\n')):
if j < len(viterbi_sequences[i]):
file_result.write('%s\t%s\n' % (item, label_voc[viterbi_sequences[i][j]]))
else:
file_result.write('%s\tO\n' % item)
file_result.write('\n')
file_result.close()
def main():
# 加载配置文件
print("config5")
with open('./train_config/config_b2b_tag_5_only_jieba.yml') as file_config:
config = yaml.load(file_config)
feature_names = config['model_params']['feature_names']
use_char_feature = config['model_params']['use_char_feature']
# 初始化embedding shape, dropouts, 预训练的embedding也在这里初始化)
feature_weight_shape_dict, feature_weight_dropout_dict, \
feature_init_weight_dict = dict(), dict(), dict()
for feature_name in feature_names:
feature_weight_shape_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['shape']
feature_weight_dropout_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['dropout_rate']
path_pre_train = config['model_params']['embed_params'][feature_name][
'path']
if path_pre_train:
with open(path_pre_train, 'rb') as file_r:
feature_init_weight_dict[feature_name] = pickle.load(file_r)
# char embedding shape
if use_char_feature:
feature_weight_shape_dict['char'] = \
config['model_params']['embed_params']['char']['shape']
conv_filter_len_list = config['model_params']['conv_filter_len_list']
conv_filter_size_list = config['model_params']['conv_filter_size_list']
else:
conv_filter_len_list = None
conv_filter_size_list = None
# 加载数据
# 加载vocs
path_vocs = []
if use_char_feature:
path_vocs.append(config['data_params']['voc_params']['char']['path'])
for feature_name in feature_names:
path_vocs.append(
config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
# 加载训练数据
sep_str = config['data_params']['sep']
assert sep_str in ['table', 'space']
sep = '\t' if sep_str == 'table' else ' '
max_len = config['model_params']['sequence_length']
word_len = config['model_params']['word_length']
data_dict = init_data(path=config['data_params']['path_train'],
feature_names=feature_names,
sep=sep,
vocs=vocs,
max_len=max_len,
model='train',
use_char_feature=use_char_feature,
word_len=word_len)
# 训练模型
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'],
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
num_layers=config['model_params']['bilstm_params']['num_layers'],
rnn_dropout=config['model_params']['bilstm_params']['rnn_dropout'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'],
feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
clip=config['model_params']['clip'],
use_char_feature=use_char_feature,
conv_filter_size_list=conv_filter_size_list,
conv_filter_len_list=conv_filter_len_list,
cnn_dropout_rate=config['model_params']['conv_dropout'],
word_length=word_len,
path_model=config['model_params']['path_model'])
model.fit(data_dict=data_dict, dev_size=config['model_params']['dev_size'])
for feature_name in feature_names:
path_vocs.append(config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
print(vocs[-1])
print(len(vocs))
# 加载模型
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'],
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'],
feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
path_model=config['model_params']['path_model'])
def predict(string):
choiceAction = []
choiceTarget = []
choiceData = []
lab = writetxt(string)
def main():
# 加载配置文件
with open('./config.yml') as file_config:
config = yaml.load(file_config)
feature_names = config['model_params']['feature_names']
logger.info(feature_names)
use_char_feature = config['model_params']['use_char_feature']
logger.info(use_char_feature)
# 初始化embedding shape, dropouts, 预训练的embedding也在这里初始化)
feature_weight_shape_dict = dict()
feature_weight_dropout_dict = dict()
feature_init_weight_dict = dict()
for feature_name in feature_names:
feature_weight_shape_dict[feature_name] = config['model_params'][
'embed_params'][feature_name]['shape']
feature_weight_dropout_dict[feature_name] = config['model_params'][
'embed_params'][feature_name]['dropout_rate']
# embeding mat, 比voc多了两行， 因为voc从2开始编序， 0， 1行用0填充
path_pre_train = config['model_params']['embed_params'][feature_name][
'path'] # 词嵌矩阵位置
# logger.info("%s init mat path: %s" % (feature_name, path_pre_train))
with open(path_pre_train, 'rb') as file_r:
feature_init_weight_dict[feature_name] = pickle.load(file_r)
logger.info(feature_weight_dropout_dict)
logger.info(feature_weight_shape_dict)
logger.info(feature_init_weight_dict)
# char embedding shape
if use_char_feature:
# 暂时不考虑
feature_weight_shape_dict['char'] = config['model_params'][
'embed_params']['char']['shape']
conv_filter_len_list = config['model_params']['conv_filter_len_list']
