def build_acoustic_training_rnn(is_mpi,is_chief, hyper_params, prog_params, train_set, test_set):
model = AcousticModel(hyper_params["num_layers"], hyper_params["hidden_size"], hyper_params["batch_size"],
hyper_params["max_input_seq_length"], hyper_params["max_target_seq_length"],
hyper_params["input_dim"], hyper_params["batch_normalization"],
hyper_params["char_map_length"])
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(train_set, hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"], hyper_params["signal_processing"],
hyper_params["char_map"])
#train_dataset = train_dataset.shuffle(10,reshuffle_each_iteration=True)
v_iterator = None
if test_set is []:
t_iterator = model.add_dataset_input(train_dataset)
else:
test_dataset = model.build_dataset(test_set, hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"], hyper_params["signal_processing"],
hyper_params["char_map"])
# Build the input stream from the different datasets
t_iterator, v_iterator = model.add_datasets_input(train_dataset, test_dataset)
# Create the model
#tensorboard_dir
model.create_training_rnn(is_mpi, hyper_params["dropout_input_keep_prob"], hyper_params["dropout_output_keep_prob"],
hyper_params["grad_clip"], hyper_params["learning_rate"],
hyper_params["lr_decay_factor"], use_iterator=True,is_sync=prog_params['is_sync'],is_chief=is_chief)
if is_chief:
model.add_tensorboard(prog_params["train_dir"], prog_params["timeline"])
return model, t_iterator, v_iterator
def build_acoustic_training_rnn(is_chief, is_ditributed, sess, hyper_params,
prog_params, train_set, test_set):
model = AcousticModel(
hyper_params["num_layers"], hyper_params["hidden_size"],
hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"], hyper_params["input_dim"],
hyper_params["batch_normalization"], hyper_params["char_map_length"])
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(train_set, hyper_params["batch_size"],
hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"],
hyper_params["signal_processing"],
hyper_params["char_map"])
train_dataset = train_dataset.shuffle(10, reshuffle_each_iteration=True)
v_iterator = None
if test_set is []:
t_iterator = model.add_dataset_input(train_dataset)
else:
test_dataset = model.build_dataset(
test_set, hyper_params["batch_size"],
hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"],
hyper_params["signal_processing"], hyper_params["char_map"])
# Build the input stream from the different datasets
t_iterator, v_iterator = model.add_datasets_input(
train_dataset, test_dataset)
# Create the model
#tensorboard_dir
model.create_training_rnn(is_chief,
is_ditributed,
hyper_params["dropout_input_keep_prob"],
hyper_params["dropout_output_keep_prob"],
hyper_params["grad_clip"],
hyper_params["learning_rate"],
hyper_params["lr_decay_factor"],
use_iterator=True)
model.add_tensorboard(sess, prog_params["train_dir"],
prog_params["timeline"])
sv = None
if is_ditributed:
init_op = tf.global_variables_initializer()
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=prog_params["train_dir"],
init_op=init_op,
recovery_wait_secs=1,
summary_op=None,
global_step=model.global_step)
model.supervisor = sv
else:
model.initialize(sess)
model.restore(sess, prog_params["train_dir"])
# Override the learning rate if given on the command line
if prog_params["learn_rate"] is not None:
model.set_learning_rate(sess, prog_params["learn_rate"])
return sv, model, t_iterator, v_iterator
def build_acoustic_training_rnn(sess, hyper_params, prog_params, train_set,
test_set):
model = AcousticModel(
hyper_params["num_layers"], hyper_params["hidden_size"],
hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"], hyper_params["input_dim"],
hyper_params["batch_normalization"], hyper_params["char_map_length"])
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(train_set, hyper_params["batch_size"],
hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"],
hyper_params["signal_processing"],
hyper_params["char_map"])
v_iterator = None
if test_set is []:
t_iterator = model.add_dataset_input(train_dataset)
sess.run(t_iterator.initializer)
else:
test_dataset = model.build_dataset(
test_set, hyper_params["batch_size"],
hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"],
hyper_params["signal_processing"], hyper_params["char_map"])
# Build the input stream from the different datasets
t_iterator, v_iterator = model.add_datasets_input(
train_dataset, test_dataset)
sess.run(t_iterator.initializer)
sess.run(v_iterator.initializer)
# Create the model
model.create_training_rnn(hyper_params["dropout_input_keep_prob"],
hyper_params["dropout_output_keep_prob"],
hyper_params["grad_clip"],
hyper_params["learning_rate"],
hyper_params["lr_decay_factor"],
use_iterator=True)
model.add_tensorboard(sess, hyper_params["tensorboard_dir"],
prog_params["tb_name"], prog_params["timeline"])
model.initialize(sess)
model.restore(sess, hyper_params["checkpoint_dir"] + "/acoustic/")
# Override the learning rate if given on the command line
if prog_params["learn_rate"] is not None:
model.set_learning_rate(sess, prog_params["learn_rate"])
return model, t_iterator, v_iterator
def build_acoustic_training_rnn(sess, hyper_params, prog_params, train_set, test_set):
model = AcousticModel(hyper_params["num_layers"], hyper_params["hidden_size"], hyper_params["batch_size"],
hyper_params["max_input_seq_length"], hyper_params["max_target_seq_length"],
hyper_params["input_dim"], hyper_params["batch_normalization"],
hyper_params["char_map_length"])
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset(train_set, hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"], hyper_params["signal_processing"],
hyper_params["char_map"])
v_iterator = None
if test_set is []:
t_iterator = model.add_dataset_input(train_dataset)
sess.run(t_iterator.initializer)
else:
test_dataset = model.build_dataset(test_set, hyper_params["batch_size"], hyper_params["max_input_seq_length"],
hyper_params["max_target_seq_length"], hyper_params["signal_processing"],
hyper_params["char_map"])
# Build the input stream from the different datasets
t_iterator, v_iterator = model.add_datasets_input(train_dataset, test_dataset)
sess.run(t_iterator.initializer)
sess.run(v_iterator.initializer)
# Create the model
model.create_training_rnn(hyper_params["dropout_input_keep_prob"], hyper_params["dropout_output_keep_prob"],
hyper_params["grad_clip"], hyper_params["learning_rate"],
hyper_params["lr_decay_factor"], use_iterator=True)
model.add_tensorboard(sess, hyper_params["tensorboard_dir"], prog_params["tb_name"], prog_params["timeline"])
model.initialize(sess)
model.restore(sess, hyper_params["checkpoint_dir"] + "/acoustic/")
# Override the learning rate if given on the command line
if prog_params["learn_rate"] is not None:
model.set_learning_rate(sess, prog_params["learn_rate"])
return model, t_iterator, v_iterator
def test_create_training_rnn_with_iterators(self):
tf.reset_default_graph()
with tf.Session():
model = AcousticModel(self.num_layers, self.hidden_size, self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim, self.normalization, self.num_labels)
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset([["/file/path", "Test", 10]], self.batch_size,
self.max_input_seq_length, self.max_target_seq_length,
self.signal_processing, ENGLISH_CHAR_MAP)
model.add_dataset_input(train_dataset)
model.create_training_rnn(self.input_keep_prob, self.output_keep_prob, self.grad_clip,
self.learning_rate, self.lr_decay_factor, use_iterator=True)
def test_create_training_rnn_with_iterators(self):
tf.reset_default_graph()
with tf.Session():
model = AcousticModel(self.num_layers, self.hidden_size,
self.batch_size, self.max_input_seq_length,
self.max_target_seq_length, self.input_dim,
self.normalization, self.num_labels)
# Create a Dataset from the train_set and the test_set
train_dataset = model.build_dataset([["/file/path", "Test", 10]],
self.batch_size,
self.max_input_seq_length,
self.max_target_seq_length,
self.signal_processing,
ENGLISH_CHAR_MAP)
model.add_dataset_input(train_dataset)
model.create_training_rnn(self.input_keep_prob,
self.output_keep_prob,
self.grad_clip,
self.learning_rate,
self.lr_decay_factor,
use_iterator=True)