def check(self, encoder_type, attention_type, label_type='character'):
print('==================================================')
print(' encoder_type: %s' % encoder_type)
print(' attention_type: %s' % attention_type)
print(' label_type: %s' % label_type)
print('==================================================')
tf.reset_default_graph()
with tf.Graph().as_default():
# Load batch data
batch_size = 4
inputs, labels, inputs_seq_len, labels_seq_len = generate_data(
label_type=label_type,
model='attention',
batch_size=batch_size)
# Define model graph
num_classes = 27 if label_type == 'character' else 61
model = AttentionSeq2Seq(input_size=inputs[0].shape[1],
encoder_type=encoder_type,
encoder_num_units=256,
encoder_num_layers=2,
encoder_num_proj=None,
attention_type=attention_type,
attention_dim=128,
decoder_type='lstm',
decoder_num_units=256,
decoder_num_layers=1,
embedding_dim=64,
num_classes=num_classes,
sos_index=num_classes,
eos_index=num_classes + 1,
max_decode_length=100,
use_peephole=True,
splice=1,
parameter_init=0.1,
clip_grad_norm=5.0,
clip_activation_encoder=50,
clip_activation_decoder=50,
weight_decay=1e-8,
time_major=True,
sharpening_factor=1.0,
logits_temperature=1.0)
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation
loss_op, logits, decoder_outputs_train, decoder_outputs_infer = model.compute_loss(
model.inputs_pl_list[0],
model.labels_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.labels_seq_len_pl_list[0],
model.keep_prob_encoder_pl_list[0],
model.keep_prob_decoder_pl_list[0],
model.keep_prob_embedding_pl_list[0])
train_op = model.train(loss_op,
optimizer='adam',
learning_rate=learning_rate_pl)
decode_op_train, decode_op_infer = model.decode(
decoder_outputs_train, decoder_outputs_infer)
ler_op = model.compute_ler(model.labels_st_true_pl,
model.labels_st_pred_pl)
# Define learning rate controller
learning_rate = 1e-3
lr_controller = Controller(learning_rate_init=learning_rate,
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# 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)))
# Make feed dict
feed_dict = {
model.inputs_pl_list[0]: inputs,
model.labels_pl_list[0]: labels,
model.inputs_seq_len_pl_list[0]: inputs_seq_len,
model.labels_seq_len_pl_list[0]: labels_seq_len,
model.keep_prob_encoder_pl_list[0]: 0.8,
model.keep_prob_decoder_pl_list[0]: 1.0,
model.keep_prob_embedding_pl_list[0]: 1.0,
learning_rate_pl: learning_rate
}
idx2phone = Idx2phone(map_file_path='./phone61.txt')
with tf.Session() as sess:
# Initialize parameters
sess.run(init_op)
# Wrapper for tfdbg
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Train model
max_steps = 1000
start_time_step = time.time()
for step in range(max_steps):
# Compute loss
_, loss_train = sess.run(
[train_op, loss_op], feed_dict=feed_dict)
# Gradient check
# grads = sess.run(model.clipped_grads,
# feed_dict=feed_dict)
# for grad in grads:
# print(np.max(grad))
if (step + 1) % 10 == 0:
# Change to evaluation mode
feed_dict[model.keep_prob_encoder_pl_list[0]] = 1.0
feed_dict[model.keep_prob_decoder_pl_list[0]] = 1.0
feed_dict[model.keep_prob_embedding_pl_list[0]] = 1.0
# Predict class ids
predicted_ids_train, predicted_ids_infer = sess.run(
[decode_op_train, decode_op_infer],
feed_dict=feed_dict)
# Compute accuracy
try:
feed_dict_ler = {
model.labels_st_true_pl: list2sparsetensor(
labels, padded_value=model.eos_index),
model.labels_st_pred_pl: list2sparsetensor(
predicted_ids_infer, padded_value=model.eos_index)
}
ler_train = sess.run(
ler_op, feed_dict=feed_dict_ler)
except IndexError:
ler_train = 1
duration_step = time.time() - start_time_step
print('Step %d: loss = %.3f / ler = %.3f (%.3f sec) / lr = %.5f' %
(step + 1, loss_train, ler_train, duration_step, learning_rate))
start_time_step = time.time()
# Visualize
if label_type == 'character':
print('True : %s' %
idx2alpha(labels[0]))
print('Pred (Training) : <%s' %
idx2alpha(predicted_ids_train[0]))
print('Pred (Inference): <%s' %
idx2alpha(predicted_ids_infer[0]))
else:
print('True : %s' %
idx2phone(labels[0]))
print('Pred (Training) : < %s' %
idx2phone(predicted_ids_train[0]))
print('Pred (Inference): < %s' %
idx2phone(predicted_ids_infer[0]))
if ler_train < 0.1:
print('Model is Converged.')
break
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=step,
value=ler_train)
feed_dict[learning_rate_pl] = learning_rate
def main():
args = parser.parse_args()
##################################################
# DATSET
##################################################
if args.model_save_path is not None:
# Load a config file (.yml)
params = load_config(args.config_path)
# NOTE: Retrain the saved model from the last checkpoint
elif args.saved_model_path is not None:
params = load_config(os.path.join(args.saved_model_path, 'config.yml'))
else:
raise ValueError("Set model_save_path or saved_model_path.")
# Load dataset
train_data = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_channel=params['input_channel'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type='train',
data_size=params['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'],
tool=params['tool'],
num_enque=None,
dynamic_batching=params['dynamic_batching'])
dev_clean_data = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_channel=params['input_channel'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type='dev_clean',
data_size=params['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'],
shuffle=True,
tool=params['tool'])
dev_other_data = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_channel=params['input_channel'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type='dev_other',
data_size=params['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'],
shuffle=True,
tool=params['tool'])
test_clean_data = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_channel=params['input_channel'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type='test_clean',
data_size=params['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'],
tool=params['tool'])
test_other_data = Dataset(data_save_path=args.data_save_path,
backend=params['backend'],
input_channel=params['input_channel'],
use_delta=params['use_delta'],
use_double_delta=params['use_double_delta'],
data_type='test_other',
data_size=params['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'],
tool=params['tool'])
params['num_classes'] = train_data.num_classes
##################################################
# MODEL
##################################################
# Model setting
model = load(model_type=params['model_type'],
params=params,
backend=params['backend'])
if args.model_save_path is not None:
# Set save path
save_path = mkdir_join(args.model_save_path, params['backend'],
params['model_type'], params['label_type'],
params['data_size'], model.name)
model.set_save_path(save_path)
# Save config file
save_config(config_path=args.config_path, save_path=model.save_path)
# Setting for logging
logger = set_logger(model.save_path)
if os.path.isdir(params['char_init']):
# NOTE: Start training from the pre-trained character model
model.load_checkpoint(save_path=params['char_init'],
epoch=-1,
load_pretrained_model=True)
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
logger.info("%s %d" % (name, num_params))
logger.info("Total %.3f M parameters" %
(model.total_parameters / 1000000))
# Define optimizer
model.set_optimizer(optimizer=params['optimizer'],
learning_rate_init=float(params['learning_rate']),
weight_decay=float(params['weight_decay']),
clip_grad_norm=params['clip_grad_norm'],
lr_schedule=False,
factor=params['decay_rate'],
patience_epoch=params['decay_patient_epoch'])
epoch, step = 1, 0
learning_rate = float(params['learning_rate'])
metric_dev_best = 1
# NOTE: Retrain the saved model from the last checkpoint
elif args.saved_model_path is not None:
# Set save path
model.save_path = args.saved_model_path
# Setting for logging
logger = set_logger(model.save_path, restart=True)
# Define optimizer
model.set_optimizer(
optimizer=params['optimizer'],
learning_rate_init=float(params['learning_rate']), # on-the-fly
weight_decay=float(params['weight_decay']),
clip_grad_norm=params['clip_grad_norm'],
lr_schedule=False,
factor=params['decay_rate'],
patience_epoch=params['decay_patient_epoch'])
# Restore the last saved model
epoch, step, learning_rate, metric_dev_best = model.load_checkpoint(
save_path=args.saved_model_path, epoch=-1, restart=True)
else:
raise ValueError("Set model_save_path or saved_model_path.")
train_data.epoch = epoch - 1
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
logger.info('PID: %s' % os.getpid())
logger.info('USERNAME: %s' % os.uname()[1])
# Set process name
setproctitle('libri_' + params['backend'] + '_' + params['model_type'] +
'_' + params['label_type'] + '_' + params['data_size'])
##################################################
# TRAINING LOOP
##################################################
# Define learning rate controller
lr_controller = Controller(
learning_rate_init=learning_rate,
backend=params['backend'],
decay_start_epoch=params['decay_start_epoch'],
decay_rate=params['decay_rate'],
decay_patient_epoch=params['decay_patient_epoch'],
lower_better=True)
# Setting for tensorboard
if params['backend'] == 'pytorch':
tf_writer = SummaryWriter(model.save_path)
# Train model
csv_steps, csv_loss_train, csv_loss_dev = [], [], []
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
not_improved_epoch = 0
best_model = model
loss_train_mean = 0.
pbar_epoch = tqdm(total=len(train_data))
while True:
# Compute loss in the training set (including parameter update)
batch_train, is_new_epoch = train_data.next()
model, loss_train_val = train_step(model,
batch_train,
params['clip_grad_norm'],
backend=params['backend'])
loss_train_mean += loss_train_val
pbar_epoch.update(len(batch_train['xs']))
if (step + 1) % params['print_step'] == 0:
# Compute loss in the dev set
batch_dev = dev_clean_data.next()[0]
loss_dev = model(batch_dev['xs'],
batch_dev['ys'],
batch_dev['x_lens'],
batch_dev['y_lens'],
is_eval=True)
loss_train_mean /= params['print_step']
csv_steps.append(step)
csv_loss_train.append(loss_train_mean)
csv_loss_dev.append(loss_dev)
# Logging by tensorboard
if params['backend'] == 'pytorch':
tf_writer.add_scalar('train/loss', loss_train_mean, step + 1)
tf_writer.add_scalar('dev/loss', loss_dev, step + 1)
for name, param in model.named_parameters():
name = name.replace('.', '/')
tf_writer.add_histogram(name,
param.data.cpu().numpy(), step + 1)
tf_writer.add_histogram(name + '/grad',
param.grad.data.cpu().numpy(),
step + 1)
duration_step = time.time() - start_time_step
logger.info(
"...Step:%d(epoch:%.3f) loss:%.3f(%.3f)/lr:%.5f/batch:%d/x_lens:%d (%.3f min)"
% (step + 1, train_data.epoch_detail, loss_train_mean,
loss_dev, learning_rate, train_data.current_batch_size,
max(batch_train['x_lens']) * params['num_stack'],
duration_step / 60))
start_time_step = time.time()
loss_train_mean = 0.
