def do_eval(model, params, epoch, beam_width, eval_batch_size, temperature):
"""Evaluate the model.
Args:
model: the model to restore
params (dict): A dictionary of parameters
epoch (int): the epoch to restore
beam_width (int): beam width for beam search.
1 disables beam search, which mean greedy decoding.
eval_batch_size (int): the size of mini-batch when evaluation
temperature (int):
"""
if 'temp1' in params['teacher_model_path']:
teacher_train_temperature = 1
elif 'temp2' in params['teacher_model_path']:
teacher_train_temperature = 2
print('=' * 40)
print(' frame stack %d' % int(params['num_stack']))
print(' splice %d' % int(params['splice']))
print(' beam width: %d' % beam_width)
print(' temperature (teacher, training): %d' % teacher_train_temperature)
print(' temperature (teacher, inference): %d' %
params['teacher_temperature'])
print(' temperature (training): %d' % params['student_temperature'])
print(' temperature (inference): %d' % temperature)
print('=' * 40)
# Load dataset
test_clean_data = Dataset(data_type='test_clean',
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)
test_other_data = Dataset(data_type='test_other',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
shuffle=False)
with tf.name_scope('tower_gpu0') as scope:
# Define placeholders
model.create_placeholders_ctc()
# Add to the graph each operation (including model definition)
_, logits = model.compute_ctc_loss(
model.inputs_pl_list[0],
model.labels_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.keep_prob_pl_list[0],
scope,
softmax_temperature=temperature, # this is for training
# is_training=False)
is_training=True)
logits /= temperature
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_clean_test, wer_clean_test = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=test_clean_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (clean): %f %%' % (cer_clean_test * 100))
print(' WER (clean): %f %%' % (wer_clean_test * 100))
cer_other_test, wer_other_test = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=test_other_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=eval_batch_size,
progressbar=True)
print(' CER (other): %f %%' % (cer_other_test * 100))
print(' WER (other): %f %%' % (wer_other_test * 100))
def check_loading(self, label_type, data_type='dev_clean',
shuffle=False, sort_utt=False, sort_stop_epoch=None,
frame_stacking=False, splice=1, num_gpu=1):
print('========================================')
print(' label_type: %s' % label_type)
print(' data_type: %s' % data_type)
print(' shuffle: %s' % str(shuffle))
print(' sort_utt: %s' % str(sort_utt))
print(' sort_stop_epoch: %s' % str(sort_stop_epoch))
print(' frame_stacking: %s' % str(frame_stacking))
print(' splice: %d' % splice)
print(' num_gpu: %d' % num_gpu)
print('========================================')
num_stack = 3 if frame_stacking else 1
num_skip = 3 if frame_stacking else 1
dataset = Dataset(
data_type=data_type, train_data_size='train_clean100',
label_type=label_type,
batch_size=64, max_epoch=1, splice=splice,
num_stack=num_stack, num_skip=num_skip,
shuffle=shuffle, sort_utt=sort_utt, sort_stop_epoch=sort_stop_epoch,
progressbar=True, num_gpu=num_gpu)
print('=> Loading mini-batch...')
if label_type == 'character':
map_file_path = '../../metrics/mapping_files/ctc/character.txt'
elif label_type == 'character_capital_divide':
map_file_path = '../../metrics/mapping_files/ctc/character_capital.txt'
elif label_type == 'word':
map_file_path = '../../metrics/mapping_files/ctc/word_' + \
dataset.train_data_size + '.txt'
idx2char = Idx2char(map_file_path)
idx2word = Idx2word(map_file_path)
for data, is_new_epoch in dataset:
inputs, labels, inputs_seq_len, input_names = data
if not self.length_check:
for i, l in zip(inputs[0], labels[0]):
if len(i) < len(l):
raise ValueError(
'input length must be longer than label length.')
self.length_check = True
if num_gpu > 1:
for inputs_gpu in inputs:
print(inputs_gpu.shape)
if label_type == 'word':
if 'test' not in data_type:
str_true = ' '.join(idx2word(labels[0][0]))
else:
word_list = np.delete(labels[0][0], np.where(
labels[0][0] == None), axis=0)
str_true = ' '.join(word_list)
else:
str_true = idx2char(labels[0][0])
str_true = re.sub(r'_', ' ', str_true)
print('----- %s (epoch: %.3f) -----' %
(input_names[0][0], dataset.epoch_detail))
print(inputs[0].shape)
print(str_true)
if dataset.epoch_detail >= 0.05:
break
def do_decode(model, params, epoch, beam_width):
"""Decode the CTC outputs.
