def val(self):
var_list = tf.global_variables()
saver = tf.train.Saver(var_list,
max_to_keep=5,
filename=self.ckpt_name)
logit_tensor = self.logit
length_tensor = self.L
with tf.Session(config=self.gpu_config) as sess:
ckpt = tf.train.latest_checkpoint(self.ckpt_path, self.ckpt_name)
saver.restore(sess, ckpt)
logit_list = []
length_list = []
for _ in range(self.val_steps):
logit_array, length_array = sess.run(
[logit_tensor, length_tensor],
feed_dict={self.drop_flag: False})
logit_list.extend(sigmoid(logit_array))
length_list.extend(length_array)
with open('./data/dev_data_char.json', encoding='utf-8') as file:
data = [json.loads(line) for line in file]
for t in np.arange(0.1, 0.6, 0.1):
decode_fn(logit_list,
length_list,
data,
'./result_dev_%.3f.json' % t,
num_target=self.num_target,
threshold=t)
p, r, f = compute_Fscore('./result_dev_%.3f.json' % t,
'./data/dev_data.json')
print(p, r, f, t)
def val_with_threshold(self, threshold=0.5):
var_list = tf.global_variables()
saver = tf.train.Saver(var_list,
max_to_keep=5,
filename=self.ckpt_name)
logit_tensor = tf.sigmoid(self.logit)
length_tensor = self.L
e = 0
with tf.Session(config=self.gpu_config) as sess:
last_ckpt = ''
while True:
ckpt = tf.train.latest_checkpoint(self.ckpt_path,
self.ckpt_name)
if not ckpt or ckpt == last_ckpt:
time.sleep(60)
continue
saver.restore(sess, ckpt)
e = e + 1
last_ckpt = ckpt
print(ckpt)
logit_list = []
length_list = []
for i in range(self.val_steps):
logit_array, length_array = sess.run(
[logit_tensor, length_tensor],
feed_dict={self.drop_flag: False})
logit_list.extend(logit_array)
length_list.extend(length_array)
# print(i)
print(e)
# logit_list = [sigmoid(logit) for logit in logit_list]
with open('./data/dev_data_char.json',
encoding='utf-8') as file:
data = [json.loads(line) for line in file]
decode_fn(logit_list,
length_list,
data,
'./result_dev_epoch_%d.json' % e,
num_target=self.num_target,
threshold=threshold)
p, r, f = compute_Fscore('./result_dev_epoch_%d.json' % e,
'./data/dev_data.json')
def train_val(self, val_fn=None):
assert self.loss != None, 'must compute loss before train'
var_list = tf.global_variables()
train_variables = tf.trainable_variables()
global_step = self.global_step
run_ops = {'step': global_step, 'loss': self.loss}
grads_and_vars = self.optimizer.compute_gradients(
self.loss, train_variables)
run_ops['train_ops'] = self.optimizer.apply_gradients(
grads_and_vars, global_step=global_step)
saver = tf.train.Saver(var_list,
max_to_keep=5,
filename=self.ckpt_name)
initializer = tf.global_variables_initializer()
with tf.Session(config=self.gpu_config) as sess:
sess.run(initializer)
ckpt = tf.train.latest_checkpoint(self.ckpt_path, self.ckpt_name)
if ckpt:
saver.restore(sess, ckpt)
train_handle = sess.run(self.train_iterator)
val_handle = sess.run(self.val_iterator)
best_score = 0
with open('./data/dev_data_char.json', encoding='utf-8') as file:
data = [json.loads(line) for line in file]
for e in range(self.train_epochs):
result = {}
# train steps
for i in range(self.steps_each_epoch):
result = sess.run(run_ops,
feed_dict={
self.handle_holder: train_handle,
self.drop_flag: True
})
if result['step'] % 100 == 0:
print('%d:\t%f' % (result['step'], result['loss']))
logit_list = []
label_list = []
loss_list = []
length_list = []
#validation steps:
for j in range(self.val_steps):
logit_array,label_array,length_array,loss_ =\
sess.run([self.logit,self.target,self.L,self.loss],
feed_dict={self.handle_holder:val_handle,self.drop_flag:False})
logit_list.extend(logit_array)
label_list.extend(label_array)
length_list.extend(length_array)
loss_list.append(loss_)
logit_list = [sigmoid(logit) for logit in logit_list]
print('epoch %d' % e)
# score = val_fn(label_list,logit_list)
print('loss:\t%f' % np.mean(loss_list))
out_file = './reslut_dev_epoch_%d.json' % e
decode_fn(logit_list,
length_list,
data,
out_file,
num_target=self.num_target)
p, r, f = compute_Fscore(out_file, './data/dev_data.json')
print('P:\t%f,R:\t%f,F1:\t%f' % (p, r, f))
if f > best_score:
best_score = f
saver.save(sess,
self.ckpt_path,
global_step=result['step'],
latest_filename=self.ckpt_name)