Closed BridgetteSong closed 1 year ago
HiFiGAN's
feature_loss
MPD
PD
all convolution's outputs
the last one
the last output
middle outputs
out = [] for i, layer in enumerate(self.convs):
out = []
for i, layer in enumerate(self.convs):
x = layer(x)
x = self.activation(x)
out.append(tf.reshape(x, [shape[0], -1, min(self.filters * (self.filter_scales ** (i + 1)), self.max_filters)]))
x = self.conv_post(x) x = tf.reshape(x, [shape[0], -1, self.out_filters]) out.append(x) return out
x = self.conv_post(x)
x = tf.reshape(x, [shape[0], -1, self.out_filters])
out.append(x)
return out
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.
HiFiGAN's
stable training,feature_loss
ofMPD
should be computed, so in the eachPD
, it should returnall convolution's outputs
, not onlythe last one
. But in eachPD
, it only returnsthe last output
, andmiddle outputs
are not saved. https://github.com/TensorSpeech/TensorFlowTTS/blob/136877136355c82d7ba474ceb7a8f133bd84767e/tensorflow_tts/models/hifigan.py#L309-L329out = []
for i, layer in enumerate(self.convs):
x = layer(x)
x = self.activation(x)
out.append(tf.reshape(x, [shape[0], -1, min(self.filters * (self.filter_scales ** (i + 1)), self.max_filters)]))
x = self.conv_post(x)
x = tf.reshape(x, [shape[0], -1, self.out_filters])
out.append(x)
return out