Args:
model: the model to restore
params (dict): A dictionary of parameters
epoch (int): the epoch to restore
beam_width (int): beam_width (int, optional): beam width for beam search.
1 disables beam search, which mean greedy decoding.
"""
# Load dataset
test_clean_data = Dataset(data_type='test_clean',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=1,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
shuffle=False)
test_other_data = Dataset(data_type='test_other',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=1,
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_input_pl_list[0],
model.keep_prob_hidden_pl_list[0],
model.keep_prob_output_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:
# Use last saved model
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.')
# Visualize
decode_test(session=sess,
decode_op=decode_op,
model=model,
dataset=test_clean_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
save_path=None)
# save_path=model.save_path)
decode_test(session=sess,
decode_op=decode_op,
model=model,
dataset=test_other_data,
label_type=params['label_type'],
train_data_size=params['train_data_size'],
save_path=None)
def do_eval(save_paths, params, beam_width, temperature_infer,
result_save_path):
"""Evaluate the model.
Args:
save_paths (list):
params (dict): A dictionary of parameters
epoch_list (list): list of the epoch to restore
beam_width (int): beam width for beam search.
1 disables beam search, which mean greedy decoding.
eval_batch_size (int): the size of mini-batch when evaluation
temperature_infer (int): temperature in the inference stage
result_save_path (string, optional):
"""
if 'temp1' in save_paths[0]:
temperature_train = 1
elif 'temp2' in save_paths[0]:
temperature_train = 2
else:
raise ValueError
if result_save_path is not None:
sys.stdout = open(
join(
result_save_path,
'4models_traintemp' + str(temperature_train) + '_inftemp' +
str(temperature_infer) + '.log'), 'w')
print('=' * 30)
print(' frame stack %d' % int(params['num_stack']))
print(' beam width: %d' % beam_width)
print(' ensemble: %d' % len(save_paths))
print(' temperature (training): %d' % temperature_train)
print(' temperature (inference): %d' % temperature_infer)
print('=' * 30)
# Load dataset
test_clean_data = Dataset(data_type='test_clean',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=1,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=True)
test_other_data = Dataset(data_type='test_other',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=1,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=True)
print('Test Data Evaluation:')
cer_clean_test, wer_clean_test = do_eval_cer(
save_paths=save_paths,
dataset=test_clean_data,
data_type='test_clean',
label_type=params['label_type'],
num_classes=params['num_classes'] + 1,
beam_width=beam_width,
temperature_infer=temperature_infer,
is_test=True,
progressbar=True)
print(' CER (clean): %f %%' % (cer_clean_test * 100))
print(' WER (clean): %f %%' % (wer_clean_test * 100))
cer_other_test, wer_other_test = do_eval_cer(
save_paths=save_paths,
dataset=test_other_data,
data_type='test_other',
label_type=params['label_type'],
num_classes=params['num_classes'] + 1,
beam_width=beam_width,
temperature_infer=temperature_infer,
is_test=True,
progressbar=True)
print(' CER (other): %f %%' % (cer_other_test * 100))
print(' WER (other): %f %%' % (wer_other_test * 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
test_clean_data = Dataset(data_type='test_clean',
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)
test_other_data = Dataset(data_type='test_other',
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],
is_training=False)
decode_op = model.decoder(logits,
model.inputs_seq_len_pl_list[0],
beam_width=beam_width)
posteriors_op = model.posteriors(logits)
# 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:')
if 'char' in params['label_type']:
if DECODER_TYPE == 1:
cer_clean_test, wer_clean_test = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=test_clean_data,
label_type=params['label_type'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' WER (clean): %f %%' % (wer_clean_test * 100))
print(' CER (clean): %f %%' % (cer_clean_test * 100))
cer_other_test, wer_other_test = do_eval_cer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=test_other_data,
label_type=params['label_type'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' WER (other): %f %%' % (wer_other_test * 100))
print(' CER (other): %f %%' % (cer_other_test * 100))
elif DECODER_TYPE == 2:
cer_clean_test, wer_clean_test = do_eval_cer2(
