def do_train(network, param):
"""Run training.
Args:
network: network to train
param: A dictionary of parameters
"""
# Load dataset
train_data = Dataset(data_type='train',
label_type=param['label_type'],
train_data_size=param['train_data_size'],
batch_size=param['batch_size'],
num_stack=param['num_stack'],
num_skip=param['num_skip'],
is_sorted=True)
dev_data_step = Dataset(data_type='dev',
label_type=param['label_type'],
train_data_size=param['train_data_size'],
batch_size=param['batch_size'],
num_stack=param['num_stack'],
num_skip=param['num_skip'],
is_sorted=False)
dev_data_epoch = Dataset(data_type='dev',
label_type=param['label_type'],
train_data_size=param['train_data_size'],
batch_size=param['batch_size'],
num_stack=param['num_stack'],
num_skip=param['num_skip'],
is_sorted=False)
# Tell TensorFlow that the model will be built into the default graph
with tf.Graph().as_default():
# Define placeholders
network.inputs = tf.placeholder(tf.float32,
shape=[None, None, network.input_size],
name='input')
indices_pl = tf.placeholder(tf.int64, name='indices')
values_pl = tf.placeholder(tf.int32, name='values')
shape_pl = tf.placeholder(tf.int64, name='shape')
network.labels = tf.SparseTensor(indices_pl, values_pl, shape_pl)
network.inputs_seq_len = tf.placeholder(tf.int64,
shape=[None],
name='inputs_seq_len')
network.keep_prob_input = tf.placeholder(tf.float32,
name='keep_prob_input')
network.keep_prob_hidden = tf.placeholder(tf.float32,
name='keep_prob_hidden')
# Add to the graph each operation (including model definition)
loss_op, logits = network.compute_loss(network.inputs, network.labels,
network.inputs_seq_len,
network.keep_prob_input,
network.keep_prob_hidden)
train_op = network.train(loss_op,
optimizer=param['optimizer'],
learning_rate_init=float(
param['learning_rate']),
is_scheduled=False)
decode_op = network.decoder(logits,
network.inputs_seq_len,
decode_type='beam_search',
beam_width=20)
ler_op = network.compute_ler(decode_op, network.labels)
# Build the summary tensor based on the TensorFlow collection of
# summaries
summary_train = tf.summary.merge(network.summaries_train)
summary_dev = tf.summary.merge(network.summaries_dev)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# Create a saver for writing training checkpoints
saver = tf.train.Saver(max_to_keep=None)
# Count total parameters
parameters_dict, total_parameters = count_total_parameters(
tf.trainable_variables())
for parameter_name in sorted(parameters_dict.keys()):
print("%s %d" % (parameter_name, parameters_dict[parameter_name]))
print("Total %d variables, %s M parameters" %
(len(parameters_dict.keys()), "{:,}".format(
total_parameters / 1000000)))
csv_steps, csv_train_loss, csv_dev_loss = [], [], []
csv_ler_train, csv_ler_dev = [], []
# Create a session for running operation on the graph
with tf.Session() as sess:
# Instantiate a SummaryWriter to output summaries and the graph
summary_writer = tf.summary.FileWriter(network.model_dir,
sess.graph)
# Initialize parameters
sess.run(init_op)
# Make mini-batch generator
mini_batch_train = train_data.next_batch()
mini_batch_dev = dev_data_step.next_batch()
# Train model
iter_per_epoch = int(train_data.data_num / param['batch_size'])
train_step = train_data.data_num / param['batch_size']
if (train_step) != int(train_step):
iter_per_epoch += 1
max_steps = iter_per_epoch * param['num_epoch']
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
error_best = 1
for step in range(max_steps):
# Create feed dictionary for next mini batch (train)
with tf.device('/cpu:0'):
inputs, labels, inputs_seq_len, _ = mini_batch_train.__next__(
)
feed_dict_train = {
network.inputs: inputs,
network.labels: list2sparsetensor(labels, padded_value=-1),
network.inputs_seq_len: inputs_seq_len,
network.keep_prob_input: network.dropout_ratio_input,
network.keep_prob_hidden: network.dropout_ratio_hidden,
network.lr: float(param['learning_rate'])
}
# Update parameters
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % 200 == 0:
# Create feed dictionary for next mini batch (dev)
with tf.device('/cpu:0'):
inputs, labels, inputs_seq_len, _ = mini_batch_dev.__next__(
)
feed_dict_dev = {
network.inputs: inputs,
network.labels: list2sparsetensor(labels,
padded_value=-1),
network.inputs_seq_len: inputs_seq_len,
network.keep_prob_input: network.dropout_ratio_input,
network.keep_prob_hidden: network.dropout_ratio_hidden
}
# Compute loss_
loss_train = sess.run(loss_op, feed_dict=feed_dict_train)
loss_dev = sess.run(loss_op, feed_dict=feed_dict_dev)
csv_steps.append(step)
csv_train_loss.append(loss_train)
csv_dev_loss.append(loss_dev)
# Change to evaluation mode
feed_dict_train[network.keep_prob_input] = 1.0
feed_dict_train[network.keep_prob_hidden] = 1.0
feed_dict_dev[network.keep_prob_input] = 1.0
feed_dict_dev[network.keep_prob_hidden] = 1.0
# Compute accuracy & update event file
ler_train, summary_str_train = sess.run(
[ler_op, summary_train], feed_dict=feed_dict_train)
ler_dev, summary_str_dev = sess.run(
[ler_op, summary_dev], feed_dict=feed_dict_dev)
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