step += 1
# Save checkpoint and evaluate model per epoch
if is_new_epoch:
duration_epoch = time.time() - start_time_epoch
logger.info('===== EPOCH:%d (%.3f min) =====' %
(epoch, duration_epoch / 60))
# Save fugure of loss
plot_loss(csv_loss_train,
csv_loss_dev,
csv_steps,
save_path=model.save_path)
if epoch < params['eval_start_epoch']:
# Save the model
model.save_checkpoint(model.save_path, epoch, step,
learning_rate, metric_dev_best)
else:
start_time_eval = time.time()
# dev
if 'word' in params['label_type']:
metric_dev_epoch, _ = do_eval_wer(
models=[model],
dataset=dev_clean_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_WORD,
eval_batch_size=1)
logger.info(' WER (dev-clean): %.3f %%' %
(metric_dev_epoch * 100))
else:
metric_dev_epoch, wer_dev_clean_epoch, _ = do_eval_cer(
models=[model],
dataset=dev_clean_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_CHAR,
eval_batch_size=1)
logger.info(' CER / WER (dev-clean): %.3f %% / %.3f %%' %
((metric_dev_epoch * 100),
(wer_dev_clean_epoch * 100)))
if metric_dev_epoch < metric_dev_best:
metric_dev_best = metric_dev_epoch
not_improved_epoch = 0
best_model = copy.deepcopy(model)
logger.info('||||| Best Score |||||')
# Save the model
model.save_checkpoint(model.save_path, epoch, step,
learning_rate, metric_dev_best)
# dev-other & test
if 'word' in params['label_type']:
metric_dev_other_epoch, _ = do_eval_wer(
models=[model],
dataset=dev_other_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_WORD,
eval_batch_size=1)
logger.info(' WER (dev-other): %.3f %%' %
(metric_dev_other_epoch * 100))
wer_test_clean, _ = do_eval_wer(
models=[model],
dataset=test_clean_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_WORD,
eval_batch_size=1)
logger.info(' WER (test-clean): %.3f %%' %
(wer_test_clean * 100))
wer_test_other, _ = do_eval_wer(
models=[model],
dataset=test_other_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_WORD,
eval_batch_size=1)
logger.info(' WER (test-other): %.3f %%' %
(wer_test_other * 100))
logger.info(
' WER (test-mean): %.3f %%' %
((wer_test_clean + wer_test_other) * 100 / 2))
else:
metric_dev_other_epoch, wer_dev_other_epoch, _ = do_eval_cer(
models=[model],
dataset=dev_other_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_CHAR,
eval_batch_size=1)
logger.info(
' CER / WER (dev-other): %.3f %% / %.3f %%' %
((metric_dev_other_epoch * 100),
(wer_dev_other_epoch * 100)))
cer_test_clean, wer_test_clean, _ = do_eval_cer(
models=[model],
dataset=test_clean_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_CHAR,
eval_batch_size=1)
logger.info(
' CER / WER (test-clean): %.3f %% / %.3f %%' %
((cer_test_clean * 100), (wer_test_clean * 100)))
cer_test_other, wer_test_other, _ = do_eval_cer(
models=[model],
dataset=test_other_data,
beam_width=1,
max_decode_len=MAX_DECODE_LEN_CHAR,
eval_batch_size=1)
logger.info(
' CER / WER (test-other): %.3f %% / %.3f %%' %
((cer_test_other * 100), (wer_test_other * 100)))
logger.info(
' CER / WER (test-mean): %.3f %% / %.3f %%' %
(((cer_test_clean + cer_test_other) * 100 / 2),
((wer_test_clean + wer_test_other) * 100 / 2)))
else:
not_improved_epoch += 1
duration_eval = time.time() - start_time_eval
logger.info('Evaluation time: %.3f min' % (duration_eval / 60))
# Early stopping
if not_improved_epoch == params['not_improved_patient_epoch']:
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=epoch,
value=metric_dev_epoch)
if epoch == params['convert_to_sgd_epoch']:
# Convert to fine-tuning stage
model.set_optimizer(
'sgd',
learning_rate_init=learning_rate,
weight_decay=float(params['weight_decay']),
clip_grad_norm=params['clip_grad_norm'],
lr_schedule=False,
factor=params['decay_rate'],
patience_epoch=params['decay_patient_epoch'])
logger.info('========== Convert to SGD ==========')
# Inject Gaussian noise to all parameters
if float(params['weight_noise_std']) > 0:
model.weight_noise_injection = True
pbar_epoch = tqdm(total=len(train_data))
print('========== EPOCH:%d (%.3f min) ==========' %
(epoch, duration_epoch / 60))
if epoch == params['num_epoch']:
break
start_time_step = time.time()
start_time_epoch = time.time()
epoch += 1
# TODO: evaluate the best model by beam search here
duration_train = time.time() - start_time_train
logger.info('Total time: %.3f hour' % (duration_train / 3600))
if params['backend'] == 'pytorch':
tf_writer.close()
# Training was finished correctly
with open(os.path.join(model.save_path, 'COMPLETE'), 'w') as f:
f.write('')
def do_train(model, params):
"""Run training.
Args:
model: the model to train
params (dict): A dictionary of parameters
"""
# Load dataset
train_data = Dataset(
data_type='train', 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'],
shuffle=True)
dev_data = Dataset(
data_type='dev', label_type=params['label_type'],
batch_size=params['batch_size'], splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=False)
test_data = Dataset(
data_type='dev', label_type=params['label_type'],
batch_size=params['batch_size'], splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
shuffle=False)
# Tell TensorFlow that the model will be built into the default graph
with tf.Graph().as_default():
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation (including model definition)
loss_op, 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])
train_op = model.train(
loss_op,
optimizer=params['optimizer'],
learning_rate=learning_rate_pl)
decode_op = model.decoder(logits,
model.inputs_seq_len_pl_list[0],
beam_width=params['beam_width'])
ler_op = model.compute_ler(decode_op, model.labels_pl_list[0])
posteriors_op = model.posteriors(logits, blank_prior=1)
# 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=False)
# 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
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)
# Train model
start_time_train = time.time()
start_time_epoch = time.time()
start_time_step = time.time()
fmean_dev_best = 0
fmean_time_dev_best = 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 = {
model.inputs_pl_list[0]: inputs[0],
model.labels_pl_list[0]: list2sparsetensor(
labels[0], padded_value=train_data.padded_value),
model.inputs_seq_len_pl_list[0]: inputs_seq_len[0],
model.keep_prob_pl_list[0]: 1 - float(params['dropout']),
learning_rate_pl: learning_rate
}
# Update parameters
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % params['print_step'] == 0:
# Create feed dictionary for next mini batch (dev)
(inputs, labels, inputs_seq_len, _), _ = dev_data.next()
feed_dict_dev = {
model.inputs_pl_list[0]: inputs[0],
model.labels_pl_list[0]: list2sparsetensor(
labels[0], padded_value=dev_data.padded_value),
model.inputs_seq_len_pl_list[0]: inputs_seq_len[0],
model.keep_prob_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_pl_list[0]] = 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 ===')
fmean_dev_epoch, df_acc = do_eval_fmeasure(
session=sess,
decode_op=decode_op,
model=model,
dataset=dev_data,
eval_batch_size=params['batch_size'])
print(df_acc)
print(' F-measure: %f %%' % (fmean_dev_epoch))
if fmean_dev_epoch > fmean_dev_best:
fmean_dev_best = fmean_dev_epoch
print('■■■ ↑Best Score (F-measure)↑ ■■■')
# Save model only when best accuracy is
# obtained (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 ===')
fmean_test_epoch, df_acc = do_eval_fmeasure(
session=sess,
decode_op=decode_op,
model=model,
dataset=test_data,
eval_batch_size=params['batch_size'])
print(df_acc)
print(' F-measure: %f %%' % (fmean_test_epoch))
# fmean_time_dev_epoch, df_acc = do_eval_fmeasure_time(
# session=sess,
# decode_op=decode_op,
# posteriors_op=posteriors_op,
# model=model,
# dataset=dev_data,
# eval_batch_size=params['batch_size'])
# print(df_acc)
# print(' Time F-measure: %f %%' %
# (fmean_time_dev_epoch))
# if fmean_time_dev_best < fmean_time_dev_epoch:
# fmean_time_dev_best = fmean_time_dev_epoch
# print('■■■ ↑Best Score (Time F-measure)↑ ■■■')
# fmean_time_test_epoch, df_acc = do_eval_fmeasure_time(
# session=sess,
# decode_op=decode_op,
# posteriors_op=posteriors_op,
# model=model,
# dataset=test_data,
# eval_batch_size=params['batch_size'])
# print(df_acc)
# print(' Time F-measure: %f %%' %
# (fmean_time_test_epoch))
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=fmean_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.save_path, 'complete.txt'), 'w') as f:
f.write('')
def check(self,
encoder_type,
decoder_type,
bidirectional=False,
attention_type='location',
label_type='char',
subsample=False,
projection=False,
init_dec_state='first',
ctc_loss_weight=0,
conv=False,
batch_norm=False,
residual=False,
dense_residual=False,
decoding_order='bahdanau_attention',
backward_loss_weight=0,
num_heads=1,
beam_width=1):
print('==================================================')
print(' label_type: %s' % label_type)
print(' encoder_type: %s' % encoder_type)
print(' bidirectional: %s' % str(bidirectional))
print(' projection: %s' % str(projection))
print(' decoder_type: %s' % decoder_type)
print(' init_dec_state: %s' % init_dec_state)
print(' attention_type: %s' % attention_type)
print(' subsample: %s' % str(subsample))
print(' ctc_loss_weight: %s' % str(ctc_loss_weight))
print(' conv: %s' % str(conv))
print(' batch_norm: %s' % str(batch_norm))
print(' residual: %s' % str(residual))
print(' dense_residual: %s' % str(dense_residual))
print(' decoding_order: %s' % decoding_order)
print(' backward_loss_weight: %s' % str(backward_loss_weight))
print(' num_heads: %s' % str(num_heads))
print(' beam_width: %s' % str(beam_width))
print('==================================================')
if conv or encoder_type == 'cnn':
# pattern 1
# conv_channels = [32, 32]
# conv_kernel_sizes = [[41, 11], [21, 11]]
# conv_strides = [[2, 2], [2, 1]]
# poolings = [[], []]
# pattern 2 (VGG like)
conv_channels = [64, 64]
conv_kernel_sizes = [[3, 3], [3, 3]]
conv_strides = [[1, 1], [1, 1]]
poolings = [[2, 2], [2, 2]]
else:
conv_channels = []
conv_kernel_sizes = []
conv_strides = []
poolings = []
# Load batch data
splice = 1
num_stack = 1 if subsample or conv or encoder_type == 'cnn' else 3
xs, ys, x_lens, y_lens = generate_data(label_type=label_type,
batch_size=2,
num_stack=num_stack,
splice=splice)
if label_type == 'char':
num_classes = 27
map_fn = idx2char
elif label_type == 'word':
num_classes = 11
map_fn = idx2word
# Load model
model = AttentionSeq2seq(
input_size=xs.shape[-1] // splice // num_stack, # 120
encoder_type=encoder_type,
encoder_bidirectional=bidirectional,
encoder_num_units=256,
encoder_num_proj=256 if projection else 0,
encoder_num_layers=1 if not subsample else 2,
attention_type=attention_type,
attention_dim=128,
decoder_type=decoder_type,
decoder_num_units=256,
decoder_num_layers=1,
embedding_dim=32,
dropout_input=0.1,
dropout_encoder=0.1,
dropout_decoder=0.1,
dropout_embedding=0.1,
num_classes=num_classes,
parameter_init_distribution='uniform',
parameter_init=0.1,
recurrent_weight_orthogonal=False,
init_forget_gate_bias_with_one=True,
subsample_list=[] if not subsample else [True, False],
subsample_type='concat' if not subsample else subsample,
bridge_layer=True,
init_dec_state=init_dec_state,
sharpening_factor=1,
logits_temperature=1,
sigmoid_smoothing=False,
coverage_weight=0,
ctc_loss_weight=ctc_loss_weight,
attention_conv_num_channels=10,
attention_conv_width=201,
num_stack=num_stack,
splice=splice,
input_channel=3,
conv_channels=conv_channels,
conv_kernel_sizes=conv_kernel_sizes,
conv_strides=conv_strides,
poolings=poolings,
activation='relu',
batch_norm=batch_norm,
scheduled_sampling_prob=0.1,
scheduled_sampling_max_step=200,
label_smoothing_prob=0.1,
weight_noise_std=1e-9,
encoder_residual=residual,
encoder_dense_residual=dense_residual,
decoder_residual=residual,
decoder_dense_residual=dense_residual,
decoding_order=decoding_order,
bottleneck_dim=256,
backward_loss_weight=backward_loss_weight,
num_heads=num_heads)
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
print("%s %d" % (name, num_params))
print("Total %.3f M parameters" % (model.total_parameters / 1000000))
# Define optimizer
learning_rate = 1e-3
model.set_optimizer('adam',
learning_rate_init=learning_rate,
weight_decay=1e-8,
lr_schedule=False,
factor=0.1,
patience_epoch=5)
# Define learning rate controller
lr_controller = Controller(learning_rate_init=learning_rate,
backend='pytorch',
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
# Train model
max_step = 300
start_time_step = time.time()
for step in range(max_step):
# Step for parameter update
model.optimizer.zero_grad()
loss = model(xs, ys, x_lens, y_lens)
loss.backward()
# torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
torch.nn.utils.clip_grad_norm(model.parameters(), 5)
model.optimizer.step()
# Inject Gaussian noise to all parameters
# if loss.item() < 50:
if loss.data[0] < 50:
model.weight_noise_injection = True
if (step + 1) % 10 == 0:
# Compute loss
loss = model(xs, ys, x_lens, y_lens, is_eval=True)
# Decode
best_hyps, _, perm_idx = model.decode(xs,
x_lens,
beam_width,
max_decode_len=60)
str_ref = map_fn(ys[0])
str_hyp = map_fn(best_hyps[0][:-1])
# Compute accuracy
try:
if label_type == 'char':
ler, _, _, _ = compute_wer(
ref=list(str_ref.replace('_', '')),
hyp=list(str_hyp.replace('_', '')),
normalize=True)
elif label_type == 'word':
ler, _, _, _ = compute_wer(ref=str_ref.split('_'),
hyp=str_hyp.split('_'),
normalize=True)
except:
ler = 1
duration_step = time.time() - start_time_step
print('Step %d: loss=%.3f / ler=%.3f / lr=%.5f (%.3f sec)' %
(step + 1, loss, ler, learning_rate, duration_step))
start_time_step = time.time()
# Visualize
print('Ref: %s' % str_ref)
print('Hyp: %s' % str_hyp)
# Decode by the CTC decoder
if model.ctc_loss_weight >= 0.1:
best_hyps_ctc, perm_idx = model.decode_ctc(xs,
x_lens,
beam_width=1)
str_pred_ctc = map_fn(best_hyps_ctc[0])
print('Hyp (CTC): %s' % str_pred_ctc)
if ler < 0.1:
print('Modle is Converged.')