session=sess,
# beam_width=beam_width,
beam_width=20,
posteriors_ops=[posteriors_op],
model=model,
dataset=test_clean_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=20,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' WER (clean): %f %%' % (wer_clean_test * 100))
print(' CER (clean): %f %%' % (cer_clean_test * 100))
cer_other_test, wer_other_test = do_eval_cer2(
session=sess,
# beam_width=beam_width,
beam_width=20,
posteriors_ops=[posteriors_op],
model=model,
dataset=test_other_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=20,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' WER (other): %f %%' % (wer_other_test * 100))
print(' CER (other): %f %%' % (cer_other_test * 100))
else:
wer_clean_test = do_eval_wer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=test_clean_data,
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' WER (clean): %f %%' % (wer_clean_test * 100))
wer_other_test = do_eval_wer(
session=sess,
decode_ops=[decode_op],
model=model,
dataset=test_other_data,
train_data_size=params['train_data_size'],
is_test=True,
# eval_batch_size=eval_batch_size,
progressbar=True)
print(' WER (other): %f %%' % (wer_other_test * 100))
def check(self,
label_type,
data_type='dev_clean',
shuffle=False,
sort_utt=False,
sort_stop_epoch=None,
frame_stacking=False,
splice=1,
num_gpu=1):
print('========================================')
print(' label_type: %s' % label_type)
print(' data_type: %s' % data_type)
print(' shuffle: %s' % str(shuffle))
print(' sort_utt: %s' % str(sort_utt))
print(' sort_stop_epoch: %s' % str(sort_stop_epoch))
print(' frame_stacking: %s' % str(frame_stacking))
print(' splice: %d' % splice)
print(' num_gpu: %d' % num_gpu)
print('========================================')
num_stack = 3 if frame_stacking else 1
num_skip = 3 if frame_stacking else 1
dataset = Dataset(data_type=data_type,
train_data_size='train100h',
label_type=label_type,
batch_size=64,
max_epoch=2,
splice=splice,
num_stack=num_stack,
num_skip=num_skip,
shuffle=shuffle,
sort_utt=sort_utt,
sort_stop_epoch=sort_stop_epoch,
progressbar=True,
num_gpu=num_gpu)
print('=> Loading mini-batch...')
if label_type == 'character':
map_file_path = '../../metrics/mapping_files/character.txt'
else:
map_file_path = '../../metrics/mapping_files/' + label_type + '_' + \
dataset.train_data_size + '.txt'
idx2char = Idx2char(map_file_path)
idx2word = Idx2word(map_file_path)
for data, is_new_epoch in dataset:
inputs, labels, inputs_seq_len, input_names = data
if data_type == 'train':
for i, l in zip(inputs[0], labels[0]):
if len(i) < len(l):
raise ValueError(
'input length must be longer than label length.')
if num_gpu > 1:
for inputs_gpu in inputs:
print(inputs_gpu.shape)
if 'test' in data_type:
str_true = labels[0][0][0]
else:
if 'word' in label_type:
str_true = '_'.join(idx2word(labels[0][0]))
else:
str_true = idx2char(labels[0][0])
print('----- %s (epoch: %.3f) -----' %
(input_names[0][0], dataset.epoch_detail))
print(inputs[0].shape)
print(str_true)
if dataset.epoch_detail >= 0.1:
break
def do_train(model, params, gpu_indices):
"""Run CTC 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_clean = Dataset(data_type='dev_clean',
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))
dev_data_other = Dataset(data_type='dev_other',
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_input_pl_list[i_gpu],
model.keep_prob_hidden_pl_list[i_gpu],
model.keep_prob_output_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, _clip_norm=False)
# 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
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_input_pl_list[
i_gpu]] = params['dropout_input']
feed_dict_train[model.keep_prob_hidden_pl_list[
i_gpu]] = params['dropout_hidden']
feed_dict_train[model.keep_prob_output_pl_list[
i_gpu]] = params['dropout_output']
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_other.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_other.padded_value)
feed_dict_dev[model.inputs_seq_len_pl_list[
i_gpu]] = inputs_seq_len[i_gpu]
feed_dict_dev[
model.keep_prob_input_pl_list[i_gpu]] = 1.0
feed_dict_dev[
model.keep_prob_hidden_pl_list[i_gpu]] = 1.0
feed_dict_dev[
model.keep_prob_output_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_input_pl_list[i_gpu]] = 1.0
feed_dict_train[
model.keep_prob_hidden_pl_list[i_gpu]] = 1.0
feed_dict_train[
model.keep_prob_output_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)
# 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)
if train_data.epoch >= params['eval_start_epoch']:
start_time_eval = time.time()
if params['label_type'] != 'word':
print('=== Dev Data Evaluation ===')
# Dev-clean
ler_dev_clean_epoch, wer_dev_clean_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_data_clean,
label_type=params['label_type'],
eval_batch_size=params['batch_size'])
print(' CER (clean): %f %%' %
(ler_dev_clean_epoch * 100))
print(' WER (clean): %f %%' %
(wer_dev_clean_epoch * 100))
# Dev-other
ler_dev_other_epoch, wer_dev_other_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_data_other,
label_type=params['label_type'],
eval_batch_size=params['batch_size'])
print(' CER (other): %f %%' %
(ler_dev_other_epoch * 100))
print(' WER (other): %f %%' %
(wer_dev_other_epoch * 100))
if ler_dev_other_epoch < ler_dev_best:
ler_dev_best = ler_dev_other_epoch
print('■■■ ↑Best Score (CER)↑ ■■■')
else:
print('=== Dev Data Evaluation ===')
# Dev-clean
ler_dev_clean_epoch = do_eval_wer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_data_clean,
train_data_size=params['train_data_size'],
eval_batch_size=params['batch_size'])
print(' WER (clean): %f %%' %
(ler_dev_clean_epoch * 100))
# Dev-other
ler_dev_other_epoch = do_eval_wer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_data_other,
train_data_size=params['train_data_size'],
eval_batch_size=params['batch_size'])
print(' WER (other): %f %%' %
(ler_dev_other_epoch * 100))
if ler_dev_other_epoch < ler_dev_best:
ler_dev_best = ler_dev_other_epoch
print('■■■ ↑Best Score (WER)↑ ■■■')
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=train_data.epoch,
value=ler_dev_other_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.save_path, 'complete.txt'), 'w') as f:
f.write('')
def do_save(model, params, epoch, eval_batch_size):
"""Save the CTC outputs.
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 in evaluation
"""
# Load dataset
train_data = Dataset(data_type='train',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
batch_size=eval_batch_size,
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=True)
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_input_pl_list[0],
model.keep_prob_hidden_pl_list[0],
model.keep_prob_output_pl_list[0],
softmax_temperature=params['softmax_temperature'])
posteriors_op = model.posteriors(logits, blank_prior=1)
# 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:
# Use last saved model
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.')
for data, is_new_epoch in train_data:
# Create feed dictionary for next mini batch
inputs, _, inputs_seq_len, input_names = data
feed_dict = {
model.inputs_pl_list[0]: inputs[0],
model.inputs_seq_len_pl_list[0]: inputs_seq_len[0],
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
}
batch_size, max_frame_num = inputs[0].shape[:2]
posteriors = sess.run(posteriors_op, feed_dict=feed_dict)
posteriors = posteriors.reshape(-1, max_frame_num,
model.num_classes)
for i_batch in range(batch_size):
prob = posteriors[i_batch][:int(inputs_seq_len[0][i_batch]), :]
# Save as a npy file
np.save(
mkdir_join(model.save_path, 'probs',
input_names[0][i_batch]), prob)
if is_new_epoch:
break
def do_save(model, params, epoch, eval_batch_size, temperature):
"""Save the CTC outputs.
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 in evaluation
temperature (int):
"""
print('=' * 30)
print(' frame stack %d' % int(params['num_stack']))
print(' splice %d' % int(params['splice']))
print(' temperature (training): %d' % temperature)
print('=' * 30)
# 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'] if eval_batch_size == -
1 else eval_batch_size,
max_epoch=3, splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=True, num_gpu=1)
dev_clean_data = Dataset(
data_type='dev_clean', 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,
max_epoch=3, splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=True, num_gpu=1)
dev_other_data = Dataset(
data_type='dev_other', 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,
max_epoch=3, splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=True, num_gpu=1)
test_clean_data = Dataset(
data_type='test_clean', 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,
max_epoch=3, splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=True, num_gpu=1)
test_other_data = Dataset(
data_type='test_other', 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,
max_epoch=3, splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=True, num_gpu=1)
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])
logits /= temperature
posteriors_op = model.posteriors(logits, blank_prior=1)
