in models/wavenet.py [0:0]
def generate(self, spectrograms: Tensor, training: bool = False) -> Tensor:
"""
Generate a sample from this model.
Returns:
A 1D float tensor containing the output waveform.
"""
self.model.eval()
self.model.module.clear_buffer()
with torch.no_grad():
spectrograms = self.model.module.upsample_net(spectrograms)
seq_len = (
22050 if training else spectrograms.size(-1)
) # synthesize the first second only during training
batch_size = spectrograms.shape[0]
spectrograms = spectrograms.transpose(1, 2).contiguous()
if self.scalar_input:
x = spectrograms.new_zeros(batch_size, 1, 1)
else:
x = spectrograms.new_zeros(
batch_size, 1, self.config.model.out_channels
)
output = []
for t in tqdm(range(seq_len)):
# Conditioning features for single time step
ct = spectrograms[:, t, :].unsqueeze(1)
x = self.model.module.first_conv.incremental_forward(x)
skips = 0
for f in self.model.module.conv_layers:
x, h = f.incremental_forward(x, ct, None)
skips += h
skips *= math.sqrt(1.0 / len(self.model.module.conv_layers))
x = skips
for f in self.model.module.last_conv_layers:
try:
x = f.incremental_forward(x)
except AttributeError:
x = f(x)
x, output = self.get_x_from_dist(
self.config.model.output_distribution,
x,
history=output,
B=batch_size,
)
output = torch.stack(output).transpose(0, 1)
if self.config.model.input_type in ["mulaw", "mulaw-quantize"]:
if self.config.model.input_type == "mulaw-quantize":
output = torch.argmax(output, dim=2)
else:
output = self.float_2_label(output, self.config.model.quantize_channels)
output = self.expand(output.long())
elif self.config.model.input_type != "raw":
raise RuntimeError(
"Not supported input type: {}".format(self.config.model.input_type)
)
self.model.module.clear_buffer()
self.model.train()
return output.flatten()