conv_filter_size_list = config['model_params']['conv_filter_size_list']
else:
# 利用卷集层来提取char的信息
conv_filter_len_list = None
conv_filter_size_list = None
# 加载vocs
path_vocs = []
if use_char_feature:
path_vocs.append(config['data_params']['voc_params']['char']
['path']) # vocs用于将文本数字序列化
for feature_name in feature_names:
path_vocs.append(
config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
# 加载训练数据
sep_str = config['data_params']['sep']
assert sep_str in ['table', 'space'] # 数据的分隔方式
sep = '\t' if sep_str == 'table' else ' '
max_len = config['model_params']['sequence_length']
word_len = config['model_params']['word_length']
# 通过voc 将input f1 和输出 label 数字序列化 得到训练的输入和输出
# data_dict = None
data_dict = init_data(path=config['data_params']['path_train'],
feature_names=feature_names,
sep=sep,
vocs=vocs,
max_len=max_len,
model='train',
use_char_feature=use_char_feature,
word_len=word_len)
logger.info(data_dict) # 每个特征序列化后的数据
# 训练模型
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'], # 句子被固定长度
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
num_layers=config['model_params']['bilstm_params']['num_layers'],
rnn_dropout=config['model_params']['bilstm_params']['rnn_dropout'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'],
feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
clip=config['model_params']['clip'],
use_char_feature=use_char_feature,
conv_filter_size_list=conv_filter_size_list,
conv_filter_len_list=conv_filter_len_list,
cnn_dropout_rate=config['model_params']['conv_dropout'],
word_length=word_len,
path_model=config['model_params']['path_model'],
last_train_sess_path=None, # 为了加快训练的速度我们继续载入前面训练的参数
transfer=False) # 是否对前面载入的参数进行迁移学习，True的话就重置LSTM的输出层
model.fit(data_dict=data_dict, dev_size=config['model_params']['dev_size'])
"""
def predict(testlist):
# 加载配置文件
with open('./config.yml') as file_config:
config = yaml.load(file_config)
feature_names = config['model_params']['feature_names']
use_char_feature = config['model_params']['use_char_feature']
# 初始化embedding shape, dropouts, 预训练的embedding也在这里初始化)
feature_weight_shape_dict, feature_weight_dropout_dict, \
feature_init_weight_dict = dict(), dict(), dict()
for feature_name in feature_names:
feature_weight_shape_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['shape']
feature_weight_dropout_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['dropout_rate']
path_pre_train = config['model_params']['embed_params'][feature_name][
'path']
if path_pre_train:
with open(path_pre_train, 'rb') as file_r:
feature_init_weight_dict[feature_name] = pickle.load(file_r)
# char embedding shape
if use_char_feature:
feature_weight_shape_dict['char'] = \
config['model_params']['embed_params']['char']['shape']
conv_filter_len_list = config['model_params']['conv_filter_len_list']
conv_filter_size_list = config['model_params']['conv_filter_size_list']
else:
conv_filter_len_list = None
conv_filter_size_list = None
# 加载vocs
path_vocs = []
if use_char_feature:
path_vocs.append(config['data_params']['voc_params']['char']['path'])
for feature_name in feature_names:
path_vocs.append(
config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
# 加载数据
sep_str = config['data_params']['sep']
assert sep_str in ['table', 'space']
sep = '\t' if sep_str == 'table' else ' '
max_len = config['model_params']['sequence_length']
word_len = config['model_params']['word_length']
data_dict = init_data(path=config['data_params']['path_test'],
feature_names=feature_names,
sep=sep,
test_sens=testlist,
vocs=vocs,
max_len=max_len,
model='test',
use_char_feature=use_char_feature,
word_len=word_len)
# 加载模型
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'],
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
num_layers=config['model_params']['bilstm_params']['num_layers'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'],
feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
use_char_feature=use_char_feature,
conv_filter_size_list=conv_filter_size_list,
conv_filter_len_list=conv_filter_len_list,
word_length=word_len,
path_model=config['model_params']['path_model'])
saver = tf.train.Saver()
saver.restore(model.sess, config['model_params']['path_model'])
# print('data_dict', data_dict)
# 标记
result_sequences = model.predict(data_dict)
#print('result_sequences', result_sequences)
# 输出结果
label_voc = dict()
for key in vocs[-1]:
label_voc[vocs[-1][key]] = key
outlist = []
for i, sentence in enumerate(testlist):
templist = []
for j, item in enumerate(sentence):
#char = recheck_char(item[0])
char = item[0]
if j < len(result_sequences[i]):
out = [char, label_voc[result_sequences[i][j]]]
else:
out = [char, 'O']