# 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.')
#########################
# Save soft targets
#########################
# train100h
# save(session=sess,
# posteriors_op=posteriors_op,
# model=model,
# dataset=train_data,
# data_type='train',
# num_stack=params['num_stack'],
# save_prob=False,
# save_soft_targets=True,
# save_path=mkdir_join(model.save_path, 'temp' + str(temperature), 'train'))
# dev
# save(session=sess,
# posteriors_op=posteriors_op,
# model=model,
# dataset=dev_clean_data,
# data_type='dev_clean',
# num_stack=params['num_stack'],
# save_prob=False,
# save_soft_targets=True,
# save_path=mkdir_join(model.save_path, 'temp' + str(temperature), 'dev_clean'))
# save(session=sess,
# posteriors_op=posteriors_op,
# model=model,
# dataset=dev_other_data,
# data_type='dev_other',
# num_stack=params['num_stack'],
# save_prob=False,
# save_soft_targets=True,
# save_path=mkdir_join(model.save_path, 'temp' + str(temperature), 'dev_other'))
# test
save(session=sess,
posteriors_op=posteriors_op,
model=model,
dataset=test_clean_data,
data_type='test_clean',
num_stack=params['num_stack'],
save_prob=True,
save_soft_targets=False,
save_path=mkdir_join(model.save_path, 'temp' + str(temperature), 'test_clean'))
save(session=sess,
posteriors_op=posteriors_op,
model=model,
dataset=test_other_data,
data_type='test_other',
num_stack=params['num_stack'],
save_prob=True,
save_soft_targets=False,
save_path=mkdir_join(model.save_path, 'temp' + str(temperature), 'test_other'))
def do_train(model, params, gpu_indices):
"""Run CTC 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_clean_data = Dataset(
data_type='dev_clean', 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))
dev_other_data = Dataset(
data_type='dev_other', 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))
test_clean_data = Dataset(
data_type='test_clean', 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)
test_other_data = Dataset(
data_type='test_other', 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,
softmax_temperature=params['softmax_temperature'])
# NOTE: tower_logits have NOT been divided by
# softmax_temperature
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)
if params['train_data_size'] in ['train100h', 'train460h']:
inputs, labels, inputs_seq_len, _ = dev_clean_data.next()[
0]
else:
inputs, labels, inputs_seq_len, _ = dev_other_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_other_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
cer_dev_clean_epoch, wer_dev_clean_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_clean_data,
label_type=params['label_type'],
eval_batch_size=params['batch_size'])
print(' CER (clean): %f %%' %
(cer_dev_clean_epoch * 100))
print(' WER (clean): %f %%' %
(wer_dev_clean_epoch * 100))
# dev-other
cer_dev_other_epoch, wer_dev_other_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=dev_other_data,
label_type=params['label_type'],
eval_batch_size=params['batch_size'])
print(' CER (other): %f %%' %
(cer_dev_other_epoch * 100))
print(' WER (other): %f %%' %
(wer_dev_other_epoch * 100))
if params['train_data_size'] in ['train100h', 'train460h']:
metric_epoch = cer_dev_clean_epoch
else:
metric_epoch = cer_dev_other_epoch
if metric_epoch < ler_dev_best:
ler_dev_best = metric_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)
print('=== Test Data Evaluation ===')
# test-clean
cer_test_clean_epoch, wer_test_clean_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=test_clean_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=params['batch_size'])
print(' CER (clean): %f %%' %
(cer_test_clean_epoch * 100))
print(' WER (clean): %f %%' %
(wer_test_clean_epoch * 100))
# test-other
cer_test_other_epoch, wer_test_other_epoch = do_eval_cer(
session=sess,
decode_ops=decode_ops,
model=model,
dataset=test_other_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=params['batch_size'])
print(' CER (other): %f %%' %
(cer_test_other_epoch * 100))
print(' WER (other): %f %%' %
(wer_test_other_epoch * 100))
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=metric_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.save_path, 'complete.txt'), 'w') as f:
f.write('')
def do_plot(model, params, epoch, eval_batch_size):
"""Plot the CTC posteriors.
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 in evaluation
"""
# Load dataset
test_clean_data = Dataset(
data_type='test_clean',
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'],
sort_utt=True)
test_other_data = Dataset(
data_type='test_other',
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_hidden_pl_list[0],
# softmax_temperature=params['softmax_temperature'])
softmax_temperature=10)
posteriors_op = model.posteriors(logits, blank_prior=1)
# 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:
# Use last saved model
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.')
plot(session=sess,
posteriors_op=posteriors_op,
model=model,
dataset=test_clean_data,
label_type=params['label_type'],
num_stack=params['num_stack'],
# save_path=None)
save_path=mkdir_join(model.save_path, 'ctc_output', 'test-clean'))
plot(session=sess,
posteriors_op=posteriors_op,
model=model,
dataset=test_other_data,
label_type=params['label_type'],
num_stack=params['num_stack'],
# save_path=None)
save_path=mkdir_join(model.save_path, 'ctc_output', 'test-other'))