summary_writer.add_summary(summary_str_train, step + 1)
summary_writer.add_summary(summary_str_dev, step + 1)
summary_writer.flush()
duration_step = time.time() - start_time_step
print(
'Step %d: loss = %.3f (%.3f) / ler = %.4f (%.4f) (%.3f min)'
% (step + 1, loss_train, loss_dev, ler_train, ler_dev,
duration_step / 60))
sys.stdout.flush()
start_time_step = time.time()
# Save checkpoint and evaluate model per epoch
if (step + 1) % iter_per_epoch == 0 or (step + 1) == max_steps:
duration_epoch = time.time() - start_time_epoch
epoch = (step + 1) // iter_per_epoch
print('-----EPOCH:%d (%.3f min)-----' %
(epoch, duration_epoch / 60))
# Save model (check point)
checkpoint_file = join(network.model_dir, 'model.ckpt')
save_path = saver.save(sess,
checkpoint_file,
global_step=epoch)
print("Model saved in file: %s" % save_path)
if epoch >= 5:
start_time_eval = time.time()
print('=== Dev Evaluation ===')
cer_dev_epoch = do_eval_cer(
session=sess,
decode_op=decode_op,
network=network,
dataset=dev_data_epoch,
label_type=param['label_type'],
eval_batch_size=param['batch_size'])
if param['label_type'] in ['kana', 'kanji']:
print(' CER: %f %%' % (cer_dev_epoch * 100))
else:
print(' PER: %f %%' % (cer_dev_epoch * 100))
if cer_dev_epoch < error_best:
error_best = cer_dev_epoch
print('■■■ ↑Best Score↑ ■■■')
duration_eval = time.time() - start_time_eval
print('Evaluation time: %.3f min' %
(duration_eval / 60))
start_time_epoch = time.time()
start_time_step = time.time()
duration_train = time.time() - start_time_train
print('Total time: %.3f hour' % (duration_train / 3600))
# Save train & dev loss, ler
save_loss(csv_steps,
csv_train_loss,
csv_dev_loss,
save_path=network.model_dir)
save_ler(csv_steps,
csv_ler_train,
csv_ler_dev,
save_path=network.model_dir)
# Training was finished correctly
with open(join(network.model_dir, 'complete.txt'), 'w') as f:
f.write('')
def do_eval(network, param, epoch=None):
"""Evaluate the model.
Args:
network: model to restore
param: A dictionary of parameters
epoch: int, the epoch to restore
"""
# Load dataset
eval1_data = Dataset(data_type='eval1',
label_type=param['label_type'],
batch_size=1,
train_data_size=param['train_data_size'],
num_stack=param['num_stack'],
num_skip=param['num_skip'],
is_sorted=False,
is_progressbar=True)
eval2_data = Dataset(data_type='eval2',
label_type=param['label_type'],
batch_size=1,
train_data_size=param['train_data_size'],
num_stack=param['num_stack'],
num_skip=param['num_skip'],
is_sorted=False,
is_progressbar=True)
eval3_data = Dataset(data_type='eval3',
label_type=param['label_type'],
batch_size=1,
train_data_size=param['train_data_size'],
num_stack=param['num_stack'],
num_skip=param['num_skip'],
is_sorted=False,
is_progressbar=True)
# Define placeholders
network.inputs = tf.placeholder(tf.float32,
shape=[None, None, network.input_size],
name='input')
indices_pl = tf.placeholder(tf.int64, name='indices')
values_pl = tf.placeholder(tf.int32, name='values')
shape_pl = tf.placeholder(tf.int64, name='shape')
network.labels = tf.SparseTensor(indices_pl, values_pl, shape_pl)
network.inputs_seq_len = tf.placeholder(tf.int64,
shape=[None],
name='inputs_seq_len')
network.keep_prob_input = tf.placeholder(tf.float32,
name='keep_prob_input')
network.keep_prob_hidden = tf.placeholder(tf.float32,
name='keep_prob_hidden')
# Add to the graph each operation (including model definition)
_, logits = network.compute_loss(network.inputs, network.labels,
network.inputs_seq_len,
network.keep_prob_input,
network.keep_prob_hidden)
decode_op = network.decoder(logits,
network.inputs_seq_len,
decode_type='beam_search',
beam_width=20)
per_op = network.compute_ler(decode_op, network.labels)
# Create a saver for writing training checkpoints
saver = tf.train.Saver()
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(network.model_dir)
# If check point exists
if ckpt:
# Use last saved model
model_path = ckpt.model_checkpoint_path
if epoch is not None:
model_path = model_path.split('/')[:-1]
model_path = '/'.join(model_path) + '/model.ckpt-' + str(epoch)
saver.restore(sess, model_path)
print("Model restored: " + model_path)
else:
raise ValueError('There are not any checkpoints.')
if param['label_type'] in ['kana', 'kanji']:
print('=== eval1 Evaluation ===')
cer_eval1 = do_eval_cer(session=sess,
decode_op=decode_op,
network=network,
dataset=eval1_data,
label_type=param['label_type'],
is_test=True,
eval_batch_size=1,
is_progressbar=True)
print(' CER: %f %%' % (cer_eval1 * 100))
print('=== eval2 Evaluation ===')
cer_eval2 = do_eval_cer(session=sess,
decode_op=decode_op,
network=network,
dataset=eval2_data,
label_type=param['label_type'],
is_test=True,
eval_batch_size=1,
is_progressbar=True)
print(' CER: %f %%' % (cer_eval2 * 100))
print('=== eval3 Evaluation ===')
cer_eval3 = do_eval_cer(session=sess,
decode_op=decode_op,
network=network,
dataset=eval3_data,
label_type=param['label_type'],
is_test=True,
eval_batch_size=1,
is_progressbar=True)
print(' CER: %f %%' % (cer_eval3 * 100))
print('=== Mean ===')
cer_mean = (cer_eval1 + cer_eval2 + cer_eval3) / 3.