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=step,
value=ler)
def check_training(self, label_type):
print('----- ' + label_type + ' -----')
tf.reset_default_graph()
with tf.Graph().as_default():
# Load batch data
batch_size = 1
inputs, att_labels, inputs_seq_len, att_labels_seq_len, ctc_labels_st = generate_data(
label_type=label_type,
model='joint_ctc_attention',
batch_size=batch_size)
# Define model graph
att_num_classes = 26 + 2 if label_type == 'character' else 61 + 2
ctc_num_classes = 26 if label_type == 'character' else 61
# model = load(model_type=model_type)
network = JointCTCAttention(input_size=inputs[0].shape[1],
encoder_num_unit=256,
encoder_num_layer=2,
attention_dim=128,
attention_type='content',
decoder_num_unit=256,
decoder_num_layer=1,
embedding_dim=20,
att_num_classes=att_num_classes,
ctc_num_classes=ctc_num_classes,
att_task_weight=0.5,
sos_index=att_num_classes - 2,
eos_index=att_num_classes - 1,
max_decode_length=50,
attention_weights_tempareture=1.0,
logits_tempareture=1.0,
parameter_init=0.1,
clip_grad=5.0,
clip_activation_encoder=50,
clip_activation_decoder=50,
dropout_ratio_input=0.9,
dropout_ratio_hidden=0.9,
dropout_ratio_output=1.0,
weight_decay=1e-8,
beam_width=1,
time_major=False)
# Define placeholders
network.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation
loss_op, att_logits, ctc_logits, decoder_outputs_train, decoder_outputs_infer = network.compute_loss(
network.inputs_pl_list[0], network.att_labels_pl_list[0],
network.inputs_seq_len_pl_list[0],
network.att_labels_seq_len_pl_list[0],
network.ctc_labels_pl_list[0],
network.keep_prob_input_pl_list[0],
network.keep_prob_hidden_pl_list[0],
network.keep_prob_output_pl_list[0])
train_op = network.train(loss_op,
optimizer='adam',
learning_rate=learning_rate_pl)
decode_op_train, decode_op_infer = network.decoder(
decoder_outputs_train, decoder_outputs_infer)
ler_op = network.compute_ler(network.att_labels_st_true_pl,
network.att_labels_st_pred_pl)
# Define learning rate controller
learning_rate = 1e-3
lr_controller = Controller(learning_rate_init=learning_rate,
decay_start_epoch=10,
decay_rate=0.99,
decay_patient_epoch=5,
lower_better=True)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# 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)))
# Make feed dict
feed_dict = {
network.inputs_pl_list[0]: inputs,
network.att_labels_pl_list[0]: att_labels,
network.inputs_seq_len_pl_list[0]: inputs_seq_len,
network.att_labels_seq_len_pl_list[0]: att_labels_seq_len,
network.ctc_labels_pl_list[0]: ctc_labels_st,
network.keep_prob_input_pl_list[0]:
network.dropout_ratio_input,
network.keep_prob_hidden_pl_list[0]:
network.dropout_ratio_hidden,
network.keep_prob_output_pl_list[0]:
network.dropout_ratio_output,
learning_rate_pl: learning_rate
}
map_file_path = '../../experiments/timit/metrics/mapping_files/attention/phone61_to_num.txt'
with tf.Session() as sess:
# Initialize parameters
sess.run(init_op)
# Wrapper for tfdbg
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Train model
max_steps = 400
start_time_global = time.time()
start_time_step = time.time()
ler_train_pre = 1
not_improved_count = 0
for step in range(max_steps):
# Compute loss
_, loss_train = sess.run([train_op, loss_op],
feed_dict=feed_dict)
# Gradient check
# grads = sess.run(network.clipped_grads,
# feed_dict=feed_dict)
# for grad in grads:
# print(np.max(grad))
if (step + 1) % 10 == 0:
# Change to evaluation mode
feed_dict[network.keep_prob_input_pl_list[0]] = 1.0
feed_dict[network.keep_prob_hidden_pl_list[0]] = 1.0
feed_dict[network.keep_prob_output_pl_list[0]] = 1.0
# Predict class ids
predicted_ids_train, predicted_ids_infer = sess.run(
[decode_op_train, decode_op_infer],
feed_dict=feed_dict)
# Compute accuracy
try:
feed_dict_ler = {
network.att_labels_st_true_pl:
list2sparsetensor(att_labels, padded_value=27),
network.att_labels_st_pred_pl:
list2sparsetensor(predicted_ids_infer,
padded_value=27)
}
ler_train = sess.run(ler_op,
feed_dict=feed_dict_ler)
except ValueError:
ler_train = 1
duration_step = time.time() - start_time_step
print(
'Step %d: loss = %.3f / ler = %.4f (%.3f sec) / lr = %.5f'
% (step + 1, loss_train, ler_train, duration_step,
learning_rate))
start_time_step = time.time()
# Visualize
if label_type == 'character':
print('True : %s' %
idx2alpha(att_labels[0]))
print('Pred (Training) : <%s' %
idx2alpha(predicted_ids_train[0]))
print('Pred (Inference): <%s' %
idx2alpha(predicted_ids_infer[0]))
else:
print('True : %s' %
idx2phone(att_labels[0], map_file_path))
print('Pred (Training) : < %s' % idx2phone(
predicted_ids_train[0], map_file_path))
print('Pred (Inference): < %s' % idx2phone(
predicted_ids_infer[0], map_file_path))
if ler_train >= ler_train_pre:
not_improved_count += 1
else:
not_improved_count = 0
if not_improved_count >= 10:
print('Model is Converged.')
break
ler_train_pre = ler_train
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=step,
value=ler_train)
feed_dict[learning_rate_pl] = learning_rate
duration_global = time.time() - start_time_global
print('Total time: %.3f sec' % (duration_global))
def check(self, usage_dec_sub='all', att_reg_weight=1,
main_loss_weight=0.5, ctc_loss_weight_sub=0,
dec_attend_temperature=1,
dec_sigmoid_smoothing=False,
backward_sub=False, num_heads=1, second_pass=False,
relax_context_vec_dec=False):
print('==================================================')
print(' usage_dec_sub: %s' % usage_dec_sub)
print(' att_reg_weight: %s' % str(att_reg_weight))
print(' main_loss_weight: %s' % str(main_loss_weight))
print(' ctc_loss_weight_sub: %s' % str(ctc_loss_weight_sub))
print(' dec_attend_temperature: %s' % str(dec_attend_temperature))
print(' dec_sigmoid_smoothing: %s' % str(dec_sigmoid_smoothing))
print(' backward_sub: %s' % str(backward_sub))
print(' num_heads: %s' % str(num_heads))
print(' second_pass: %s' % str(second_pass))
print(' relax_context_vec_dec: %s' % str(relax_context_vec_dec))
print('==================================================')
# Load batch data
splice = 1
num_stack = 1
xs, ys, ys_sub, x_lens, y_lens, y_lens_sub = generate_data(
label_type='word_char',
batch_size=2,
num_stack=num_stack,
splice=splice)
# Load model
model = NestedAttentionSeq2seq(
input_size=xs.shape[-1] // splice // num_stack, # 120
encoder_type='lstm',
encoder_bidirectional=True,
encoder_num_units=256,
encoder_num_proj=0,
encoder_num_layers=2,
encoder_num_layers_sub=2,
attention_type='location',
attention_dim=128,
decoder_type='lstm',
decoder_num_units=256,
decoder_num_layers=1,
decoder_num_units_sub=256,
decoder_num_layers_sub=1,
embedding_dim=64,
embedding_dim_sub=32,
dropout_input=0.1,
dropout_encoder=0.1,
dropout_decoder=0.1,
dropout_embedding=0.1,
main_loss_weight=0.8,
sub_loss_weight=0.2 if ctc_loss_weight_sub == 0 else 0,
num_classes=11,
num_classes_sub=27 if not second_pass else 11,
parameter_init_distribution='uniform',
parameter_init=0.1,
recurrent_weight_orthogonal=False,
init_forget_gate_bias_with_one=True,
subsample_list=[True, False],
subsample_type='drop',
init_dec_state='first',
sharpening_factor=1,
logits_temperature=1,
sigmoid_smoothing=False,
ctc_loss_weight_sub=ctc_loss_weight_sub,
attention_conv_num_channels=10,
attention_conv_width=201,
num_stack=num_stack,
splice=1,
conv_channels=[],
conv_kernel_sizes=[],
conv_strides=[],
poolings=[],
batch_norm=False,
scheduled_sampling_prob=0.1,
scheduled_sampling_max_step=200,
label_smoothing_prob=0.1,
weight_noise_std=0,
encoder_residual=False,
encoder_dense_residual=False,
decoder_residual=False,
decoder_dense_residual=False,
decoding_order='attend_generate_update',
# decoding_order='attend_update_generate',
# decoding_order='conditional',
bottleneck_dim=256,
bottleneck_dim_sub=256,
backward_sub=backward_sub,
num_heads=num_heads,
num_heads_sub=num_heads,
num_heads_dec=num_heads,
usage_dec_sub=usage_dec_sub,
att_reg_weight=att_reg_weight,
dec_attend_temperature=dec_attend_temperature,
dec_sigmoid_smoothing=dec_attend_temperature,
relax_context_vec_dec=relax_context_vec_dec,
dec_attention_type='location')
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
print("%s %d" % (name, num_params))
print("Total %.3f M parameters" % (model.total_parameters / 1000000))
# Define optimizer
learning_rate = 1e-3
model.set_optimizer('adam',
learning_rate_init=learning_rate,
weight_decay=1e-6,
lr_schedule=False,
factor=0.1,
patience_epoch=5)
# Define learning rate controller
lr_controller = Controller(learning_rate_init=learning_rate,
backend='pytorch',
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
# Train model
max_step = 300
start_time_step = time.time()
for step in range(max_step):
# Step for parameter update
model.optimizer.zero_grad()
if second_pass:
loss = model(xs, ys, x_lens, y_lens)
else:
loss, loss_main, loss_sub = model(
xs, ys, x_lens, y_lens, ys_sub, y_lens_sub)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
model.optimizer.step()
if (step + 1) % 10 == 0:
# Compute loss
if second_pass:
loss = model(xs, ys, x_lens, y_lens, is_eval=True)
else:
loss, loss_main, loss_sub = model(
xs, ys, x_lens, y_lens, ys_sub, y_lens_sub, is_eval=True)
best_hyps, _, best_hyps_sub, _, perm_idx = model.decode(
xs, x_lens, beam_width=1,
max_decode_len=30,
max_decode_len_sub=60)
str_hyp = idx2word(best_hyps[0][:-1])
str_ref = idx2word(ys[0])
if second_pass:
str_hyp_sub = idx2word(best_hyps_sub[0][:-1])
str_ref_sub = idx2word(ys[0])
else:
str_hyp_sub = idx2char(best_hyps_sub[0][:-1])
str_ref_sub = idx2char(ys_sub[0])
# Compute accuracy
try:
wer, _, _, _ = compute_wer(ref=str_ref.split('_'),
hyp=str_hyp.split('_'),
normalize=True)
if second_pass:
cer, _, _, _ = compute_wer(ref=str_ref.split('_'),
hyp=str_hyp_sub.split('_'),
normalize=True)
else:
cer, _, _, _ = compute_wer(
ref=list(str_ref_sub.replace('_', '')),
hyp=list(str_hyp_sub.replace('_', '')),
normalize=True)
except:
wer = 1
cer = 1
duration_step = time.time() - start_time_step
if second_pass:
print('Step %d: loss=%.3f / wer=%.3f / cer=%.3f / lr=%.5f (%.3f sec)' %
(step + 1, loss, wer, cer, learning_rate, duration_step))
else:
print('Step %d: loss=%.3f(%.3f/%.3f) / wer=%.3f / cer=%.3f / lr=%.5f (%.3f sec)' %
(step + 1, loss, loss_main, loss_sub,
wer, cer, learning_rate, duration_step))
start_time_step = time.time()
# Visualize
print('Ref: %s' % str_ref)
print('Hyp (word): %s' % str_hyp)
print('Hyp (char): %s' % str_hyp_sub)
if cer < 0.1:
print('Modle is Converged.')
