def DiscretizedMixtureLogisticLoss(outputs,
targets,
hparams,
lengths=None,
mask=None,
max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, time_length, 1]
if mask is None:
mask = sequence_mask(lengths, max_len, True)
#[batch_size, time_length, dimension]
ones = tf.ones(
[tf.shape(mask)[0],
tf.shape(mask)[1],
tf.shape(targets)[-1]], tf.float32)
mask_ = mask * ones
losses = discretized_mix_logistic_loss(
outputs,
targets,
num_classes=hparams.quantize_channels,
log_scale_min=hparams.log_scale_min,
reduce=False)
with tf.control_dependencies(
[tf.assert_equal(tf.shape(losses), tf.shape(targets))]):
return tf.reduce_sum(losses * mask_) / tf.reduce_sum(mask_)
def GaussianMaximumLikelihoodEstimation(outputs,
targets,
hparams,
lengths=None,
mask=None,
max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, time_length, 1]
if mask is None:
mask = sequence_mask(lengths, max_len, True)
#[batch_size, time_length, dimension]
ones = tf.ones(
[tf.shape(mask)[0],
tf.shape(mask)[1],
tf.shape(targets)[-1]], tf.float32)
mask_ = mask * ones
losses = gaussian_maximum_likelihood_estimation_loss(
outputs,
targets,
log_scale_min_gauss=hparams.log_scale_min_gauss,
reduce=False)
with tf.control_dependencies(
[tf.assert_equal(tf.shape(losses), tf.shape(targets))]):
return tf.reduce_sum(losses * mask_) / tf.reduce_sum(mask_)
def MaskedCrossEntropyLoss(outputs,
targets,
lengths=None,
mask=None,
max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, time_length]
if mask is None:
mask = sequence_mask(lengths, max_len, False)
#One hot encode targets (outputs.shape[-1] = hparams.quantize_channels)
targets_ = tf.one_hot(targets, depth=tf.shape(outputs)[-1])
with tf.control_dependencies(
[tf.assert_equal(tf.shape(outputs), tf.shape(targets_))]):
losses = tf.nn.softmax_cross_entropy_with_logits_v2(logits=outputs,
labels=targets_)
with tf.control_dependencies(
[tf.assert_equal(tf.shape(mask), tf.shape(losses))]):
masked_loss = losses * mask
return tf.reduce_sum(masked_loss) / tf.count_nonzero(masked_loss,
dtype=tf.float32)
def GaussianMaximumLikelihoodEstimation(outputs,
targets,
hparams,
lengths=None,
mask=None,
max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, time_length, 1]
if mask is None:
mask = sequence_mask(lengths, max_len, True)
#[batch_size, time_length, dimension]
ones = tf.ones(
[tf.shape(mask)[0],
tf.shape(mask)[1],
tf.shape(targets)[-1]], tf.float32)
mask_ = mask * ones
## as the return value of gaussian_maximum_likelihood_estimation_loss(), lossses could be a number (if reduce=True)
#or could be a tensor with last element is [1] (if reduce=False)
losses = gaussian_maximum_likelihood_estimation_loss(
outputs,
targets,
log_scale_min_gauss=hparams.log_scale_min_gauss,
reduce=False)
print(
'losses in modules.py, gaussianMaximumlikehoodestimation function. {}'.
format(losses))
with tf.control_dependencies(
[tf.assert_equal(tf.shape(losses), tf.shape(targets))]):
return tf.reduce_sum(losses * mask_) / tf.reduce_sum(mask_)
def MaskedMeanSquaredError(outputs, targets, lengths=None, mask=None, max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, frames, 1]
if mask is None:
mask = sequence_mask(lengths, max_len, True)
#[batch_size, frames, freq]
ones = tf.ones([tf.shape(mask)[0], tf.shape(mask)[1], tf.shape(targets)[-1]], tf.float32)
mask_ = mask * ones
with tf.control_dependencies([tf.assert_equal(tf.shape(targets), tf.shape(mask_))]):
return tf.losses.mean_squared_error(labels=targets, predictions=outputs, weights=mask_)
def DiscretizedMixtureLogisticLoss(outputs, targets, hparams, lengths=None, mask=None, max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, time_length, 1]
if mask is None:
mask = sequence_mask(lengths, max_len, True)
#[batch_size, time_length, dimension]
ones = tf.ones([tf.shape(mask)[0], tf.shape(mask)[1], tf.shape(targets)[-1]], tf.float32)
mask_ = mask * ones
losses = discretized_mix_logistic_loss(
outputs, targets, num_classes=hparams.quantize_channels,
log_scale_min=hparams.log_scale_min, reduce=False)
with tf.control_dependencies([tf.assert_equal(tf.shape(losses), tf.shape(targets))]):
return tf.reduce_sum(losses * mask_) / tf.reduce_sum(mask_)
def MaskedCrossEntropyLoss(outputs, targets, lengths=None, mask=None, max_len=None):
if lengths is None and mask is None:
raise RuntimeError('Please provide either lengths or mask')
#[batch_size, time_length]
if mask is None:
mask = sequence_mask(lengths, max_len, False)
#One hot encode targets (outputs.shape[-1] = hparams.quantize_channels)
targets_ = tf.one_hot(targets, depth=tf.shape(outputs)[-1])
with tf.control_dependencies([tf.assert_equal(tf.shape(outputs), tf.shape(targets_))]):
losses = tf.nn.softmax_cross_entropy_with_logits_v2(logits=outputs, labels=targets_)
with tf.control_dependencies([tf.assert_equal(tf.shape(mask), tf.shape(losses))]):
masked_loss = losses * mask
return tf.reduce_sum(masked_loss) / tf.count_nonzero(masked_loss, dtype=tf.float32)