print(' CER: %f %%' % (cer_mean * 100))
else:
print('=== eval1 Evaluation ===')
per_eval1 = do_eval_per(session=sess,
per_op=per_op,
network=network,
dataset=eval1_data,
eval_batch_size=1,
is_progressbar=True)
print(' PER: %f %%' % (per_eval1 * 100))
print('=== eval2 Evaluation ===')
per_eval2 = do_eval_per(session=sess,
per_op=per_op,
network=network,
dataset=eval2_data,
eval_batch_size=1,
is_progressbar=True)
print(' PER: %f %%' % (per_eval2 * 100))
print('=== eval3 Evaluation ===')
per_eval3 = do_eval_per(session=sess,
per_op=per_op,
network=network,
dataset=eval3_data,
eval_batch_size=1,
is_progressbar=True)
print(' PER: %f %%' % (per_eval3 * 100))
print('=== Mean ===')
per_mean = (per_eval1 + per_eval2 + per_eval3) / 3.
print(' PER: %f %%' % 1 (per_mean * 100))
def do_eval(model, params, epoch, eval_batch_size, beam_width):
"""Evaluate the model.
Args:
model: the model to restore
params (dict): A dictionary of parameters
epoch (int): the epoch to restore
eval_batch_size (int): the size of mini-batch when evaluation
beam_width (int): beam_width (int, optional): beam width for beam search.
1 disables beam search, which mean greedy decoding.
"""
# Load dataset
eval1_data = Dataset(
data_type='eval1', train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'] if eval_batch_size == -
1 else eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=False)
eval2_data = Dataset(
data_type='eval2', train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'] if eval_batch_size == -
1 else eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=False)
eval3_data = Dataset(
data_type='eval3', train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'] if eval_batch_size == -
1 else eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=False)
with tf.name_scope('tower_gpu0'):
# Define placeholders
model.create_placeholders()
# Add to the graph each operation (including model definition)
_, logits = model.compute_loss(model.inputs_pl_list[0],
model.labels_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.keep_prob_pl_list[0])
decode_op = model.decoder(logits,
model.inputs_seq_len_pl_list[0],
beam_width=beam_width)
# Create a saver for writing training checkpoints
saver = tf.train.Saver()
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(model.save_path)
# If check point exists
if ckpt:
model_path = ckpt.model_checkpoint_path
if epoch != -1:
model_path = model_path.split('/')[:-1]
model_path = '/'.join(model_path) + '/model.ckpt-' + str(epoch)
saver.restore(sess, model_path)
print("Model restored: " + model_path)
else:
raise ValueError('There are not any checkpoints.')
print('Test Data Evaluation:')
cer_eval1 = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=eval1_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (eval1): %f %%' % (cer_eval1 * 100))
cer_eval2 = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=eval2_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (eval2): %f %%' % (cer_eval2 * 100))
cer_eval3 = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=eval3_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (eval3): %f %%' % (cer_eval3 * 100))
cer_mean = (cer_eval1 + cer_eval2 + cer_eval3) / 3.
print(' CER (mean): %f %%' % (cer_mean * 100))
def do_eval(model, params, epoch, eval_batch_size, beam_width):
"""Evaluate the model.
Args:
model: the model to restore
params (dict): A dictionary of parameters
epoch (int): the epoch to restore
eval_batch_size (int): the size of mini-batch when evaluation
beam_width (int): beam_width (int, optional): beam width for beam search.
1 disables beam search, which mean greedy decoding.
"""
# Load dataset
eval1_data = Dataset(data_type='eval1',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size']
if eval_batch_size == -1 else eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
shuffle=False)
eval2_data = Dataset(data_type='eval2',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size']
if eval_batch_size == -1 else eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
shuffle=False)
eval3_data = Dataset(data_type='eval3',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size']
if eval_batch_size == -1 else eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
shuffle=False)
with tf.name_scope('tower_gpu0'):
# Define placeholders
model.create_placeholders()
# Add to the graph each operation (including model definition)
_, logits = model.compute_loss(model.inputs_pl_list[0],
model.labels_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.keep_prob_pl_list[0])
decode_op = model.decoder(logits,
model.inputs_seq_len_pl_list[0],
beam_width=beam_width)
# Create a saver for writing training checkpoints
saver = tf.train.Saver()
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(model.save_path)
# If check point exists
if ckpt:
model_path = ckpt.model_checkpoint_path
if epoch != -1:
model_path = model_path.split('/')[:-1]
model_path = '/'.join(model_path) + '/model.ckpt-' + str(epoch)
saver.restore(sess, model_path)
print("Model restored: " + model_path)
else:
raise ValueError('There are not any checkpoints.')
print('Test Data Evaluation:')
cer_eval1 = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=eval1_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (eval1): %f %%' % (cer_eval1 * 100))
cer_eval2 = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=eval2_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (eval2): %f %%' % (cer_eval2 * 100))
cer_eval3 = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=eval3_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (eval3): %f %%' % (cer_eval3 * 100))
cer_mean = (cer_eval1 + cer_eval2 + cer_eval3) / 3.
print(' CER (mean): %f %%' % (cer_mean * 100))
def do_train(model, params, gpu_indices):
"""Run training.
