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.
_ = deepSpeech.loss(logits, labels, 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.summary.scalar(loss_name + '(raw)', loss)
tf.summary.scalar(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 tower_loss(sess, 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(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)
tf.summary.scalar(loss_name, loss_averages.average(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
def evaluate():
""" Evaluate deepSpeech modelfor a number of steps."""
with tf.Graph().as_default() as graph:
# Get feats and labels for deepSpeech.
feats, labels, seq_lens = deepSpeech.inputs(ARGS.eval_data,
data_dir=ARGS.data_dir,
batch_size=ARGS.batch_size,
use_fp16=ARGS.use_fp16,
shuffle=True)
session = tf.Session()
# Build ops that computes the logits predictions from the
# inference model.
ARGS.keep_prob = 1.0 # Disable dropout during testing.
logits = deepSpeech.inference(session, feats, seq_lens, ARGS)
# Calculate predictions.
output_log_prob = tf.nn.log_softmax(logits)
decoder = tf.nn.ctc_greedy_decoder
strided_seq_lens = deepSpeech.get_rnn_seqlen(seq_lens)
predictions = decoder(output_log_prob, strided_seq_lens)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
ARGS.moving_avg_decay)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(ARGS.eval_dir, graph)
while True:
eval_once(session, saver, summary_writer, predictions, summary_op,
labels)
if ARGS.run_once:
break
time.sleep(ARGS.eval_interval_secs)