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=step,
value=wer)
def check(self, encoder_type, lstm_impl, time_major=False):
print('==================================================')
print(' encoder_type: %s' % str(encoder_type))
print(' lstm_impl: %s' % str(lstm_impl))
print(' time_major: %s' % str(time_major))
print('==================================================')
tf.reset_default_graph()
with tf.Graph().as_default():
# Load batch data
batch_size = 2
inputs, labels_char, labels_phone, inputs_seq_len = generate_data(
label_type='multitask', model='ctc', batch_size=batch_size)
# Define model graph
num_classes_main = 27
num_classes_sub = 61
model = MultitaskCTC(
encoder_type=encoder_type,
input_size=inputs[0].shape[1],
num_units=256,
num_layers_main=2,
num_layers_sub=1,
num_classes_main=num_classes_main,
num_classes_sub=num_classes_sub,
main_task_weight=0.8,
lstm_impl=lstm_impl,
parameter_init=0.1,
clip_grad_norm=5.0,
clip_activation=50,
num_proj=256,
weight_decay=1e-8,
# bottleneck_dim=50,
bottleneck_dim=None,
time_major=time_major)
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# 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_pl_list[0])
train_op = model.train(loss_op,
optimizer='adam',
learning_rate=learning_rate_pl)
decode_op_main, decode_op_sub = model.decoder(
logits_main,
logits_sub,
model.inputs_seq_len_pl_list[0],
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
learning_rate = 1e-3
lr_controller = Controller(learning_rate_init=learning_rate,
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=5,
lower_better=True)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# 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)))
# Make feed dict
feed_dict = {
model.inputs_pl_list[0]:
inputs,
model.labels_pl_list[0]:
list2sparsetensor(labels_char, padded_value=-1),
model.labels_sub_pl_list[0]:
list2sparsetensor(labels_phone, padded_value=-1),
model.inputs_seq_len_pl_list[0]:
inputs_seq_len,
model.keep_prob_pl_list[0]:
0.9,
learning_rate_pl:
learning_rate
}
idx2phone = Idx2phone(map_file_path='./phone61.txt')
with tf.Session() as sess:
# Initialize parameters
sess.run(init_op)
# Wrapper for tfdbg
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Train model
max_steps = 1000
start_time_step = time.time()
for step in range(max_steps):
# Compute loss
_, loss_train = sess.run([train_op, loss_op],
feed_dict=feed_dict)
# Gradient check
# grads = sess.run(model.clipped_grads,
# feed_dict=feed_dict)
# for grad in grads:
# print(np.max(grad))
if (step + 1) % 10 == 0:
# Change to evaluation mode
feed_dict[model.keep_prob_pl_list[0]] = 1.0
# Compute accuracy
ler_train_char, ler_train_phone = sess.run(
[ler_op_main, ler_op_sub], feed_dict=feed_dict)
duration_step = time.time() - start_time_step
print(
'Step %d: loss = %.3f / cer = %.3f / per = %.3f (%.3f sec) / lr = %.5f'
% (step + 1, loss_train, ler_train_char,
ler_train_phone, duration_step, learning_rate))
start_time_step = time.time()
# Visualize
labels_pred_char_st, labels_pred_phone_st = sess.run(
[decode_op_main, decode_op_sub],
feed_dict=feed_dict)
labels_pred_char = sparsetensor2list(
labels_pred_char_st, batch_size=batch_size)
labels_pred_phone = sparsetensor2list(
labels_pred_phone_st, batch_size=batch_size)
print('Character')
try:
print(' Ref: %s' % idx2alpha(labels_char[0]))
print(' Hyp: %s' % idx2alpha(labels_pred_char[0]))
except IndexError:
print('Character')
print(' Ref: %s' % idx2alpha(labels_char[0]))
print(' Hyp: %s' % '')
print('Phone')
try:
print(' Ref: %s' % idx2phone(labels_phone[0]))
print(' Hyp: %s' %
idx2phone(labels_pred_phone[0]))
except IndexError:
print(' Ref: %s' % idx2phone(labels_phone[0]))
print(' Hyp: %s' % '')
# NOTE: This is for no prediction
print('-' * 30)
if ler_train_char < 0.1:
print('Modle is Converged.')
break
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=step,
value=ler_train_char)
feed_dict[learning_rate_pl] = learning_rate
def do_train(model, params):
"""Run training. If target labels are phone, the model is evaluated by PER
with 39 phones.
Args:
model: the model to train
params (dict): A dictionary of parameters
"""
map_file_path_train = '../metrics/mapping_files/' + \
params['label_type'] + '.txt'
if 'phone' in params['label_type']:
map_file_path_eval = '../metrics/mapping_files/phone39.txt'
else:
map_file_path_eval = '../metrics/mapping_files/' + \
params['label_type'] + '.txt'
# Load dataset
train_data = Dataset(
data_type='train', label_type=params['label_type'],
batch_size=params['batch_size'], map_file_path=map_file_path_train,
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'])
dev_data = Dataset(
data_type='dev', label_type=params['label_type'],
batch_size=params['batch_size'], map_file_path=map_file_path_train,
splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
sort_utt=False)
if 'char' in params['label_type']:
test_data = Dataset(
data_type='test', label_type=params['label_type'],
batch_size=1, map_file_path=map_file_path_eval,
splice=params['splice'],
num_stack=params['num_stack'], num_skip=params['num_skip'],
sort_utt=False)
else:
test_data = Dataset(
data_type='test', label_type='phone39',
batch_size=1, map_file_path=map_file_path_eval,
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():
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation (including model definition)
loss_op, logits, decoder_outputs_train, decoder_outputs_infer = model.compute_loss(
model.inputs_pl_list[0],
model.labels_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.labels_seq_len_pl_list[0],
model.keep_prob_encoder_pl_list[0],
model.keep_prob_decoder_pl_list[0],
model.keep_prob_embedding_pl_list[0])
train_op = model.train(loss_op,
optimizer=params['optimizer'],
learning_rate=learning_rate_pl)
_, decode_op_infer = model.decode(
decoder_outputs_train,
decoder_outputs_infer)
ler_op = model.compute_ler(model.labels_st_true_pl,
model.labels_st_pred_pl)
# 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 param
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 param" %
(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
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 param
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_train, inputs_seq_len, labels_seq_len, _ = data
feed_dict_train = {
model.inputs_pl_list[0]: inputs[0],
model.labels_pl_list[0]: labels_train[0],
model.inputs_seq_len_pl_list[0]: inputs_seq_len[0],
model.labels_seq_len_pl_list[0]: labels_seq_len[0],
model.keep_prob_encoder_pl_list[0]: 1 - float(params['dropout_encoder']),
model.keep_prob_decoder_pl_list[0]: 1 - float(params['dropout_decoder']),
model.keep_prob_embedding_pl_list[0]: 1 - float(params['dropout_embedding']),
learning_rate_pl: learning_rate
}
# Update parameters
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % params['print_step'] == 0:
# Create feed dictionary for next mini batch (dev)
(inputs, labels_dev, inputs_seq_len,
labels_seq_len, _), _ = dev_data.next()
feed_dict_dev = {
model.inputs_pl_list[0]: inputs[0],
model.labels_pl_list[0]: labels_dev[0],
model.inputs_seq_len_pl_list[0]: inputs_seq_len[0],
model.labels_seq_len_pl_list[0]: labels_seq_len[0],
model.keep_prob_encoder_pl_list[0]: 1.0,
model.keep_prob_decoder_pl_list[0]: 1.0,
model.keep_prob_embedding_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_encoder_pl_list[0]] = 1.0
feed_dict_train[model.keep_prob_decoder_pl_list[0]] = 1.0
feed_dict_train[model.keep_prob_embedding_pl_list[0]] = 1.0
# Predict class ids & update even files
predicted_ids_train, summary_str_train = sess.run(
[decode_op_infer, summary_train], feed_dict=feed_dict_train)
predicted_ids_dev, summary_str_dev = sess.run(
[decode_op_infer, summary_dev], feed_dict=feed_dict_dev)
summary_writer.add_summary(summary_str_train, step + 1)
summary_writer.add_summary(summary_str_dev, step + 1)
summary_writer.flush()
# Convert to sparsetensor to compute LER
feed_dict_ler_train = {
model.labels_st_true_pl: list2sparsetensor(
labels_train[0], padded_value=train_data.padded_value),
model.labels_st_pred_pl: list2sparsetensor(
predicted_ids_train, padded_value=train_data.padded_value)
}
feed_dict_ler_dev = {
model.labels_st_true_pl: list2sparsetensor(
labels_dev[0], padded_value=dev_data.padded_value),
model.labels_st_pred_pl: list2sparsetensor(
predicted_ids_dev, padded_value=dev_data.padded_value)
}
# Compute accuracy
ler_train = sess.run(ler_op, feed_dict=feed_dict_ler_train)
ler_dev = sess.run(ler_op, feed_dict=feed_dict_ler_dev)
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
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()
if 'char' in params['label_type']:
print('=== Dev Data Evaluation ===')
ler_dev_epoch, wer_dev_epoch = do_eval_cer(
session=sess,
decode_op=decode_op_infer,
model=model,
dataset=dev_data,
label_type=params['label_type'],
eval_batch_size=1)
print(' CER: %f %%' % (ler_dev_epoch * 100))
print(' WER: %f %%' % (wer_dev_epoch * 100))
if ler_dev_epoch < ler_dev_best:
ler_dev_best = ler_dev_epoch
print('■■■ ↑Best Score (CER)↑ ■■■')
# Save model only when best accuracy is
# obtained (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 ===')
ler_test, wer_test = do_eval_cer(
session=sess,
decode_op=decode_op_infer,
model=model,
dataset=test_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=1)
print(' CER: %f %%' % (ler_test * 100))
print(' WER: %f %%' % (wer_test * 100))
else:
print('=== Dev Data Evaluation ===')
ler_dev_epoch = do_eval_per(
session=sess,
decode_op=decode_op_infer,
per_op=ler_op,
model=model,
dataset=dev_data,
label_type=params['label_type'],
eval_batch_size=1)
print(' PER: %f %%' % (ler_dev_epoch * 100))
if ler_dev_epoch < ler_dev_best:
ler_dev_best = ler_dev_epoch
print('■■■ ↑Best Score (PER)↑ ■■■')
# Save model only when best accuracy is
# obtained (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 ===')
ler_test = do_eval_per(
session=sess,
decode_op=decode_op_infer,
per_op=ler_op,
model=model,
dataset=test_data,
label_type=params['label_type'],
is_test=True,
eval_batch_size=1)
print(' PER: %f %%' % (ler_test * 100))
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_epoch)
start_time_step = time.time()
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 check(self, encoder_type, lstm_impl, time_major=False):
print('==================================================')
print(' encoder_type: %s' % str(encoder_type))
print(' lstm_impl: %s' % str(lstm_impl))
print(' time_major: %s' % str(time_major))