Args:
model: the model to train
params (dict): A dictionary of parameters
gpu_indices (list): GPU indices
"""
# Load dataset
train_data = Dataset(
data_type='train', train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'], max_epoch=params['num_epoch'],
splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
sort_utt=True, sort_stop_epoch=params['sort_stop_epoch'],
num_gpu=len(gpu_indices))
dev_data = Dataset(
data_type='dev', train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'], splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
sort_utt=False, num_gpu=len(gpu_indices))
# Tell TensorFlow that the model will be built into the default graph
with tf.Graph().as_default(), tf.device('/cpu:0'):
# Create a variable to track the global step
global_step = tf.Variable(0, name='global_step', trainable=False)
# Set optimizer
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
optimizer = model._set_optimizer(
params['optimizer'], learning_rate_pl)
# Calculate the gradients for each model tower
total_grads_and_vars, total_losses = [], []
decode_ops, ler_ops = [], []
all_devices = ['/gpu:%d' % i_gpu for i_gpu in range(len(gpu_indices))]
# NOTE: /cpu:0 is prepared for evaluation
with tf.variable_scope(tf.get_variable_scope()):
for i_gpu in range(len(all_devices)):
with tf.device(all_devices[i_gpu]):
with tf.name_scope('tower_gpu%d' % i_gpu) as scope:
# Define placeholders in each tower
model.create_placeholders()
# Calculate the total loss for the current tower of the
# model. This function constructs the entire model but
# shares the variables across all towers.
tower_loss, tower_logits = model.compute_loss(
model.inputs_pl_list[i_gpu],
model.labels_pl_list[i_gpu],
model.inputs_seq_len_pl_list[i_gpu],
model.keep_prob_pl_list[i_gpu],
scope)
tower_loss = tf.expand_dims(tower_loss, axis=0)
total_losses.append(tower_loss)
# Reuse variables for the next tower
tf.get_variable_scope().reuse_variables()
# Calculate the gradients for the batch of data on this
# tower
tower_grads_and_vars = optimizer.compute_gradients(
tower_loss)
# Gradient clipping
tower_grads_and_vars = model._clip_gradients(
tower_grads_and_vars)
# TODO: Optionally add gradient noise
# Keep track of the gradients across all towers
total_grads_and_vars.append(tower_grads_and_vars)
# Add to the graph each operation per tower
decode_op_tower = model.decoder(
tower_logits,
model.inputs_seq_len_pl_list[i_gpu],
beam_width=params['beam_width'])
decode_ops.append(decode_op_tower)
ler_op_tower = model.compute_ler(
decode_op_tower, model.labels_pl_list[i_gpu])
ler_op_tower = tf.expand_dims(ler_op_tower, axis=0)
ler_ops.append(ler_op_tower)
# Aggregate losses, then calculate average loss
total_losses = tf.concat(axis=0, values=total_losses)
loss_op = tf.reduce_mean(total_losses, axis=0)
ler_ops = tf.concat(axis=0, values=ler_ops)
ler_op = tf.reduce_mean(ler_ops, axis=0)
# We must calculate the mean of each gradient. Note that this is the
# synchronization point across all towers
average_grads_and_vars = average_gradients(total_grads_and_vars)
# Apply the gradients to adjust the shared variables.
train_op = optimizer.apply_gradients(average_grads_and_vars,
global_step=global_step)
# Define learning rate controller
lr_controller = Controller(
learning_rate_init=params['learning_rate'],
decay_start_epoch=params['decay_start_epoch'],
decay_rate=params['decay_rate'],
decay_patient_epoch=params['decay_patient_epoch'],
lower_better=True)
# Build the summary tensor based on the TensorFlow collection of
# summaries
summary_train = tf.summary.merge(model.summaries_train)
summary_dev = tf.summary.merge(model.summaries_dev)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# Create a saver for writing training checkpoints
saver = tf.train.Saver(max_to_keep=None)
# Count total parameters
parameters_dict, total_parameters = count_total_parameters(
tf.trainable_variables())
for parameter_name in sorted(parameters_dict.keys()):
print("%s %d" % (parameter_name, parameters_dict[parameter_name]))
print("Total %d variables, %s M parameters" %
(len(parameters_dict.keys()),
"{:,}".format(total_parameters / 1000000)))
csv_steps, csv_loss_train, csv_loss_dev = [], [], []
csv_ler_train, csv_ler_dev = [], []
