def wrapper(**kwargs):
opt = tf.sg_opt(kwargs)
# default training options
opt += tf.sg_opt(lr=0.001,
save_dir='asset/train',
max_ep=1000,
ep_size=100000,
save_interval=600,
log_interval=60,
early_stop=True,
lr_reset=False,
eval_metric=[],
max_keep=5,
keep_interval=1,
tqdm=True,
console_log=False)
# make directory if not exist
if not os.path.exists(opt.save_dir + '/log'):
os.makedirs(opt.save_dir + '/log')
if not os.path.exists(opt.save_dir + '/ckpt'):
os.makedirs(opt.save_dir + '/ckpt')
# find last checkpoint
last_file = tf.train.latest_checkpoint(opt.save_dir + '/ckpt')
if last_file:
ep = start_ep = int(last_file.split('-')[1]) + 1
start_step = int(last_file.split('-')[2])
else:
ep = start_ep = 1
start_step = 0
# checkpoint saver
saver = tf.train.Saver(max_to_keep=opt.max_keep,
keep_checkpoint_every_n_hours=opt.keep_interval)
# summary writer
summary_writer = tf.train.SummaryWriter(opt.save_dir + '/log',
graph=tf.get_default_graph())
# add learning rate summary
with tf.name_scope('summary'):
tf.scalar_summary('60. learning_rate/learning_rate',
_learning_rate)
# add evaluation metric summary
for m in opt.eval_metric:
tf.sg_summary_metric(m)
# summary op
summary_op = tf.merge_all_summaries()
# create session
if opt.sess:
sess = opt.sess
else:
# session with multiple GPU support
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# initialize variables
sg_init(sess)
# restore last checkpoint
if last_file:
saver.restore(sess, last_file)
# set learning rate
if start_ep == 1 or opt.lr_reset:
sess.run(_learning_rate.assign(opt.lr))
# logging
tf.sg_info('Training started from epoch[%03d]-step[%d].' %
(start_ep, start_step))
try:
# start data queue runner
with tf.sg_queue_context(sess):
# set session mode to train
tf.sg_set_train(sess)
# loss history for learning rate decay
loss, loss_prev, early_stopped = None, None, False
# time stamp for saving and logging
last_saved = last_logged = time.time()
# epoch loop
for ep in range(start_ep, opt.max_ep + 1):
# show progressbar
if opt.tqdm:
iterator = tqdm(range(opt.ep_size),
desc='train',
ncols=70,
unit='b',
leave=False)
else:
iterator = range(opt.ep_size)
# batch loop
for _ in iterator:
# call train function
batch_loss = func(sess, opt)
# loss history update
if batch_loss is not None:
if loss is None:
loss = np.mean(batch_loss)
else:
loss = loss * 0.9 + np.mean(batch_loss) * 0.1
# saving
if time.time() - last_saved > opt.save_interval:
last_saved = time.time()
saver.save(sess,
opt.save_dir + '/ckpt/model-%03d' % ep,
write_meta_graph=False,
global_step=sess.run(
tf.sg_global_step()))
# logging
if time.time() - last_logged > opt.log_interval:
last_logged = time.time()
# set session mode to infer
tf.sg_set_infer(sess)
# run evaluation op
if len(opt.eval_metric) > 0:
sess.run(opt.eval_metric)
if opt.console_log: # console logging
# log epoch information
tf.sg_info(
'\tEpoch[%03d:lr=%7.5f:gs=%d] - loss = %s'
% (ep, sess.run(_learning_rate),
sess.run(tf.sg_global_step()),
('NA' if loss is None else '%8.6f' %
loss)))
else: # tensorboard logging
# run logging op
summary_writer.add_summary(
sess.run(summary_op),
global_step=sess.run(tf.sg_global_step()))
# learning rate decay
if opt.early_stop and loss_prev:
# if loss stalling
if loss >= 0.95 * loss_prev:
# early stopping
current_lr = sess.run(_learning_rate)
if current_lr < 5e-6:
early_stopped = True
break
else:
# decrease learning rate by half
sess.run(
_learning_rate.assign(current_lr /
2.))