print('==================================================')
tf.reset_default_graph()
with tf.Graph().as_default():
# Load batch data
batch_size = 2
inputs, labels_char, labels_phone, inputs_seq_len = generate_data(
label_type='multitask',
model='ctc',
batch_size=batch_size)
# Define model graph
num_classes_main = 27
num_classes_sub = 61
model = MultitaskCTC(
encoder_type=encoder_type,
input_size=inputs[0].shape[1],
num_units=256,
num_layers_main=2,
num_layers_sub=1,
num_classes_main=num_classes_main,
num_classes_sub=num_classes_sub,
main_task_weight=0.8,
lstm_impl=lstm_impl,
parameter_init=0.1,
clip_grad_norm=5.0,
clip_activation=50,
num_proj=256,
weight_decay=1e-8,
# bottleneck_dim=50,
bottleneck_dim=None,
time_major=time_major)
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# 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_pl_list[0])
train_op = model.train(
loss_op,
optimizer='adam',
learning_rate=learning_rate_pl)
decode_op_main, decode_op_sub = model.decoder(
logits_main,
logits_sub,
model.inputs_seq_len_pl_list[0],
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
learning_rate = 1e-3
lr_controller = Controller(learning_rate_init=learning_rate,
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=5,
lower_better=True)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# 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)))
# Make feed dict
feed_dict = {
model.inputs_pl_list[0]: inputs,
model.labels_pl_list[0]: list2sparsetensor(labels_char, padded_value=-1),
model.labels_sub_pl_list[0]: list2sparsetensor(labels_phone, padded_value=-1),
model.inputs_seq_len_pl_list[0]: inputs_seq_len,
model.keep_prob_pl_list[0]: 0.9,
learning_rate_pl: learning_rate
}
idx2phone = Idx2phone(map_file_path='./phone61.txt')
with tf.Session() as sess:
# Initialize parameters
sess.run(init_op)
# Wrapper for tfdbg
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Train model
max_steps = 1000
start_time_step = time.time()
for step in range(max_steps):
# Compute loss
_, loss_train = sess.run(
[train_op, loss_op], feed_dict=feed_dict)
# Gradient check
# grads = sess.run(model.clipped_grads,
# feed_dict=feed_dict)
# for grad in grads:
# print(np.max(grad))
if (step + 1) % 10 == 0:
# Change to evaluation mode
feed_dict[model.keep_prob_pl_list[0]] = 1.0
# Compute accuracy
ler_train_char, ler_train_phone = sess.run(
[ler_op_main, ler_op_sub], feed_dict=feed_dict)
duration_step = time.time() - start_time_step
print('Step %d: loss = %.3f / cer = %.3f / per = %.3f (%.3f sec) / lr = %.5f' %
(step + 1, loss_train, ler_train_char,
ler_train_phone, duration_step, learning_rate))
start_time_step = time.time()
# Visualize
labels_pred_char_st, labels_pred_phone_st = sess.run(
[decode_op_main, decode_op_sub],
feed_dict=feed_dict)
labels_pred_char = sparsetensor2list(
labels_pred_char_st, batch_size=batch_size)
labels_pred_phone = sparsetensor2list(
labels_pred_phone_st, batch_size=batch_size)
print('Character')
try:
print(' Ref: %s' % idx2alpha(labels_char[0]))
print(' Hyp: %s' % idx2alpha(labels_pred_char[0]))
except IndexError:
print('Character')
print(' Ref: %s' % idx2alpha(labels_char[0]))
print(' Hyp: %s' % '')
print('Phone')
try:
print(' Ref: %s' % idx2phone(labels_phone[0]))
print(' Hyp: %s' %
idx2phone(labels_pred_phone[0]))
except IndexError:
print(' Ref: %s' % idx2phone(labels_phone[0]))
print(' Hyp: %s' % '')
# NOTE: This is for no prediction
print('-' * 30)
if ler_train_char < 0.1:
print('Modle is Converged.')
break
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=step,
value=ler_train_char)
feed_dict[learning_rate_pl] = learning_rate
def check(self,
rnn_type,
bidirectional=False,
label_type='char',
tie_weights=False):
print('==================================================')
print(' label_type: %s' % label_type)
print(' rnn_type: %s' % rnn_type)
print(' bidirectional: %s' % str(bidirectional))
print(' tie_weights: %s' % str(tie_weights))
print('==================================================')
# Load batch data
_, ys, _, y_lens = generate_data(model_type='lm',
label_type=label_type,
batch_size=2)
if label_type == 'char':
num_classes = 27
map_fn = idx2char
elif label_type == 'word':
num_classes = 11
map_fn = idx2word
# Load model
model = RNNLM(num_classes,
embedding_dim=128,
rnn_type=rnn_type,
bidirectional=bidirectional,
num_units=1024,
num_layers=1,
dropout_embedding=0.1,
dropout_hidden=0.1,
dropout_output=0.1,
parameter_init_distribution='uniform',
parameter_init=0.1,
tie_weights=False)
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
print("%s %d" % (name, num_params))
print("Total %.3f M parameters" % (model.total_parameters / 1000000))
# Define optimizer
learning_rate = 1e-3
model.set_optimizer('adam',
learning_rate_init=learning_rate,
weight_decay=1e-8,
lr_schedule=False,
factor=0.1,
patience_epoch=5)
# Define learning rate controller
lr_controller = Controller(learning_rate_init=learning_rate,
backend='pytorch',
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
# Train model
max_step = 1000
start_time_step = time.time()
for step in range(max_step):
# Step for parameter update
model.optimizer.zero_grad()
loss = model(ys, y_lens)
loss.backward()
nn.utils.clip_grad_norm(model.parameters(), 5)
model.optimizer.step()
# Inject Gaussian noise to all parameters
if loss.data[0] < 50:
model.weight_noise_injection = True
if (step + 1) % 10 == 0:
# Compute loss
loss = model(ys, y_lens, is_eval=True)
# Compute PPL
ppl = math.exp(loss)
# Decode
# best_hyps, perm_idx = model.decode(
# xs, x_lens,
# # beam_width=1,
# beam_width=2,
# max_decode_len=60)
# Compute accuracy
# if label_type == 'char':
# str_true = map_fn(ys[0, :y_lens[0]][1:-1])
# str_pred = map_fn(best_hyps[0][0:-1]).split('>')[0]
# ler = compute_cer(ref=str_true.replace('_', ''),
# hyp=str_pred.replace('_', ''),
# normalize=True)
# elif label_type == 'word':
# str_true = map_fn(ys[0, : y_lens[0]][1: -1])
# str_pred = map_fn(best_hyps[0][0: -1]).split('>')[0]
# ler, _, _, _ = compute_wer(ref=str_true.split('_'),
# hyp=str_pred.split('_'),
# normalize=True)
duration_step = time.time() - start_time_step
print('Step %d: loss=%.3f / ppl=%.3f / lr=%.5f (%.3f sec)' %
(step + 1, loss, ppl, learning_rate, duration_step))
start_time_step = time.time()
# Visualize
# print('Ref: %s' % str_true)
# print('Hyp: %s' % str_pred)
if ppl == 0:
print('Modle is Converged.')
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=step,
value=ppl)
def check_training(self,
encoder_type,
label_type,
lstm_impl='LSTMBlockCell',
save_params=False):
print('==================================================')
print(' encoder_type: %s' % encoder_type)
print(' label_type: %s' % label_type)
print(' lstm_impl: %s' % lstm_impl)
print('==================================================')
tf.reset_default_graph()
with tf.Graph().as_default():
# Load batch data
batch_size = 1
splice = 11 if encoder_type in [
'vgg_blstm', 'vgg_lstm', 'vgg_wang', 'resnet_wang', 'cnn_zhang'
] else 1
inputs, labels_true_st, inputs_seq_len = generate_data(
label_type=label_type,
model='ctc',
batch_size=batch_size,
splice=splice)
# NOTE: input_size must be even number when using CudnnLSTM
# Define model graph
num_classes = 26 if label_type == 'character' else 61
model = CTC(
encoder_type=encoder_type,
input_size=inputs[0].shape[-1] // splice,
splice=splice,
num_units=256,
num_layers=2,
num_classes=num_classes,
lstm_impl=lstm_impl,
parameter_init=0.1,
clip_grad=5.0,
clip_activation=50,
num_proj=256,
# bottleneck_dim=50,
bottleneck_dim=None,
weight_decay=1e-8)
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation
loss_op, 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])
train_op = model.train(loss_op,
optimizer='adam',
learning_rate=learning_rate_pl)
# NOTE: Adam does not run on CudnnLSTM
decode_op = model.decoder(logits,
model.inputs_seq_len_pl_list[0],
beam_width=20)
ler_op = model.compute_ler(decode_op, model.labels_pl_list[0])
# Define learning rate controller
learning_rate = 1e-3
lr_controller = Controller(learning_rate_init=learning_rate,
decay_start_epoch=10,
decay_rate=0.98,
decay_patient_epoch=5,
lower_better=True)
if save_params:
# Create a saver for writing training checkpoints
saver = tf.train.Saver(max_to_keep=None)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# Count total parameters
if lstm_impl != 'CudnnLSTM':
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)))
# Make feed dict
feed_dict = {
model.inputs_pl_list[0]: inputs,
model.labels_pl_list[0]: labels_true_st,
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,
learning_rate_pl: learning_rate
}
idx2phone = Idx2phone(map_file_path='./phone61_ctc.txt')
with tf.Session() as sess:
# Initialize parameters
sess.run(init_op)
# Wrapper for tfdbg
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Train model
max_steps = 1000
start_time_global = time.time()
start_time_step = time.time()
ler_train_pre = 1
not_improved_count = 0
for step in range(max_steps):
# Compute loss
_, loss_train = sess.run([train_op, loss_op],
feed_dict=feed_dict)
# Gradient check
# grads = sess.run(model.clipped_grads,
# feed_dict=feed_dict)
# for grad in grads:
# print(np.max(grad))
if (step + 1) % 10 == 0:
# Change to evaluation mode
feed_dict[model.keep_prob_input_pl_list[0]] = 1.0
feed_dict[model.keep_prob_hidden_pl_list[0]] = 1.0
feed_dict[model.keep_prob_output_pl_list[0]] = 1.0
# Compute accuracy
ler_train = sess.run(ler_op, feed_dict=feed_dict)
duration_step = time.time() - start_time_step
print(
'Step %d: loss = %.3f / ler = %.3f (%.3f sec) / lr = %.5f'
% (step + 1, loss_train, ler_train, duration_step,
learning_rate))
start_time_step = time.time()
# Decode
labels_pred_st = sess.run(decode_op,
feed_dict=feed_dict)
labels_true = sparsetensor2list(labels_true_st,
batch_size=batch_size)
# Visualize
try:
labels_pred = sparsetensor2list(
labels_pred_st, batch_size=batch_size)
if label_type == 'character':
print('Ref: %s' % idx2alpha(labels_true[0]))
print('Hyp: %s' % idx2alpha(labels_pred[0]))
else:
print('Ref: %s' % idx2phone(labels_true[0]))
print('Hyp: %s' % idx2phone(labels_pred[0]))
except IndexError:
if label_type == 'character':
print('Ref: %s' % idx2alpha(labels_true[0]))
print('Hyp: %s' % '')
else:
print('Ref: %s' % idx2phone(labels_true[0]))
print('Hyp: %s' % '')
# NOTE: This is for no prediction
if ler_train >= ler_train_pre:
not_improved_count += 1
else:
not_improved_count = 0
if ler_train < 0.05:
print('Modle is Converged.')