# Create a session for running operation on the graph
# NOTE: Start running operations on the Graph. allow_soft_placement
# must be set to True to build towers on GPU, as some of the ops do not
# have GPU implementations.
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)) as sess:
# Instantiate a SummaryWriter to output summaries and the graph
summary_writer = tf.summary.FileWriter(
model.save_path, sess.graph)
# Initialize parameters
sess.run(init_op)
# Train model
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
cer_dev_best = 1
not_improved_epoch = 0
learning_rate = float(params['learning_rate'])
for step, (data, is_new_epoch) in enumerate(train_data):
# Create feed dictionary for next mini batch (train)
inputs, labels, inputs_seq_len, _ = data
feed_dict_train = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_train[model.inputs_pl_list[i_gpu]
] = inputs[i_gpu]
feed_dict_train[model.labels_pl_list[i_gpu]] = list2sparsetensor(
labels[i_gpu], padded_value=train_data.padded_value)
feed_dict_train[model.inputs_seq_len_pl_list[i_gpu]
] = inputs_seq_len[i_gpu]
feed_dict_train[model.keep_prob_pl_list[i_gpu]
] = 1 - float(params['dropout'])
feed_dict_train[learning_rate_pl] = learning_rate
# Update parameters
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % int(params['print_step'] / len(gpu_indices)) == 0:
# Create feed dictionary for next mini batch (dev)
inputs, labels, inputs_seq_len, _ = dev_data.next()[0]
feed_dict_dev = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_dev[model.inputs_pl_list[i_gpu]
] = inputs[i_gpu]
feed_dict_dev[model.labels_pl_list[i_gpu]] = list2sparsetensor(
labels[i_gpu], padded_value=dev_data.padded_value)
feed_dict_dev[model.inputs_seq_len_pl_list[i_gpu]
] = inputs_seq_len[i_gpu]
feed_dict_dev[model.keep_prob_pl_list[i_gpu]] = 1.0
# Compute loss
loss_train = sess.run(loss_op, feed_dict=feed_dict_train)
loss_dev = sess.run(loss_op, feed_dict=feed_dict_dev)
csv_steps.append(step)
csv_loss_train.append(loss_train)
csv_loss_dev.append(loss_dev)
# Change to evaluation mode
for i_gpu in range(len(gpu_indices)):
feed_dict_train[model.keep_prob_pl_list[i_gpu]] = 1.0
# Compute accuracy & update event files
ler_train, summary_str_train = sess.run(
[ler_op, summary_train], feed_dict=feed_dict_train)
ler_dev, summary_str_dev = sess.run(
[ler_op, summary_dev], feed_dict=feed_dict_dev)
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
summary_writer.add_summary(summary_str_train, step + 1)
summary_writer.add_summary(summary_str_dev, step + 1)
summary_writer.flush()
duration_step = time.time() - start_time_step
print("Step %d (epoch: %.3f): loss = %.3f (%.3f) / ler = %.3f (%.3f) / lr = %.5f (%.3f min)" %
(step + 1, train_data.epoch_detail, loss_train, loss_dev, ler_train, ler_dev,
learning_rate, duration_step / 60))
sys.stdout.flush()
start_time_step = time.time()
# Save checkpoint and evaluate model per epoch
if is_new_epoch:
duration_epoch = time.time() - start_time_epoch
print('-----EPOCH:%d (%.3f min)-----' %
(train_data.epoch, duration_epoch / 60))
# Save fugure of loss & ler
plot_loss(csv_loss_train, csv_loss_dev, csv_steps,
save_path=model.save_path)
plot_ler(csv_ler_train, csv_ler_dev, csv_steps,
label_type=params['label_type'],
save_path=model.save_path)
if train_data.epoch >= params['eval_start_epoch']:
start_time_eval = time.time()
print('=== Dev Data Evaluation ===')
cer_dev_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
eval_batch_size=1)
print(' CER: %f %%' % (cer_dev_epoch * 100))
if cer_dev_epoch < cer_dev_best:
cer_dev_best = cer_dev_epoch
not_improved_epoch = 0
print('■■■ ↑Best Score (CER)↑ ■■■')
# Save model only (check point)
checkpoint_file = join(
model.save_path, 'model.ckpt')
save_path = saver.save(
sess, checkpoint_file, global_step=train_data.epoch)
print("Model saved in file: %s" % save_path)
else:
not_improved_epoch += 1
duration_eval = time.time() - start_time_eval
print('Evaluation time: %.3f min' %
(duration_eval / 60))
# Early stopping
if not_improved_epoch == params['not_improved_patient_epoch']:
break
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=train_data.epoch,
value=cer_dev_epoch)
start_time_epoch = time.time()
duration_train = time.time() - start_time_train
print('Total time: %.3f hour' % (duration_train / 3600))
# Training was finished correctly
with open(join(model.model_dir, 'complete.txt'), 'w') as f:
f.write('')
def do_train(model, params, gpu_indices):
"""Run training.
Args:
model: the model to train
params (dict): A dictionary of parameters
gpu_indices (list): GPU indices
"""
# Load dataset
train_data = Dataset(data_type='train',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'],
max_epoch=params['num_epoch'],
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=True,
sort_stop_epoch=params['sort_stop_epoch'],
num_gpu=len(gpu_indices))
dev_data = Dataset(data_type='dev',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=params['batch_size'],
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False,
num_gpu=len(gpu_indices))
# eval1_data = Dataset(
# data_type='eval1', train_data_size=params['train_data_size'],
# label_type=params['label_type'],
# batch_size=params['batch_size'], splice=params['splice'],
# num_stack=params['num_stack'], num_skip=params['num_skip'],
# sort_utt=False)
# eval2_data = Dataset(
# data_type='eval2', train_data_size=params['train_data_size'],
# label_type=params['label_type'],
# batch_size=params['batch_size'], splice=params['splice'],
# num_stack=params['num_stack'], num_skip=params['num_skip'],
# sort_utt=False)
# eval3_data = Dataset(
# data_type='eval3', train_data_size=params['train_data_size'],
# label_type=params['label_type'],
# batch_size=params['batch_size'], splice=params['splice'],
# num_stack=params['num_stack'], num_skip=params['num_skip'],
# sort_utt=False)
# Tell TensorFlow that the model will be built into the default graph
with tf.Graph().as_default(), tf.device('/cpu:0'):
# Create a variable to track the global step
global_step = tf.Variable(0, name='global_step', trainable=False)
# Set optimizer
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
optimizer = model._set_optimizer(params['optimizer'], learning_rate_pl)
# Calculate the gradients for each model tower
total_grads_and_vars, total_losses = [], []
decode_ops, ler_ops = [], []
all_devices = ['/gpu:%d' % i_gpu for i_gpu in range(len(gpu_indices))]
# NOTE: /cpu:0 is prepared for evaluation
with tf.variable_scope(tf.get_variable_scope()):
for i_gpu in range(len(all_devices)):
with tf.device(all_devices[i_gpu]):
with tf.name_scope('tower_gpu%d' % i_gpu) as scope:
# Define placeholders in each tower
model.create_placeholders()
