def main(args):
is_train = not args['nottrain']
is_save = args['savemodel']
model_path = args['modelpath']
epoch = args['epoch']
config = load_conf(args)
config['nepoch'] = epoch
#Load Dataset
train_data, test_data = load_tt_datas(args, config)
# setup randomer
Randomer.set_stddev(config['stddev'])
with tf.Graph().as_default():
# build model
model = Seq2SeqAttNN(config)
model.build_model()
if is_save or not is_train:
saver = tf.train.Saver(max_to_keep=30)
else:
saver = None
# run
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
if is_train:
t1 = time.time()
model.train(sess, train_data, test_data, saver)
t2 = time.time()
print('Training+Evaluation Time = ', t2 - t1)
else:
t1 = time.time()
saver.restore(sess, model_path)
model.test(sess, test_data)
t2 = time.time()
print('Testing Time = ', t2 - t1)
def __init__(self, hidden_size, stddev=None):
'''
hidden_size: int,
'''
self.stddev = stddev
self.hidden_size = hidden_size
self.W = tf.Variable(Randomer.random_normal(
[self.hidden_size, self.hidden_size]),
trainable=True)
self.U = tf.Variable(Randomer.random_normal(
[self.hidden_size, self.hidden_size]),
trainable=True)
self.b = tf.Variable(tf.zeros([1]), trainable=True)
def __init__(self, w_shape, stddev=None, params=None):
'''
:param w_shape: [input_dim, output_dim]
:param stddev: 用于初始化
:param params: 从外界制定参数
'''
if params is None:
self.w = tf.Variable(Randomer.random_normal(w_shape),
trainable=True)
else:
self.w = params['w']
def __init__(self, edim, class_num, stddev=None, params=None):
'''
class_num: class type num.
edim: the input embedding dim.
params = {'wline': wline, 'bline': bline}
'''
self.edim = edim
self.class_num = class_num
# the linear basic_layer for softmax.
if params is None:
self.wline_softmax = tf.Variable(Randomer.random_normal(
[self.edim, self.class_num]),
trainable=True)
self.bline_softmax = tf.Variable(tf.zeros([1, 1]), trainable=True)
else:
self.wline_softmax = params['wline']
self.bline_softmax = params['bline']
def __init__(self, w_shape=None, stddev=None, params=None, active='tanh'):
'''
the initialize function.
w_shape is the shape of the w param. if params is None, need.
staddev is the stddev of the tf.random_normal. if params is None, need.
params = {'wline':wline}, is use to assign the params.
active is the active function.
'''
self.w_shape = w_shape
self.stddev = stddev
if params is None:
self.wline = tf.Variable(
Randomer.random_normal(self.w_shape),
# tf.random_uniform(self.w_shape, -0.0015, 0.035),
trainable=True)
else:
self.wline = params['wline']
self.active = active
def main(options, modelconf="config/model.conf"):
'''
model: 需要加载的模型
dataset: 需要加载的数据集
reload: 是否需要重新加载数据，yes or no
modelconf: model config文件所在的路径
class_num: 分类的类别
use_term: 是否是对aspect term 进行分类
'''
model = options.model
dataset = options.dataset
reload = options.reload
class_num = options.classnum
is_train = not options.not_train
is_save = not options.not_save_model
model_path = options.model_path
input_data = options.input_data
epoch = options.epoch
module, obj, config = load_conf(model, modelconf)
config['model'] = model
print(model)
config['dataset'] = dataset
config['class_num'] = class_num
config['nepoch'] = epoch
train_data, test_data = load_tt_datas(config, reload)
module = __import__(module, fromlist=True)
# setup randomer
Randomer.set_stddev(config['stddev'])
with tf.Graph().as_default():
# build model
model = getattr(module, obj)(config)
model.build_model()
if is_save or not is_train:
saver = tf.train.Saver(max_to_keep=30)
else:
saver = None
# run
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
if is_train:
print(dataset)
if dataset == "cikm16":
model.train(sess, train_data, test_data, saver, threshold_acc=config['cikm_threshold_acc'])
elif dataset == "rsc15":
model.train(sess, train_data, test_data, saver, threshold_acc=config['recsys_threshold_acc'])
else:
model.train(sess, train_data, test_data, saver)
# if dataset == "rsc15":
# model.train(sess, train_data, test_data, saver, threshold_acc=config['recsys_threshold_acc'])
else:
if input_data is "test":
sent_data = test_data
elif input_data is "train":
sent_data = train_data
else:
sent_data = test_data
saver.restore(sess, model_path)
model.test(sess, sent_data)