# update loss history
loss_prev = loss
# revert session mode to train
tf.sg_set_train(sess)
# log epoch information
if not opt.console_log:
tf.sg_info(
'\tEpoch[%03d:lr=%7.5f:gs=%d] - loss = %s' %
(ep, sess.run(_learning_rate),
sess.run(tf.sg_global_step()),
('NA' if loss is None else '%8.6f' % loss)))
if early_stopped:
tf.sg_info('\tEarly stopped ( no loss progress ).')
break
finally:
# save last epoch
saver.save(sess,
opt.save_dir + '/ckpt/model-%03d' % ep,
write_meta_graph=False,
global_step=sess.run(tf.sg_global_step()))
# set session mode to infer
tf.sg_set_infer(sess)
# logging
tf.sg_info('Training finished at epoch[%d]-step[%d].' %
(ep, sess.run(tf.sg_global_step())))
# close session
if opt.sess is None:
sess.close()
def generate():
dev = '/cpu:0'
with tf.device(dev):
mydir = 'tfrc150char_wrd0704'
files = [f for f in listdir(mydir) if isfile(join(mydir, f))]
tfrecords_filename = []
tfrecords_filename = [join(mydir, 'short_infer3.tfrecords')
] #[join(mydir, f) for f in tfrecords_filename]
tfrecords_filename_inf = [join(mydir, '11_3.tfrecords')]
print(tfrecords_filename)
filename_queue = tf.train.string_input_producer(tfrecords_filename,
num_epochs=num_epochs,
shuffle=True,
capacity=1)
infer_queue = tf.train.string_input_producer(tfrecords_filename_inf,
num_epochs=num_epochs,
shuffle=True,
capacity=1)
optim = tf.train.AdamOptimizer(learning_rate=0.0001,
beta1=0.9,
beta2=0.99)
# Calculate the gradients for each model tower.
tower_grads = []
reuse_vars = False
with tf.variable_scope("dec_lstm") as scp:
dec_cell = BasicLSTMCell2(Hp.w_emb_size,
Hp.rnn_hd,
state_is_tuple=True)
with tf.variable_scope("contx_lstm") as scp:
cell = BasicLSTMCell2(Hp.hd, Hp.rnn_hd, state_is_tuple=True)
rnn_cell = tf.contrib.rnn.DropoutWrapper(
cell,
input_keep_prob=Hp.keep_prob,
output_keep_prob=Hp.keep_prob)
(words, chars) = read_and_decode(filename_queue,
Hp.batch_size * Hp.num_gpus)
words_splits = tf.split(axis=0,
num_or_size_splits=Hp.num_gpus,
value=words)
chars_splits = tf.split(axis=0,
num_or_size_splits=Hp.num_gpus,
value=chars)
word_emb = np.loadtxt("glove300d_0704.txt")
Hp.word_vs = word_emb.shape[0]
# --------------------------------------------------------------------------------
with tf.name_scope('%s_%d' % ("tower", 0)) as scope:
rnn_state = tower_infer_enc(chars_splits[0],
scope,
rnn_cell,
dec_cell,
word_emb,
out_reuse_vars=False,
dev='/cpu:0')
chars_pl = tf.placeholder(tf.int32, shape=(None, Hp.c_maxlen))
rnn_state_pl1 = [
tf.placeholder(tf.float32, shape=(None, Hp.rnn_hd)),
tf.placeholder(tf.float32, shape=(None, Hp.rnn_hd))
]
rnn_state_pl = tf.contrib.rnn.LSTMStateTuple(
rnn_state_pl1[0], rnn_state_pl1[1])
final_ids, rnn_state_dec = tower_infer_dec(chars_pl,
scope,
rnn_cell,
dec_cell,
word_emb,
rnn_state_pl,
out_reuse_vars=False,
dev='/cpu:0')
# --------------------------------------------------------------------------------
saver = tf.train.Saver(tf.trainable_variables())
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
session_config.gpu_options.per_process_gpu_memory_fraction = 0.94
session_config.gpu_options.allow_growth = False
restore_dir = 'tnsrbrd/hin17d08m_1313g2' # lec30d07m_1634g2 lec04d07m_2006g2 lec28d07m_1221g2 lec31d07m_1548g2