if save_params:
# Save model (check point)
checkpoint_file = './model.ckpt'
save_path = saver.save(sess,
checkpoint_file,
global_step=1)
print("Model saved in file: %s" % save_path)
break
ler_train_pre = ler_train
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=step,
value=ler_train)
feed_dict[learning_rate_pl] = learning_rate
duration_global = time.time() - start_time_global
print('Total time: %.3f sec' % (duration_global))
def check(self, encoder_type, bidirectional=False, label_type='char',
subsample=False, projection=False,
conv=False, batch_norm=False, activation='relu',
encoder_residual=False, encoder_dense_residual=False,
label_smoothing=False):
print('==================================================')
print(' label_type: %s' % label_type)
print(' encoder_type: %s' % encoder_type)
print(' bidirectional: %s' % str(bidirectional))
print(' projection: %s' % str(projection))
print(' subsample: %s' % str(subsample))
print(' conv: %s' % str(conv))
print(' batch_norm: %s' % str(batch_norm))
print(' activation: %s' % activation)
print(' encoder_residual: %s' % str(encoder_residual))
print(' encoder_dense_residual: %s' % str(encoder_dense_residual))
print(' label_smoothing: %s' % str(label_smoothing))
print('==================================================')
if conv or encoder_type == 'cnn':
# pattern 1
# conv_channels = [32, 32]
# conv_kernel_sizes = [[41, 11], [21, 11]]
# conv_strides = [[2, 2], [2, 1]]
# poolings = [[], []]
# pattern 2 (VGG like)
conv_channels = [64, 64]
conv_kernel_sizes = [[3, 3], [3, 3]]
conv_strides = [[1, 1], [1, 1]]
poolings = [[2, 2], [2, 2]]
fc_list = [786, 786]
else:
conv_channels = []
conv_kernel_sizes = []
conv_strides = []
poolings = []
fc_list = []
# Load batch data
splice = 1
num_stack = 1 if subsample or conv or encoder_type == 'cnn' else 2
xs, ys, x_lens, y_lens = generate_data(
label_type=label_type,
batch_size=2,
num_stack=num_stack,
splice=splice,
backend='chainer')
if label_type == 'char':
num_classes = 27
map_fn = idx2char
elif label_type == 'word':
num_classes = 11
map_fn = idx2word
# Load model
model = CTC(
input_size=xs[0].shape[-1] // splice // num_stack, # 120
encoder_type=encoder_type,
encoder_bidirectional=bidirectional,
encoder_num_units=256,
encoder_num_proj=256 if projection else 0,
encoder_num_layers=1 if not subsample else 2,
fc_list=fc_list,
dropout_input=0.1,
dropout_encoder=0.1,
num_classes=num_classes,
parameter_init_distribution='uniform',
parameter_init=0.1,
recurrent_weight_orthogonal=False,
init_forget_gate_bias_with_one=True,
subsample_list=[] if not subsample else [True] * 2,
num_stack=num_stack,
splice=splice,
input_channel=3,
conv_channels=conv_channels,
conv_kernel_sizes=conv_kernel_sizes,
conv_strides=conv_strides,
poolings=poolings,
activation=activation,
batch_norm=batch_norm,
label_smoothing_prob=0.1 if label_smoothing else 0,
weight_noise_std=0,
encoder_residual=encoder_residual,
encoder_dense_residual=encoder_dense_residual)
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
print("%s %d" % (name, num_params))
print("Total %.3f M parameters" % (model.total_parameters / 1000000))
# Define optimizer
learning_rate = 1e-3
model.set_optimizer(
'adam',
# 'adadelta',
learning_rate_init=learning_rate,
weight_decay=1e-6,
clip_grad_norm=5,
lr_schedule=None,
factor=None,
patience_epoch=None)
# Define learning rate controller
lr_controller = Controller(learning_rate_init=learning_rate,
backend='chainer',
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
# Train model
max_step = 300
start_time_step = time.time()
for step in range(max_step):
# Step for parameter update
loss = model(xs, ys, x_lens, y_lens)
model.optimizer.target.cleargrads()
model.cleargrads()
loss.backward()
loss.unchain_backward()
model.optimizer.update()
# Inject Gaussian noise to all parameters
if (step + 1) % 10 == 0:
# Compute loss
loss = model(xs, ys, x_lens, y_lens, is_eval=True)
# Decode
best_hyps, _, _ = model.decode(xs, x_lens, beam_width=1)
# TODO: fix beam search
str_ref = map_fn(ys[0, :y_lens[0]])
str_hyp = map_fn(best_hyps[0])
# Compute accuracy
try:
if label_type == 'char':
ler, _, _, _ = compute_wer(
ref=list(str_ref.replace('_', '')),
hyp=list(str_hyp.replace('_', '')),
normalize=True)
elif label_type == 'word':
ler, _, _, _ = compute_wer(ref=str_ref.split('_'),
hyp=str_hyp.split('_'),
normalize=True)
except:
ler = 1
duration_step = time.time() - start_time_step
print('Step %d: loss=%.3f / ler=%.3f / lr=%.5f (%.3f sec)' %
(step + 1, loss, ler, learning_rate, duration_step))
start_time_step = time.time()
# Visualize
print('Ref: %s' % str_ref)
print('Hyp: %s' % str_hyp)
if ler < 0.05:
print('Modle is Converged.')
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=step,
value=ler)
def check(self,
encoder_type,
bidirectional=False,
subsample=False,
projection=False,
conv=False,
batch_norm=False,
activation='relu',
encoder_residual=False,
encoder_dense_residual=False,
label_smoothing=False):
print('==================================================')
print(' encoder_type: %s' % encoder_type)
print(' bidirectional: %s' % str(bidirectional))
print(' projection: %s' % str(projection))
print(' subsample: %s' % str(subsample))
print(' conv: %s' % str(conv))
print(' batch_norm: %s' % str(batch_norm))
print(' encoder_residual: %s' % str(encoder_residual))
print(' encoder_dense_residual: %s' % str(encoder_dense_residual))
print(' label_smoothing: %s' % str(label_smoothing))
print('==================================================')
if conv or encoder_type == 'cnn':
# pattern 1
# conv_channels = [32, 32]
# conv_kernel_sizes = [[41, 11], [21, 11]]
# conv_strides = [[2, 2], [2, 1]]
# poolings = [[], []]
# pattern 2 (VGG like)
conv_channels = [64, 64]
conv_kernel_sizes = [[3, 3], [3, 3]]
conv_strides = [[1, 1], [1, 1]]
poolings = [[2, 2], [2, 2]]
fc_list = [786, 786]
else:
conv_channels = []
conv_kernel_sizes = []
conv_strides = []
poolings = []
fc_list = []
# Load batch data
num_stack = 1 if subsample or conv or encoder_type == 'cnn' else 2
splice = 1
xs, ys, ys_sub, x_lens, y_lens, y_lens_sub = generate_data(
label_type='word_char',
batch_size=2,
num_stack=num_stack,
splice=splice)
num_classes = 11
num_classes_sub = 27
# Load model
model = HierarchicalCTC(
input_size=xs.shape[-1] // splice // num_stack, # 120
encoder_type=encoder_type,
encoder_bidirectional=bidirectional,
encoder_num_units=256,
encoder_num_proj=256 if projection else 0,
encoder_num_layers=2,
encoder_num_layers_sub=1,
fc_list=fc_list,
fc_list_sub=fc_list,
dropout_input=0.1,
dropout_encoder=0.1,
main_loss_weight=0.8,
sub_loss_weight=0.2,
num_classes=num_classes,
num_classes_sub=num_classes_sub,
parameter_init_distribution='uniform',
parameter_init=0.1,
recurrent_weight_orthogonal=False,
init_forget_gate_bias_with_one=True,
subsample_list=[] if not subsample else [True, False],
num_stack=num_stack,
splice=splice,
input_channel=3,
conv_channels=conv_channels,
conv_kernel_sizes=conv_kernel_sizes,
conv_strides=conv_strides,
poolings=poolings,
batch_norm=batch_norm,
label_smoothing_prob=0.1 if label_smoothing else 0,
weight_noise_std=0,
encoder_residual=encoder_residual,
encoder_dense_residual=encoder_dense_residual)
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
print("%s %d" % (name, num_params))
print("Total %.3f M parameters" % (model.total_parameters / 1000000))
# Define optimizer
learning_rate = 1e-3
model.set_optimizer('adam',
learning_rate_init=learning_rate,
weight_decay=1e-6,
lr_schedule=False,
factor=0.1,
patience_epoch=5)
# Define learning rate controller
lr_controller = Controller(learning_rate_init=learning_rate,
backend='pytorch',
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
# Train model
max_step = 300
start_time_step = time.time()
for step in range(max_step):
# Step for parameter update
model.optimizer.zero_grad()
loss, loss_main, loss_sub = model(xs, ys, x_lens, y_lens, ys_sub,
y_lens_sub)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
model.optimizer.step()
if (step + 1) % 10 == 0:
# Compute loss
loss, loss_main, loss_sub = model(xs,
ys,
x_lens,
y_lens,
ys_sub,
y_lens_sub,
is_eval=True)
# Decode
best_hyps, _, _ = model.decode(xs,
x_lens,
beam_width=2,
task_index=0)
best_hyps_sub, _, _ = model.decode(xs,
x_lens,
beam_width=2,
task_index=1)
str_ref = idx2word(ys[0, :y_lens[0]])
str_hyp = idx2word(best_hyps[0])
str_ref_sub = idx2char(ys_sub[0, :y_lens_sub[0]])
str_hyp_sub = idx2char(best_hyps_sub[0])
# Compute accuracy
try:
wer, _, _, _ = compute_wer(ref=str_ref.split('_'),
hyp=str_hyp.split('_'),
normalize=True)
cer, _, _, _ = compute_wer(
ref=list(str_ref_sub.replace('_', '')),
hyp=list(str_hyp_sub.replace('_', '')),
normalize=True)
except:
wer = 1
cer = 1
duration_step = time.time() - start_time_step
print(
'Step %d: loss=%.3f(%.3f/%.3f) / wer=%.3f / cer=%.3f / lr=%.5f (%.3f sec)'
% (step + 1, loss, loss_main, loss_sub, wer, cer,
learning_rate, duration_step))
start_time_step = time.time()
# Visualize
print('Ref: %s' % str_ref)
print('Hyp (word): %s' % str_hyp)
print('Hyp (char): %s' % str_hyp_sub)
if cer < 0.1:
print('Modle is Converged.')