# Calculate the total loss for the current tower of the
# model. This function constructs the entire model but
# shares the variables across all towers.
tower_loss, tower_logits = model.compute_loss(
model.inputs_pl_list[i_gpu],
model.labels_pl_list[i_gpu],
model.inputs_seq_len_pl_list[i_gpu],
model.keep_prob_pl_list[i_gpu], scope)
tower_loss = tf.expand_dims(tower_loss, axis=0)
total_losses.append(tower_loss)
# Reuse variables for the next tower
tf.get_variable_scope().reuse_variables()
# Calculate the gradients for the batch of data on this
# tower
tower_grads_and_vars = optimizer.compute_gradients(
tower_loss)
# Gradient clipping
tower_grads_and_vars = model._clip_gradients(
tower_grads_and_vars)
# TODO: Optionally add gradient noise
# Keep track of the gradients across all towers
total_grads_and_vars.append(tower_grads_and_vars)
# Add to the graph each operation per tower
decode_op_tower = model.decoder(
tower_logits,
model.inputs_seq_len_pl_list[i_gpu],
beam_width=params['beam_width'])
decode_ops.append(decode_op_tower)
ler_op_tower = model.compute_ler(
decode_op_tower, model.labels_pl_list[i_gpu])
ler_op_tower = tf.expand_dims(ler_op_tower, axis=0)
ler_ops.append(ler_op_tower)
# Aggregate losses, then calculate average loss
total_losses = tf.concat(axis=0, values=total_losses)
loss_op = tf.reduce_mean(total_losses, axis=0)
ler_ops = tf.concat(axis=0, values=ler_ops)
ler_op = tf.reduce_mean(ler_ops, axis=0)
# We must calculate the mean of each gradient. Note that this is the
# synchronization point across all towers
average_grads_and_vars = average_gradients(total_grads_and_vars)
# Apply the gradients to adjust the shared variables.
train_op = optimizer.apply_gradients(average_grads_and_vars,
global_step=global_step)
# Define learning rate controller
lr_controller = Controller(
learning_rate_init=params['learning_rate'],
decay_start_epoch=params['decay_start_epoch'],
decay_rate=params['decay_rate'],
decay_patient_epoch=params['decay_patient_epoch'],
lower_better=True)
# Build the summary tensor based on the TensorFlow collection of
# summaries
summary_train = tf.summary.merge(model.summaries_train)
summary_dev = tf.summary.merge(model.summaries_dev)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# Create a saver for writing training checkpoints
saver = tf.train.Saver(max_to_keep=None)
# Count total parameters
parameters_dict, total_parameters = count_total_parameters(
tf.trainable_variables())
for parameter_name in sorted(parameters_dict.keys()):
print("%s %d" % (parameter_name, parameters_dict[parameter_name]))
print("Total %d variables, %s M parameters" %
(len(parameters_dict.keys()), "{:,}".format(
total_parameters / 1000000)))
csv_steps, csv_loss_train, csv_loss_dev = [], [], []
csv_ler_train, csv_ler_dev = [], []
# Create a session for running operation on the graph
# NOTE: Start running operations on the Graph. allow_soft_placement
# must be set to True to build towers on GPU, as some of the ops do not
# have GPU implementations.
with tf.Session(
config=tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)) as sess:
# Instantiate a SummaryWriter to output summaries and the graph
summary_writer = tf.summary.FileWriter(model.save_path, sess.graph)
# Initialize parameters
sess.run(init_op)
# Train model
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
ler_dev_best = 1
not_improved_epoch = 0
learning_rate = float(params['learning_rate'])
for step, (data, is_new_epoch) in enumerate(train_data):
# Create feed dictionary for next mini batch (train)
inputs, labels, inputs_seq_len, _ = data
feed_dict_train = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_train[
model.inputs_pl_list[i_gpu]] = inputs[i_gpu]
feed_dict_train[
model.labels_pl_list[i_gpu]] = list2sparsetensor(
labels[i_gpu],
padded_value=train_data.padded_value)
feed_dict_train[model.inputs_seq_len_pl_list[
i_gpu]] = inputs_seq_len[i_gpu]
feed_dict_train[
model.keep_prob_pl_list[i_gpu]] = 1 - float(
params['dropout'])
feed_dict_train[learning_rate_pl] = learning_rate
# Update parameters
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % int(
params['print_step'] / len(gpu_indices)) == 0:
# Create feed dictionary for next mini batch (dev)
(inputs, labels, inputs_seq_len, _), _ = dev_data.next()
feed_dict_dev = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_dev[
model.inputs_pl_list[i_gpu]] = inputs[i_gpu]
feed_dict_dev[
model.labels_pl_list[i_gpu]] = list2sparsetensor(
labels[i_gpu],
padded_value=dev_data.padded_value)
feed_dict_dev[model.inputs_seq_len_pl_list[
i_gpu]] = inputs_seq_len[i_gpu]
feed_dict_dev[model.keep_prob_pl_list[i_gpu]] = 1.0
# Compute loss
loss_train = sess.run(loss_op, feed_dict=feed_dict_train)
loss_dev = sess.run(loss_op, feed_dict=feed_dict_dev)
csv_steps.append(step)
csv_loss_train.append(loss_train)
csv_loss_dev.append(loss_dev)
# Change to evaluation mode
for i_gpu in range(len(gpu_indices)):
feed_dict_train[model.keep_prob_pl_list[i_gpu]] = 1.0
# Compute accuracy & update event files
ler_train, summary_str_train = sess.run(
[ler_op, summary_train], feed_dict=feed_dict_train)