csv_file = join(restore_dir, time.strftime("hin%dd%mm_%H%M.csv"))
csv_f = open(csv_file, 'a')
csv_writer = csv.writer(csv_f)
with tf.Session(config=session_config) as sess:
sess.run(
tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer()))
tf.train.start_queue_runners(sess=sess)
saver.restore(sess,
tf.train.latest_checkpoint(
join(restore_dir,
'last_chpt'))) # lec04d07m_2006g2
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for ep in range(num_epochs):
tf.sg_set_infer(sess)
rnn_state_val, w_txt, ch_txt = sess.run(
[rnn_state, words_splits[0], chars_splits[0]],
feed_dict={Hp.keep_prob: 1.0})
predictions = [] #[w_txt[:,2,:]]
for idx in range(3):
char_inpt = word2char_ids(
ids_val) if idx != 0 else ch_txt[:, 2, :]
ids_val, rnn_state_val = sess.run(
[final_ids, rnn_state_dec],
feed_dict={
Hp.keep_prob: 1.0,
rnn_state_pl1[0]: rnn_state_val[0],
rnn_state_pl1[1]: rnn_state_val[1],
chars_pl: char_inpt
})
temp = np.zeros((Hp.batch_size, Hp.w_maxlen))
for b in range(Hp.batch_size):
stop_ind = np.where(ids_val[b] == 2)[0]
if stop_ind.size > 0:
stop_ind = stop_ind[0]
ids_val[b, stop_ind +
1:] = ids_val[b, stop_ind + 1:] * 0
temp[:, :ids_val.shape[1]] = ids_val
predictions.append(temp)
# predictions are decode_sent x b x w_maxlen
predictions = np.array(predictions)
in_batches = [w_txt[b, :, :] for b in range(Hp.batch_size)]
res_batches = [
predictions[:, b, :] for b in range(Hp.batch_size)
]
for b in range(Hp.batch_size):
in_paragraph = idxword2txt(in_batches[b])
print("\n INPUT SAMPLE \n")
print(in_paragraph)
res_paragraph = idxword2txt(res_batches[b])
print("\n RESULTS \n")
print(res_paragraph)
csv_writer.writerow([
" ".join(in_paragraph[:3]), " ".join(in_paragraph[3:]),
" ".join(res_paragraph)
])
csv_f.close()
def wrapper(**kwargs):
r""" Manages arguments of `tf.sg_opt`.
Args:
**kwargs:
lr: A Python Scalar (optional). Learning rate. Default is .001.
eval_metric: A list of tensors containing the value to evaluate. Default is [].
early_stop: Boolean. If True (default), the training should stop when the following two conditions are met.
i. Current loss is less than .95 * previous loss.
ii. Current learning rate is less than 5e-6.
lr_reset: Boolean. If True, learning rate is set to opt.lr. when training restarts.
Otherwise (Default), the value of the stored `_learning_rate` is taken.
save_dir: A string. The root path to which checkpoint and log files are saved.
Default is `asset/train`.
max_ep: A positive integer. Maximum number of epochs. Default is 1000.
ep_size: A positive integer. Number of Total batches in an epoch.
For proper display of log. Default is 1e5.
save_interval: A Python scalar. The interval of saving checkpoint files.
By default, for every 600 seconds, a checkpoint file is written.
log_interval: A Python scalar. The interval of recoding logs.
By default, for every 60 seconds, logging is executed.
max_keep: A positive integer. Maximum number of recent checkpoints to keep. Default is 5.
keep_interval: A Python scalar. How often to keep checkpoints. Default is 1 hour.
tqdm: Boolean. If True (Default), progress bars are shown.
console_log: Boolean. If True, a series of loss will be shown
on the console instead of tensorboard. Default is False.