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=step,
value=wer)
def check(self,
encoder_type,
decoder_type,
bidirectional=False,
attention_type='location',
subsample=False,
projection=False,
ctc_loss_weight_sub=0,
conv=False,
batch_norm=False,
residual=False,
dense_residual=False,
num_heads=1,
backward_sub=False):
print('==================================================')
print(' encoder_type: %s' % encoder_type)
print(' bidirectional: %s' % str(bidirectional))
print(' projection: %s' % str(projection))
print(' decoder_type: %s' % decoder_type)
print(' attention_type: %s' % attention_type)
print(' subsample: %s' % str(subsample))
print(' ctc_loss_weight_sub: %s' % str(ctc_loss_weight_sub))
print(' conv: %s' % str(conv))
print(' batch_norm: %s' % str(batch_norm))
print(' residual: %s' % str(residual))
print(' dense_residual: %s' % str(dense_residual))
print(' backward_sub: %s' % str(backward_sub))
print(' num_heads: %s' % str(num_heads))
print('==================================================')
if conv or encoder_type == 'cnn':
# pattern 1
# conv_channels = [32, 32]
# conv_kernel_sizes = [[41, 11], [21, 11]]
# conv_strides = [[2, 2], [2, 1]]
# poolings = [[], []]
# pattern 2 (VGG like)
conv_channels = [64, 64]
conv_kernel_sizes = [[3, 3], [3, 3]]
conv_strides = [[1, 1], [1, 1]]
poolings = [[2, 2], [2, 2]]
else:
conv_channels = []
conv_kernel_sizes = []
conv_strides = []
poolings = []
# Load batch data
splice = 1
num_stack = 1 if subsample or conv or encoder_type == 'cnn' else 2
xs, ys, ys_sub, x_lens, y_lens, y_lens_sub = generate_data(
label_type='word_char',
batch_size=2,
num_stack=num_stack,
splice=splice,
backend='chainer')
num_classes = 11
num_classes_sub = 27
# Load model
model = HierarchicalAttentionSeq2seq(
input_size=xs[0].shape[-1] // splice // num_stack, # 120
encoder_type=encoder_type,
encoder_bidirectional=bidirectional,
encoder_num_units=320,
encoder_num_proj=320 if projection else 0,
encoder_num_layers=2,
encoder_num_layers_sub=1,
attention_type=attention_type,
attention_dim=128,
decoder_type=decoder_type,
decoder_num_units=320,
decoder_num_layers=1,
decoder_num_units_sub=320,
decoder_num_layers_sub=1,
embedding_dim=64,
embedding_dim_sub=32,
dropout_input=0.1,
dropout_encoder=0.1,
dropout_decoder=0.1,
dropout_embedding=0.1,
main_loss_weight=0.8,
sub_loss_weight=0.2 if ctc_loss_weight_sub == 0 else 0,
num_classes=num_classes,
num_classes_sub=num_classes_sub,
parameter_init_distribution='uniform',
parameter_init=0.1,
recurrent_weight_orthogonal=False,
init_forget_gate_bias_with_one=True,
subsample_list=[] if not subsample else [True, False],
subsample_type='drop' if not subsample else subsample,
bridge_layer=True,
init_dec_state='first',
sharpening_factor=1,
logits_temperature=1,
sigmoid_smoothing=False,
ctc_loss_weight_sub=ctc_loss_weight_sub,
attention_conv_num_channels=10,
attention_conv_width=201,
input_channel=3,
num_stack=num_stack,
splice=splice,
conv_channels=conv_channels,
conv_kernel_sizes=conv_kernel_sizes,
conv_strides=conv_strides,
poolings=poolings,
activation='relu',
batch_norm=batch_norm,
scheduled_sampling_prob=0.1,
scheduled_sampling_max_step=200,
label_smoothing_prob=0.1,
weight_noise_std=0,
encoder_residual=residual,
encoder_dense_residual=dense_residual,
decoder_residual=residual,
decoder_dense_residual=dense_residual,
decoding_order='attend_generate_update',
bottleneck_dim=256,
bottleneck_dim_sub=256,
backward_sub=backward_sub,
num_heads=num_heads,
num_heads_sub=num_heads)
# Count total parameters
for name in sorted(list(model.num_params_dict.keys())):
num_params = model.num_params_dict[name]
print("%s %d" % (name, num_params))
print("Total %.3f M parameters" % (model.total_parameters / 1000000))
# Define optimizer
learning_rate = 1e-3
model.set_optimizer('adam',
learning_rate_init=learning_rate,
weight_decay=1e-6,
lr_schedule=False,
factor=0.1,
patience_epoch=5)
# Define learning rate controller
lr_controller = Controller(learning_rate_init=learning_rate,
backend='chainer',
decay_start_epoch=20,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
# GPU setting
model.set_cuda(deterministic=False, benchmark=True)
# Train model
max_step = 300
start_time_step = time.time()
for step in range(max_step):
# Step for parameter update
loss, loss_main, loss_sub = model(xs, ys, x_lens, y_lens, ys_sub,
y_lens_sub)
model.optimizer.target.cleargrads()
model.cleargrads()
loss.backward()
loss.unchain_backward()
model.optimizer.update()
if (step + 1) % 10 == 0:
# Compute loss
loss, loss_main, loss_sub = model(xs,
ys,
x_lens,
y_lens,
ys_sub,
y_lens_sub,
is_eval=True)
# Decode
best_hyps, _, _ = model.decode(
xs,
x_lens,
beam_width=1,
# beam_width=2,
max_decode_len=30)
best_hyps_sub, _, _ = model.decode(
xs,
x_lens,
beam_width=1,
# beam_width=2,
max_decode_len=60,
task_index=1)
str_hyp = idx2word(best_hyps[0][:-1]).split('>')[0]
str_ref = idx2word(ys[0])
str_hyp_sub = idx2char(best_hyps_sub[0][:-1]).split('>')[0]
str_ref_sub = idx2char(ys_sub[0])
# Compute accuracy
try:
wer, _, _, _ = compute_wer(ref=str_ref.split('_'),
hyp=str_hyp.split('_'),
normalize=True)
cer, _, _, _ = compute_wer(
ref=list(str_ref_sub.replace('_', '')),
hyp=list(str_hyp_sub.replace('_', '')),
normalize=True)
except:
wer = 1
cer = 1
duration_step = time.time() - start_time_step
print(
'Step %d: loss=%.3f(%.3f/%.3f) / wer=%.3f / cer=%.3f / lr=%.5f (%.3f sec)'
% (step + 1, loss, loss_main, loss_sub, wer, cer,
learning_rate, duration_step))
start_time_step = time.time()
# Visualize
print('Ref: %s' % str_ref)
print('Hyp (word): %s' % str_hyp)
print('Hyp (char): %s' % str_hyp_sub)
if cer < 0.1:
print('Modle is Converged.')
break
# Update learning rate
model.optimizer, learning_rate = lr_controller.decay_lr(
optimizer=model.optimizer,
learning_rate=learning_rate,
epoch=step,
value=wer)
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
"""
if 'kanji' in params['label_type']:
map_file_path = '../metrics/mapping_files/' + \
params['label_type'] + '_' + params['train_data_size'] + '.txt'
elif 'kana' in params['label_type']:
map_file_path = '../metrics/mapping_files/' + \
params['label_type'] + '.txt'
# Load dataset
train_data = Dataset(
data_type='train', train_data_size=params['train_data_size'],
label_type=params['label_type'], map_file_path=map_file_path,
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'], map_file_path=map_file_path,
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_infer, 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, tower_decoder_outputs_train, tower_decoder_outputs_infer = model.compute_loss(
model.inputs_pl_list[i_gpu],
model.labels_pl_list[i_gpu],
model.inputs_seq_len_pl_list[i_gpu],
model.labels_seq_len_pl_list[i_gpu],
model.keep_prob_encoder_pl_list[i_gpu],
model.keep_prob_decoder_pl_list[i_gpu],
model.keep_prob_embedding_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_infer = model.decode(
tower_decoder_outputs_train,
tower_decoder_outputs_infer)
decode_ops_infer.append(decode_op_tower_infer)
# ler_op_tower = model.compute_ler(
# decode_op_tower, model.labels_pl_list[i_gpu])
ler_op_tower = model.compute_ler(
model.labels_st_true_pl_list[i_gpu],
model.labels_st_pred_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 param
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_train, inputs_seq_len, labels_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]
] = labels_train[i_gpu]
feed_dict_train[model.inputs_seq_len_pl_list[i_gpu]
] = inputs_seq_len[i_gpu]
feed_dict_train[model.labels_seq_len_pl_list[i_gpu]
] = labels_seq_len[i_gpu]
feed_dict_train[model.keep_prob_encoder_pl_list[i_gpu]
] = 1 - float(params['dropout_encoder'])
feed_dict_train[model.keep_prob_decoder_pl_list[i_gpu]
] = 1 - float(params['dropout_decoder'])
feed_dict_train[model.keep_prob_embedding_pl_list[i_gpu]
] = 1 - float(params['dropout_embedding'])
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_dev, inputs_seq_len, labels_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]
] = labels_dev[i_gpu]
feed_dict_dev[model.inputs_seq_len_pl_list[i_gpu]
] = inputs_seq_len[i_gpu]
feed_dict_dev[model.labels_seq_len_pl_list[i_gpu]
] = labels_seq_len[i_gpu]
feed_dict_dev[model.keep_prob_encoder_pl_list[i_gpu]
] = 1.0
feed_dict_dev[model.keep_prob_decoder_pl_list[i_gpu]
] = 1.0
feed_dict_dev[model.keep_prob_embedding_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_encoder_pl_list[i_gpu]] = 1.0
feed_dict_train[model.keep_prob_decoder_pl_list[i_gpu]] = 1.0
feed_dict_train[model.keep_prob_embedding_pl_list[i_gpu]] = 1.0
# Predict class ids
predicted_ids_train_list, summary_str_train = sess.run(
[decode_ops_infer, summary_train], feed_dict=feed_dict_train)
predicted_ids_dev_list, summary_str_dev = sess.run(
[decode_ops_infer, summary_dev], feed_dict=feed_dict_dev)
# Convert to sparsetensor to compute LER
feed_dict_ler_train = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_ler_train[model.labels_st_true_pl_list[i_gpu]] = list2sparsetensor(
labels_train[i_gpu],
padded_value=train_data.padded_value),
feed_dict_ler_train[model.labels_st_pred_pl_list[i_gpu]] = list2sparsetensor(
predicted_ids_train_list[i_gpu],
padded_value=train_data.padded_value)
feed_dict_ler_dev = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_ler_dev[model.labels_st_true_pl_list[i_gpu]] = list2sparsetensor(
labels_dev[i_gpu],
padded_value=dev_data.padded_value),
feed_dict_ler_dev[model.labels_st_pred_pl_list[i_gpu]] = list2sparsetensor(
predicted_ids_dev_list[i_gpu],
padded_value=dev_data.padded_value)
# Compute accuracy
# ler_train = sess.run(ler_op, feed_dict=feed_dict_ler_train)
# ler_dev = sess.run(ler_op, feed_dict=feed_dict_ler_dev)
ler_train = 1
ler_dev = 1
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
# TODO: fix this
# Update even files
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_infer,
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
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=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.save_path, '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
"""
if 'kanji' in params['label_type']:
map_file_path = '../metrics/mapping_files/' + \
params['label_type'] + '_' + params['train_data_size'] + '.txt'
elif 'kana' in params['label_type']:
map_file_path = '../metrics/mapping_files/' + \
params['label_type'] + '.txt'
# Load dataset
train_data = Dataset(data_type='train',
train_data_size=params['train_data_size'],
label_type=params['label_type'],
map_file_path=map_file_path,
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'],
map_file_path=map_file_path,
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_infer, 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, tower_decoder_outputs_train, tower_decoder_outputs_infer = model.compute_loss(
model.inputs_pl_list[i_gpu],
model.labels_pl_list[i_gpu],
model.inputs_seq_len_pl_list[i_gpu],
model.labels_seq_len_pl_list[i_gpu],
model.keep_prob_encoder_pl_list[i_gpu],
model.keep_prob_decoder_pl_list[i_gpu],
model.keep_prob_embedding_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_infer = model.decode(
tower_decoder_outputs_train,
tower_decoder_outputs_infer)
decode_ops_infer.append(decode_op_tower_infer)
# ler_op_tower = model.compute_ler(
# decode_op_tower, model.labels_pl_list[i_gpu])
ler_op_tower = model.compute_ler(
model.labels_st_true_pl_list[i_gpu],
model.labels_st_pred_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 param
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_train, inputs_seq_len, labels_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]] = labels_train[i_gpu]
feed_dict_train[model.inputs_seq_len_pl_list[
i_gpu]] = inputs_seq_len[i_gpu]
feed_dict_train[model.labels_seq_len_pl_list[
i_gpu]] = labels_seq_len[i_gpu]
feed_dict_train[
model.keep_prob_encoder_pl_list[i_gpu]] = 1 - float(
params['dropout_encoder'])
feed_dict_train[
model.keep_prob_decoder_pl_list[i_gpu]] = 1 - float(
params['dropout_decoder'])
feed_dict_train[
model.keep_prob_embedding_pl_list[i_gpu]] = 1 - float(
params['dropout_embedding'])
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_dev, inputs_seq_len, labels_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]] = labels_dev[i_gpu]
feed_dict_dev[model.inputs_seq_len_pl_list[
i_gpu]] = inputs_seq_len[i_gpu]
feed_dict_dev[model.labels_seq_len_pl_list[
i_gpu]] = labels_seq_len[i_gpu]
feed_dict_dev[
model.keep_prob_encoder_pl_list[i_gpu]] = 1.0
feed_dict_dev[
model.keep_prob_decoder_pl_list[i_gpu]] = 1.0
feed_dict_dev[
model.keep_prob_embedding_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_encoder_pl_list[i_gpu]] = 1.0
feed_dict_train[
model.keep_prob_decoder_pl_list[i_gpu]] = 1.0
feed_dict_train[
model.keep_prob_embedding_pl_list[i_gpu]] = 1.0
# Predict class ids
predicted_ids_train_list, summary_str_train = sess.run(
[decode_ops_infer, summary_train],
feed_dict=feed_dict_train)
predicted_ids_dev_list, summary_str_dev = sess.run(
[decode_ops_infer, summary_dev],
feed_dict=feed_dict_dev)
# Convert to sparsetensor to compute LER
feed_dict_ler_train = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_ler_train[model.labels_st_true_pl_list[
i_gpu]] = list2sparsetensor(
labels_train[i_gpu],
padded_value=train_data.padded_value),
feed_dict_ler_train[model.labels_st_pred_pl_list[
i_gpu]] = list2sparsetensor(
predicted_ids_train_list[i_gpu],
padded_value=train_data.padded_value)
feed_dict_ler_dev = {}
for i_gpu in range(len(gpu_indices)):
feed_dict_ler_dev[model.labels_st_true_pl_list[
i_gpu]] = list2sparsetensor(
labels_dev[i_gpu],
padded_value=dev_data.padded_value),
feed_dict_ler_dev[model.labels_st_pred_pl_list[
i_gpu]] = list2sparsetensor(
predicted_ids_dev_list[i_gpu],
padded_value=dev_data.padded_value)
# Compute accuracy
# ler_train = sess.run(ler_op, feed_dict=feed_dict_ler_train)
# ler_dev = sess.run(ler_op, feed_dict=feed_dict_ler_dev)
ler_train = 1
ler_dev = 1
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
# TODO: fix this
# Update even files
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_infer,
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
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=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.save_path, 'complete.txt'), 'w') as f:
f.write('')
def do_train(model, params):
"""Run training. If target labels are phone, the model is evaluated by PER
with 39 phones.