ler_dev, summary_str_dev = sess.run(
[ler_op, summary_dev], feed_dict=feed_dict_dev)
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
summary_writer.add_summary(summary_str_train, step + 1)
summary_writer.add_summary(summary_str_dev, step + 1)
summary_writer.flush()
duration_step = time.time() - start_time_step
print(
"Step %d (epoch: %.3f): loss = %.3f (%.3f) / ler = %.3f (%.3f) / lr = %.5f (%.3f min)"
% (step + 1, train_data.epoch_detail, loss_train,
loss_dev, ler_train, ler_dev, learning_rate,
duration_step / 60))
sys.stdout.flush()
start_time_step = time.time()
# Save checkpoint and evaluate model per epoch
if is_new_epoch:
duration_epoch = time.time() - start_time_epoch
print('-----EPOCH:%d (%.3f min)-----' %
(train_data.epoch, duration_epoch / 60))
# Save fugure of loss & ler
plot_loss(csv_loss_train,
csv_loss_dev,
csv_steps,
save_path=model.save_path)
plot_ler(csv_ler_train,
csv_ler_dev,
csv_steps,
label_type=params['label_type'],
save_path=model.save_path)
if train_data.epoch >= params['eval_start_epoch']:
start_time_eval = time.time()
print('=== Dev Data Evaluation ===')
# dev-clean
ler_dev_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
eval_batch_size=params['batch_size'])
print(' CER: %f %%' % (ler_dev_epoch * 100))
if ler_dev_epoch < ler_dev_best:
ler_dev_best = ler_dev_epoch
not_improved_epoch = 0
print('■■■ ↑Best Score (CER)↑ ■■■')
# Save model (check point)
checkpoint_file = join(model.save_path,
'model.ckpt')
save_path = saver.save(
sess,
checkpoint_file,
global_step=train_data.epoch)
print("Model saved in file: %s" % save_path)
else:
not_improved_epoch += 1
duration_eval = time.time() - start_time_eval
print('Evaluation time: %.3f min' %
(duration_eval / 60))
# Early stopping
if not_improved_epoch == params[
'not_improved_patient_epoch']:
break
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=train_data.epoch,
value=ler_dev_epoch)
start_time_epoch = time.time()
duration_train = time.time() - start_time_train
print('Total time: %.3f hour' % (duration_train / 3600))
# Training was finished correctly
with open(join(model.model_dir, 'complete.txt'), 'w') as f:
f.write('')
def do_train(model, params):
"""Run training.
Args:
model: model to train
params: A dictionary of parameters
"""
# Load dataset
train_data = Dataset(data_type='train',
label_type_main=params['label_type_main'],
label_type_sub=params['label_type_sub'],
train_data_size=params['train_data_size'],
batch_size=params['batch_size'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=True)
dev_data_step = Dataset(data_type='dev',
label_type_main=params['label_type_main'],
label_type_sub=params['label_type_sub'],
train_data_size=params['train_data_size'],
batch_size=params['batch_size'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False)
dev_data_epoch = Dataset(data_type='dev',
label_type_main=params['label_type_main'],
label_type_sub=params['label_type_sub'],
train_data_size=params['train_data_size'],
batch_size=params['batch_size'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False)
# Tell TensorFlow that the model will be built into the default graph
with tf.Graph().as_default():
# Define placeholders
model.create_placeholders(gpu_index=0)
# Add to the graph each operation
loss_op, logits_main, logits_sub = model.compute_loss(
model.inputs_pl_list[0],
model.labels_pl_list[0],
model.labels_sub_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.keep_prob_input_pl_list[0],
model.keep_prob_hidden_pl_list[0],
model.keep_prob_output_pl_list[0])
train_op = model.train(loss_op,
optimizer=params['optimizer'],
learning_rate=model.learning_rate_pl_list[0])
decode_op_main, decode_op_sub = model.decoder(
logits_main,
logits_sub,
model.inputs_seq_len_pl_list[0],
decode_type='beam_search',
beam_width=20)
ler_op_main, ler_op_sub = model.compute_ler(
decode_op_main, decode_op_sub,
model.labels_pl_list[0], model.labels_sub_pl_list[0])
# Define learning rate controller
lr_controller = Controller(
learning_rate_init=params['learning_rate'],
decay_start_epoch=params['decay_start_epoch'],
decay_rate=params['decay_rate'],
decay_patient_epoch=1,
lower_better=True)
# Build the summary tensor based on the TensorFlow collection of
# summaries
summary_train = tf.summary.merge(model.summaries_train)
summary_dev = tf.summary.merge(model.summaries_dev)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# Create a saver for writing training checkpoints
saver = tf.train.Saver(max_to_keep=None)
# Count total parameters
parameters_dict, total_parameters = count_total_parameters(
tf.trainable_variables())
for parameter_name in sorted(parameters_dict.keys()):
print("%s %d" % (parameter_name, parameters_dict[parameter_name]))
print("Total %d variables, %s M parameters" %
(len(parameters_dict.keys()),
"{:,}".format(total_parameters / 1000000)))
csv_steps, csv_loss_train, csv_loss_dev = [], [], []
csv_ler_main_train, csv_ler_main_dev = [], []
csv_ler_sub_train, csv_ler_sub_dev = [], []
# Create a session for running operation on the graph
with tf.Session() as sess:
# Instantiate a SummaryWriter to output summaries and the graph
summary_writer = tf.summary.FileWriter(
model.save_path, sess.graph)
# Initialize parameters
sess.run(init_op)
# Make mini-batch generator
mini_batch_train = train_data.next_batch()
mini_batch_dev = dev_data_step.next_batch()
# Train model
iter_per_epoch = int(train_data.data_num / params['batch_size'])
train_step = train_data.data_num / params['batch_size']
if (train_step) != int(train_step):
iter_per_epoch += 1
max_steps = iter_per_epoch * params['num_epoch']
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
ler_main_dev_best = 1
learning_rate = float(params['learning_rate'])
for step in range(max_steps):
# Create feed dictionary for next mini batch (train)
inputs, labels_main, labels_sub, inputs_seq_len, _ = mini_batch_train.__next__()
feed_dict_train = {
model.inputs_pl_list[0]: inputs,
model.labels_pl_list[0]: list2sparsetensor(labels_main, padded_value=-1),
model.labels_sub_pl_list[0]: list2sparsetensor(labels_sub, padded_value=-1),
model.inputs_seq_len_pl_list[0]: inputs_seq_len,
model.keep_prob_input_pl_list[0]: model.dropout_ratio_input,
model.keep_prob_hidden_pl_list[0]: model.dropout_ratio_hidden,
model.keep_prob_output_pl_list[0]: model.dropout_ratio_output,
model.learning_rate_pl_list[0]: learning_rate
}
# Update parameters
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % 200 == 0:
# Create feed dictionary for next mini batch (dev)
inputs, labels_main, labels_sub, inputs_seq_len, _ = mini_batch_dev.__next__()
feed_dict_dev = {
model.inputs_pl_list[0]: inputs,
model.labels_pl_list[0]: list2sparsetensor(labels_main, padded_value=-1),
model.labels_sub_pl_list[0]: list2sparsetensor(labels_sub, padded_value=-1),
model.inputs_seq_len_pl_list[0]: inputs_seq_len,
model.keep_prob_input_pl_list[0]: 1.0,
model.keep_prob_hidden_pl_list[0]: 1.0,
model.keep_prob_output_pl_list[0]: 1.0
}
# Compute loss
loss_train = sess.run(loss_op, feed_dict=feed_dict_train)
loss_dev = sess.run(loss_op, feed_dict=feed_dict_dev)
csv_steps.append(step)
csv_loss_train.append(loss_train)
csv_loss_dev.append(loss_dev)
# Change to evaluation mode
feed_dict_train[model.keep_prob_input_pl_list[0]] = 1.0
feed_dict_train[model.keep_prob_hidden_pl_list[0]] = 1.0
feed_dict_train[model.keep_prob_output_pl_list[0]] = 1.0
# Compute accuracy & update event file
ler_main_train, ler_sub_train, summary_str_train = sess.run(
[ler_op_main, ler_op_sub, summary_train],
feed_dict=feed_dict_train)
ler_main_dev, ler_sub_dev, summary_str_dev = sess.run(
[ler_op_main, ler_op_sub, summary_dev],
feed_dict=feed_dict_dev)
csv_ler_main_train.append(ler_main_train)
csv_ler_main_dev.append(ler_main_dev)
csv_ler_sub_train.append(ler_sub_train)
csv_ler_sub_dev.append(ler_sub_dev)
summary_writer.add_summary(summary_str_train, step + 1)
summary_writer.add_summary(summary_str_dev, step + 1)
summary_writer.flush()
duration_step = time.time() - start_time_step
print('Step %d: loss = %.3f (%.3f) / ler_main = %.4f (%.4f) / ler_sub = %.4f (%.4f) (%.3f min)' %
(step + 1, loss_train, loss_dev, ler_main_train, ler_main_dev,
ler_sub_train, ler_sub_dev, duration_step / 60))
sys.stdout.flush()
start_time_step = time.time()
# Save checkpoint and evaluate model per epoch
if (step + 1) % iter_per_epoch == 0 or (step + 1) == max_steps:
duration_epoch = time.time() - start_time_epoch
epoch = (step + 1) // iter_per_epoch
print('-----EPOCH:%d (%.3f min)-----' %
(epoch, duration_epoch / 60))
# Save model (check point)
checkpoint_file = join(model.save_path, 'model.ckpt')
save_path = saver.save(
sess, checkpoint_file, global_step=epoch)
print("Model saved in file: %s" % save_path)
if epoch >= 5:
start_time_eval = time.time()
print('=== Dev Evaluation ===')
ler_main_dev_epoch = do_eval_cer(
session=sess,
decode_op=decode_op_main,
model=model,
dataset=dev_data_epoch,
label_type=params['label_type_main'],
eval_batch_size=params['batch_size'],
is_multitask=True,
is_main=True)
print(' CER (main): %f %%' %
(ler_main_dev_epoch * 100))
ler_sub_dev_epoch = do_eval_cer(
session=sess,
decode_op=decode_op_sub,
model=model,
dataset=dev_data_epoch,
label_type=params['label_type_sub'],
eval_batch_size=params['batch_size'],
is_multitask=True,
is_main=False)
print(' CER (sub): %f %%' %
(ler_sub_dev_epoch * 100))
if ler_main_dev_epoch < ler_main_dev_best:
ler_main_dev_best = ler_main_dev_epoch
print('■■■ ↑Best Score (CER main)↑ ■■■')
duration_eval = time.time() - start_time_eval
print('Evaluation time: %.3f min' %
(duration_eval / 60))
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=epoch,
value=ler_main_dev_epoch)
start_time_epoch = time.time()
start_time_step = time.time()
duration_train = time.time() - start_time_train
print('Total time: %.3f hour' % (duration_train / 3600))
# Save train & dev loss, ler
save_loss(csv_steps, csv_loss_train, csv_loss_dev,
save_path=model.save_path)
save_ler(csv_steps, csv_ler_main_train, csv_ler_sub_dev,
save_path=model.save_path)
save_ler(csv_steps, csv_ler_sub_train, csv_ler_sub_dev,
save_path=model.save_path)
# Training was finished correctly
with open(join(model.save_path, 'complete.txt'), 'w') as f:
f.write('')