"""
opt = tf.sg_opt(kwargs)
# default training options
opt += tf.sg_opt(lr=0.001,
save_dir='asset/train',
max_ep=1000, ep_size=100000,
save_interval=600, log_interval=60,
early_stop=True, lr_reset=False,
eval_metric=[],
max_keep=5, keep_interval=1,
tqdm=True, console_log=False)
# make directory if not exist
if not os.path.exists(opt.save_dir):
os.makedirs(opt.save_dir)
# find last checkpoint
last_file = tf.train.latest_checkpoint(opt.save_dir)
if last_file:
ep = start_ep = int(last_file.split('-')[1]) + 1
start_step = int(last_file.split('-')[2])
else:
ep = start_ep = 1
start_step = 0
# checkpoint saver
saver = tf.train.Saver(max_to_keep=opt.max_keep,
keep_checkpoint_every_n_hours=opt.keep_interval)
# summary writer
summary_writer = tf.summary.FileWriter(opt.save_dir, graph=tf.get_default_graph())
# add learning rate summary
tf.summary.scalar('learning_r', _learning_rate)
# add evaluation metric summary
for m in opt.eval_metric:
tf.sg_summary_metric(m)
# summary op
summary_op = tf.summary.merge_all()
# create session
if opt.sess:
sess = opt.sess
else:
# session with multiple GPU support
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))
# initialize variables
sg_init(sess)
# restore last checkpoint
if last_file:
saver.restore(sess, last_file)
# set learning rate
if start_ep == 1 or opt.lr_reset:
sess.run(_learning_rate.assign(opt.lr))
# logging
tf.sg_info('Training started from epoch[%03d]-step[%d].' % (start_ep, start_step))
try:
# start data queue runner
with tf.sg_queue_context(sess):
# set session mode to train
tf.sg_set_train(sess)
# loss history for learning rate decay
loss, loss_prev, early_stopped = None, None, False
# time stamp for saving and logging
last_saved = last_logged = time.time()
# epoch loop
for ep in range(start_ep, opt.max_ep + 1):
# show progressbar
if opt.tqdm:
iterator = tqdm(range(opt.ep_size), desc='train', ncols=70, unit='b', leave=False)
else:
iterator = range(opt.ep_size)
# batch loop
for _ in iterator:
# call train function
batch_loss = func(sess, opt)
# loss history update
if batch_loss is not None:
if loss is None:
loss = np.mean(batch_loss)
else:
loss = loss * 0.9 + np.mean(batch_loss) * 0.1
# saving
if time.time() - last_saved > opt.save_interval:
last_saved = time.time()
saver.save(sess, opt.save_dir + '/model-%03d' % ep,
write_meta_graph=False,
global_step=sess.run(tf.sg_global_step()))
# logging
if time.time() - last_logged > opt.log_interval:
last_logged = time.time()
# set session mode to infer
tf.sg_set_infer(sess)
# run evaluation op
if len(opt.eval_metric) > 0:
sess.run(opt.eval_metric)
if opt.console_log: # console logging
# log epoch information
tf.sg_info('\tEpoch[%03d:lr=%7.5f:gs=%d] - loss = %s' %
(ep, sess.run(_learning_rate), sess.run(tf.sg_global_step()),
('NA' if loss is None else '%8.6f' % loss)))
else: # tensorboard logging
# run logging op
summary_writer.add_summary(sess.run(summary_op),
global_step=sess.run(tf.sg_global_step()))
# learning rate decay
if opt.early_stop and loss_prev:
# if loss stalling
if loss >= 0.95 * loss_prev:
# early stopping
current_lr = sess.run(_learning_rate)
if current_lr < 5e-6:
early_stopped = True
break
else:
# decrease learning rate by half
sess.run(_learning_rate.assign(current_lr / 2.))
# update loss history
loss_prev = loss
# revert session mode to train
tf.sg_set_train(sess)
# log epoch information
if not opt.console_log:
tf.sg_info('\tEpoch[%03d:lr=%7.5f:gs=%d] - loss = %s' %
(ep, sess.run(_learning_rate), sess.run(tf.sg_global_step()),
('NA' if loss is None else '%8.6f' % loss)))
if early_stopped:
tf.sg_info('\tEarly stopped ( no loss progress ).')
break
finally:
# save last epoch
saver.save(sess, opt.save_dir + '/model-%03d' % ep,
write_meta_graph=False,
global_step=sess.run(tf.sg_global_step()))
# set session mode to infer
tf.sg_set_infer(sess)
# logging
tf.sg_info('Training finished at epoch[%d]-step[%d].' % (ep, sess.run(tf.sg_global_step())))
# close session
if opt.sess is None:
sess.close()