Args:
model: the model to train
params (dict): A dictionary of parameters
"""
# Load dataset
train_data = Dataset(data_type='train',
label_type=params['label_type'],
batch_size=params['batch_size'],
eos_index=params['eos_index'],
max_epoch=params['num_epoch'],
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=True)
dev_data = Dataset(data_type='dev',
label_type=params['label_type'],
batch_size=params['batch_size'],
eos_index=params['eos_index'],
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False)
if 'char' in params['label_type']:
test_data = Dataset(data_type='test',
label_type=params['label_type'],
batch_size=1,
eos_index=params['eos_index'],
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False)
else:
test_data = Dataset(data_type='test',
label_type='phone39',
batch_size=1,
eos_index=params['eos_index'],
splice=params['splice'],
num_stack=params['num_stack'],
num_skip=params['num_skip'],
sort_utt=False)
# TODO(hirofumi): add frame_stacking and splice
# Tell TensorFlow that the model will be built into the default graph
with tf.Graph().as_default():
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation (including model definition)
loss_op, att_logits, ctc_logits, decoder_outputs_train, decoder_outputs_infer = model.compute_loss(
model.inputs_pl_list[0], model.att_labels_pl_list[0],
model.inputs_seq_len_pl_list[0],
model.att_labels_seq_len_pl_list[0], model.ctc_labels_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=learning_rate_pl)
_, decode_op_infer = model.decoder(decoder_outputs_train,
decoder_outputs_infer,
decode_type='greedy',
beam_width=20)
ler_op = model.compute_ler(model.att_labels_st_true_pl,
model.att_labels_st_pred_pl)
# 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 param
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 param" %
(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
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 param
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, att_labels_train, ctc_labels, inputs_seq_len, att_labels_seq_len, _ = data
feed_dict_train = {
model.inputs_pl_list[0]:
inputs,
model.att_labels_pl_list[0]:
att_labels_train,
model.inputs_seq_len_pl_list[0]:
inputs_seq_len,
model.att_labels_seq_len_pl_list[0]:
att_labels_seq_len,
model.ctc_labels_pl_list[0]:
list2sparsetensor(
ctc_labels, padded_value=train_data.ctc_padded_value),
model.keep_prob_input_pl_list[0]:
params['dropout_input'],
model.keep_prob_hidden_pl_list[0]:
params['dropout_hidden'],
model.keep_prob_output_pl_list[0]:
params['dropout_output'],
learning_rate_pl:
learning_rate
}
# Update param
sess.run(train_op, feed_dict=feed_dict_train)
if (step + 1) % params['print_step'] == 0:
# Create feed dictionary for next mini batch (dev)
(inputs, att_labels_dev, ctc_labels, inputs_seq_len,
att_labels_seq_len, _), _ = dev_data().next()
feed_dict_dev = {
model.inputs_pl_list[0]:
inputs,
model.att_labels_pl_list[0]:
att_labels_dev,
model.inputs_seq_len_pl_list[0]:
inputs_seq_len,
model.att_labels_seq_len_pl_list[0]:
att_labels_seq_len,
model.ctc_labels_pl_list[0]:
list2sparsetensor(
ctc_labels,
padded_value=dev_data.ctc_padded_value),
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
# Predict class ids & update event files
predicted_ids_train, summary_str_train = sess.run(
[decode_op_infer, summary_train],
feed_dict=feed_dict_train)
predicted_ids_dev, summary_str_dev = sess.run(
[decode_op_infer, summary_dev],
feed_dict=feed_dict_dev)
summary_writer.add_summary(summary_str_train, step + 1)
summary_writer.add_summary(summary_str_dev, step + 1)
summary_writer.flush()
# Convert to sparsetensor to compute LER
feed_dict_ler_train = {
model.att_labels_true_st:
list2sparsetensor(att_labels_train,
padded_value=params['eos_index']),
model.att_labels_st_pred_pl:
list2sparsetensor(predicted_ids_train,
padded_value=params['eos_index'])
}
feed_dict_ler_dev = {
model.att_labels_true_st:
list2sparsetensor(att_labels_dev,
padded_value=params['eos_index']),
model.att_labels_st_pred_pl:
list2sparsetensor(predicted_ids_dev,
padded_value=params['eos_index'])
}
# Compute accuracy
ler_train = sess.run(ler_op, feed_dict=feed_dict_ler_train)
ler_dev = sess.run(ler_op, feed_dict=feed_dict_ler_dev)
csv_ler_train.append(ler_train)
csv_ler_dev.append(ler_dev)
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()
if 'char' in params['label_type']:
print('=== Dev Data Evaluation ===')
ler_dev_epoch = do_eval_cer(
session=sess,
decode_op=decode_op_infer,
model=model,
dataset=dev_data,
eval_batch_size=1)
print(' CER: %f %%' % (ler_dev_epoch * 100))
if ler_dev_epoch < ler_dev_best:
ler_dev_best = ler_dev_epoch
print('■■■ ↑Best Score (CER)↑ ■■■')
# Save model only when best accuracy is
# obtained (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 ===')
ler_test = do_eval_cer(
session=sess,
decode_op=decode_op_infer,
model=model,
dataset=test_data,
eval_batch_size=1)
print(' CER: %f %%' % (ler_test * 100))
else:
print('=== Dev Data Evaluation ===')
ler_dev_epoch = do_eval_per(
session=sess,
decode_op=decode_op_infer,
per_op=ler_op,
model=model,
dataset=dev_data,
label_type=params['label_type'],
eval_batch_size=1)
print(' PER: %f %%' % (ler_dev_epoch * 100))
if ler_dev_epoch < ler_dev_best:
ler_dev_best = ler_dev_epoch
print('■■■ ↑Best Score (PER)↑ ■■■')
# Save model only when best accuracy is
# obtained (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 ===')
ler_test = do_eval_per(
session=sess,
decode_op=decode_op_infer,
per_op=ler_op,
model=model,
dataset=test_data,
label_type=params['label_type'],
eval_batch_size=1)
print(' PER: %f %%' % (ler_test * 100))
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_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 check(self, encoder_type, label_type='character',
lstm_impl=None, time_major=True, save_params=False):
print('==================================================')
print(' encoder_type: %s' % encoder_type)
print(' label_type: %s' % label_type)
print(' lstm_impl: %s' % lstm_impl)
print(' time_major: %s' % str(time_major))
print(' save_params: %s' % str(save_params))
print('==================================================')
tf.reset_default_graph()
with tf.Graph().as_default():
# Load batch data
batch_size = 2
splice = 11 if encoder_type in ['vgg_blstm', 'vgg_lstm', 'cnn_zhang',
'vgg_wang', 'resnet_wang', 'cldnn_wang'] else 1
num_stack = 2
inputs, labels, inputs_seq_len = generate_data(
label_type=label_type,
model='ctc',
batch_size=batch_size,
num_stack=num_stack,
splice=splice)
# NOTE: input_size must be even number when using CudnnLSTM
# Define model graph
num_classes = 27 if label_type == 'character' else 61
model = CTC(encoder_type=encoder_type,
input_size=inputs[0].shape[-1] // splice // num_stack,
splice=splice,
num_stack=num_stack,
num_units=256,
num_layers=2,
num_classes=num_classes,
lstm_impl=lstm_impl,
parameter_init=0.1,
clip_grad_norm=5.0,
clip_activation=50,
num_proj=256,
weight_decay=1e-10,
# bottleneck_dim=50,
bottleneck_dim=None,
time_major=time_major)
# Define placeholders
model.create_placeholders()
learning_rate_pl = tf.placeholder(tf.float32, name='learning_rate')
# Add to the graph each operation
loss_op, 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])
train_op = model.train(loss_op,
optimizer='nestrov',
learning_rate=learning_rate_pl)
# NOTE: Adam does not run on CudnnLSTM
decode_op = model.decoder(logits,
model.inputs_seq_len_pl_list[0],
beam_width=20)
ler_op = model.compute_ler(decode_op, model.labels_pl_list[0])
# Define learning rate controller
learning_rate = 1e-4
lr_controller = Controller(learning_rate_init=learning_rate,
decay_start_epoch=50,
decay_rate=0.9,
decay_patient_epoch=10,
lower_better=True)
if save_params:
# Create a saver for writing training checkpoints
saver = tf.train.Saver(max_to_keep=None)
# Add the variable initializer operation
init_op = tf.global_variables_initializer()
# Count total parameters
if lstm_impl != 'CudnnLSTM':
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)))
# Make feed dict
feed_dict = {
model.inputs_pl_list[0]: inputs,
model.labels_pl_list[0]: list2sparsetensor(labels, padded_value=-1),
model.inputs_seq_len_pl_list[0]: inputs_seq_len,
model.keep_prob_pl_list[0]: 1.0,
learning_rate_pl: learning_rate
}
idx2phone = Idx2phone(map_file_path='./phone61.txt')
with tf.Session() as sess:
# Initialize parameters
sess.run(init_op)
# Wrapper for tfdbg
# sess = tf_debug.LocalCLIDebugWrapperSession(sess)
# Train model
max_steps = 1000
start_time_step = time.time()
for step in range(max_steps):
# for debug
# encoder_outputs = sess.run(
# model.encoder_outputs, feed_dict)
# print(encoder_outputs.shape)
# Compute loss
_, loss_train = sess.run(
[train_op, loss_op], feed_dict=feed_dict)
# Gradient check
# grads = sess.run(model.clipped_grads,
# feed_dict=feed_dict)
# for grad in grads:
# print(np.max(grad))
if (step + 1) % 10 == 0:
# Change to evaluation mode
feed_dict[model.keep_prob_pl_list[0]] = 1.0
# Compute accuracy
ler_train = sess.run(ler_op, feed_dict=feed_dict)
duration_step = time.time() - start_time_step
print('Step %d: loss = %.3f / ler = %.3f (%.3f sec) / lr = %.5f' %
(step + 1, loss_train, ler_train, duration_step, learning_rate))
start_time_step = time.time()
# Decode
labels_pred_st = sess.run(
decode_op, feed_dict=feed_dict)
# Visualize
try:
labels_pred = sparsetensor2list(
labels_pred_st, batch_size=batch_size)
if label_type == 'character':
print('Ref: %s' % idx2alpha(labels[0]))
print('Hyp: %s' % idx2alpha(labels_pred[0]))
else:
print('Ref: %s' % idx2phone(labels[0]))
print('Hyp: %s' % idx2phone(labels_pred[0]))
except IndexError:
if label_type == 'character':
print('Ref: %s' % idx2alpha(labels[0]))
print('Hyp: %s' % '')
else:
print('Ref: %s' % idx2phone(labels[0]))
print('Hyp: %s' % '')
# NOTE: This is for no prediction
if ler_train < 0.1:
print('Modle is Converged.')
if save_params:
# Save model (check point)
checkpoint_file = './model.ckpt'
save_path = saver.save(
sess, checkpoint_file, global_step=2)
print("Model saved in file: %s" % save_path)
break
# Update learning rate
learning_rate = lr_controller.decay_lr(
learning_rate=learning_rate,
epoch=step,
value=ler_train)
feed_dict[learning_rate_pl] = learning_rate