def tower_loss(scope, feats, labels, seq_lens):
"""Calculate the total loss on a single tower running the deepSpeech model.
This function builds the graph for computing the loss per tower(GPU).
ARGS:
scope: unique prefix string identifying the
deepSpeech tower, e.g. 'tower_0'
feats: Tensor of shape BxFxT representing the
audio features (mfccs or spectrogram).
labels: sparse tensor holding labels of each utterance.
seq_lens: tensor of shape [batch_size] holding
the sequence length per input utterance.
Returns:
Tensor of shape [batch_size] containing
the total loss for a batch of data
"""
# Build inference Graph.
logits = deepSpeech.inference(feats, seq_lens, ARGS)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
strided_seq_lens = tf.div(seq_lens, ARGS.temporal_stride)
_ = deepSpeech.loss(logits, labels, strided_seq_lens)
# Assemble all of the losses for the current tower only.
losses = tf.get_collection('losses', scope)
# Calculate the total loss for the current tower.
total_loss = tf.add_n(losses, name='total_loss')
# Compute the moving average of all individual losses and the total loss.
loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
# Attach a scalar summary to all individual losses and the total loss;
# do the same for the averaged version of the losses.
for loss in losses + [total_loss]:
# Remove 'tower_[0-9]/' from the name in case this is a
# multi-GPU training session. This helps the clarity
# of presentation on tensorboard.
loss_name = re.sub('%s_[0-9]*/' % helper_routines.TOWER_NAME, '',
loss.op.name)
# Name each loss as '(raw)' and name the moving average
# version of the loss as the original loss name.
tf.scalar_summary(loss_name + '(raw)', loss)
tf.scalar_summary(loss_name, loss_averages.average(loss))
# Without this loss_averages_op would never run
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
return total_loss
def train():
"""
Train deepSpeech for a number of steps.
This function build a set of ops required to build the model and optimize
weights.
"""
# Learning rate set up
learning_rate, global_step = set_learning_rate()
# Create an optimizer that performs gradient descent.
# optimizer = tf.train.AdamOptimizer(learning_rate)
# forward pass to compute loss
freq_size = 161
inputs = tf.placeholder(tf.float32, [ARGS.batch_size, freq_size, None])
targets = tf.placeholder(tf.float32, [ARGS.batch_size, None])
max_seqlen = tf.placeholder(tf.int32, [ARGS.batch_size])
logits = deepSpeech.inference(inputs, max_seqlen, ARGS)
# ctcloss
loss = deepSpeech.loss(logits, targets, max_seqlen)
# backward optimize
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
# optimizer.minimize(loss)
grads_and_vars = optimizer.compute_gradients(loss)
clipped_grads_and_vars = [(tf.clip_by_value(grad,
clip_value_min=-400,
clip_value_max=400), var)
for grad, var in grads_and_vars]
apply_gradient_op = optimizer.apply_gradients(grads,
global_step=global_step)
train_op = apply_gradient_op
with g.as_default(), tf.device('/device:GPU:0'):
# 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.
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=ARGS.log_device_placement))
print("forbid the use of checkpoint")
sess.run(tf.global_variables_initializer())
# Start the queue runners.
tf.train.start_queue_runners(sess)
g.finalize()
# Run training loop
# run_train_loop(sess, (loss_op, train_op, summary_op), saver)
run_train_loop(sess, (logits, loss, train_op))
def tower_loss(sess, feats, labels, seq_lens):
"""Calculate the total loss on a single tower running the deepSpeech model.
This function builds the graph for computing the loss per tower(GPU).
ARGS:
feats: Tensor of shape BxFxT representing the
audio features (mfccs or spectrogram).
labels: sparse tensor holding labels of each utterance.
seq_lens: tensor of shape [batch_size] holding
the sequence length per input utterance.
Returns:
Tensor of shape [batch_size] containing
the total loss for a batch of data
"""
# Build inference Graph.
logits = deepSpeech.inference(sess, feats, seq_lens, ARGS)
# Build the portion of the Graph calculating the losses. Note that we will
# assemble the total_loss using a custom function below.
total_loss = deepSpeech.loss(logits, labels, seq_lens)
# Compute the moving average of all individual losses and the total loss.
# loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg')
# loss_averages_op = loss_averages.apply([total_loss])
# Attach a scalar summary to all individual losses and the total loss;
# do the same for the averaged version of the losses.
# loss_name = total_loss.op.name
# Name each loss as '(raw)' and name the moving average
# version of the loss as the original loss name.
# tf.summary.scalar(loss_name + '(raw)', total_loss)
# Without this loss_averages_op would never run
# with tf.control_dependencies([loss_averages_op]):
# total_loss = tf.identity(total_loss)
return total_loss