templist.append(out)
outlist.append(templist)
return outlist
def export_serving_model():
"""输出tensorserving model"""
with open('./config.yml') as file_config:
config = yaml.load(file_config)
feature_names = config['model_params']['feature_names']
use_char_feature = config['model_params']['use_char_feature']
# 初始化embedding shape, dropouts, 预训练的embedding也在这里初始化)
feature_weight_shape_dict, feature_weight_dropout_dict, \
feature_init_weight_dict = dict(), dict(), dict()
for feature_name in feature_names:
feature_weight_shape_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['shape']
feature_weight_dropout_dict[feature_name] = \
config['model_params']['embed_params'][feature_name]['dropout_rate']
path_pre_train = config['model_params']['embed_params'][feature_name][
'path']
if path_pre_train:
with open(path_pre_train, 'rb') as file_r:
feature_init_weight_dict[feature_name] = pickle.load(file_r)
# char embedding shape
if use_char_feature:
feature_weight_shape_dict['char'] = \
config['model_params']['embed_params']['char']['shape']
conv_filter_len_list = config['model_params']['conv_filter_len_list']
conv_filter_size_list = config['model_params']['conv_filter_size_list']
else:
conv_filter_len_list = None
conv_filter_size_list = None
# 加载vocs
path_vocs = []
for feature_name in feature_names:
path_vocs.append(
config['data_params']['voc_params'][feature_name]['path'])
path_vocs.append(config['data_params']['voc_params']['label']['path'])
vocs = load_vocs(path_vocs)
# 加载数据
sep_str = config['data_params']['sep']
assert sep_str in ['table', 'space']
sep = '\t' if sep_str == 'table' else ' '
max_len = config['model_params']['sequence_length']
word_len = config['model_params']['word_length']
model = SequenceLabelingModel(
sequence_length=config['model_params']['sequence_length'],
nb_classes=config['model_params']['nb_classes'],
nb_hidden=config['model_params']['bilstm_params']['num_units'],
num_layers=config['model_params']['bilstm_params']['num_layers'],
feature_weight_shape_dict=feature_weight_shape_dict,
feature_init_weight_dict=feature_init_weight_dict,
feature_weight_dropout_dict=feature_weight_dropout_dict,
dropout_rate=config['model_params']['dropout_rate'],
nb_epoch=config['model_params']['nb_epoch'],
feature_names=feature_names,
batch_size=config['model_params']['batch_size'],
train_max_patience=config['model_params']['max_patience'],
use_crf=config['model_params']['use_crf'],
l2_rate=config['model_params']['l2_rate'],
rnn_unit=config['model_params']['rnn_unit'],
learning_rate=config['model_params']['learning_rate'],
use_char_feature=use_char_feature,
conv_filter_size_list=conv_filter_size_list,
conv_filter_len_list=conv_filter_len_list,
word_length=word_len,
path_model=config['model_params']['path_model'])
session = model.sess
saver = tf.train.Saver()
saver.restore(session, config['model_params']['path_model'])
# 输出tensorserving model 过程
model_version = 1
work_dir = './Model/ner_model'
export_path_base = work_dir
export_path = os.path.join(tf.compat.as_bytes(export_path_base),
tf.compat.as_bytes(str(model_version)))
print('Exporting trained model to', export_path)
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
# 定义输入变量
tensor_info_input_x_f1 = tf.saved_model.utils.build_tensor_info(
model.input_feature_ph_dict['f1'])
tensor_info_weight_dropout_ph_dict_f1 = tf.saved_model.utils.build_tensor_info(
model.weight_dropout_ph_dict['f1'])
tensor_info_dropout_rate_ph = tf.saved_model.utils.build_tensor_info(
model.dropout_rate_ph)
tensor_info_rnn_dropout_rate_ph = tf.saved_model.utils.build_tensor_info(
model.rnn_dropout_rate_ph)
tensor_info_logits = tf.saved_model.utils.build_tensor_info(model.logits)
tensor_info_actual_length = tf.saved_model.utils.build_tensor_info(
model.sequence_actual_length)
tensor_info_transition_params = tf.saved_model.utils.build_tensor_info(
model.transition_params)
# 构建过程
prediction_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={
'input_x_f1': tensor_info_input_x_f1,
'weight_dropout_ph_dict_f1':
tensor_info_weight_dropout_ph_dict_f1,
'dropout_rate_ph': tensor_info_dropout_rate_ph,
'rnn_dropout_rate_ph': tensor_info_rnn_dropout_rate_ph
},
outputs={
'transition_params': tensor_info_transition_params,
'logits': tensor_info_logits,
'sequence_actual_length': tensor_info_actual_length,
},
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
)
builder.add_meta_graph_and_variables(
session, [tf.saved_model.tag_constants.SERVING],
signature_def_map={'ner_predict': prediction_signature},
main_op=tf.tables_initializer(),
strip_default_attrs=True)
builder.save()
print('Done exporting!')