def infer(self, x, chunk_length=None, chunk_overlap=0):
with torch.no_grad():
x = self.get_cond_input(x)
length = x.shape[-1]
if chunk_length is None or length <= chunk_length:
chunks = [x]
else:
n = math.ceil(
(length - chunk_overlap) / (chunk_length - chunk_overlap))
chunk_length = math.ceil(length / n) + chunk_overlap
chunks = [
x[..., onset:onset + chunk_length]
for onset in range(0, length -
chunk_overlap, chunk_length -
chunk_overlap)
]
audio = []
for i, xi in enumerate(chunks):
if xi.device == 'cpu':
raise NotImplementedError
else:
from .nv_wavenet.nv_wavenet import Impl
xi = self.nv_wavenet.infer(xi, Impl.AUTO)
xi = mu_law_decode(xi, self.n_out_channels)
if i > 0:
xi = xi[..., chunk_overlap:]
audio.append(xi)
audio = torch.cat(tuple(audio), dim=-1)
if self.fading is not None:
assert self.fading in ['half', 'full']
pad_width = self.upsamp_window - self.upsamp_stride
if self.fading == 'half':
pad_width //= 2
audio = audio[..., pad_width:]
return audio
def infer(self, x, chunk_length=None, chunk_overlap=0):
with torch.no_grad():
x = self.get_cond_input(x)
x = x.view(x.size(0), self.n_layers, -1, x.size(2))
# This makes the data channels x batch x num_layers x samples
x = x.permute(2, 0, 1, 3)
length = x.shape[-1]
if chunk_length is None or length <= chunk_length:
chunks = [x]
else:
n = math.ceil(
(length - chunk_overlap) / (chunk_length - chunk_overlap))
chunk_length = math.ceil(length / n) + chunk_overlap
chunks = [
x[..., onset:onset + chunk_length]
for onset in range(0, length -
chunk_overlap, chunk_length -
chunk_overlap)
]
audio = []
for i, xi in enumerate(chunks):
if xi.device == 'cpu':
raise NotImplementedError
else:
from .nv_wavenet.nv_wavenet import Impl
xi = self.nv_wavenet.infer(xi, Impl.AUTO)
torch.cuda.synchronize(xi.device)
xi = mu_law_decode(xi, self.n_out_channels)
if i > 0:
xi = xi[..., chunk_overlap:]
audio.append(xi)
audio = torch.cat(tuple(audio), dim=-1)
return audio
def review(self, inputs, outputs):
predictions, targets = outputs
ce = torch.nn.CrossEntropyLoss(reduction='none')(predictions, targets)
summary = dict(
loss=ce.mean(),
scalars=dict(),
histograms=dict(reconstruction_ce=ce),
audios=dict(target=(inputs[self.audio_key][0], self.sample_rate),
decode=(mu_law_decode(
torch.argmax(outputs[0][0], dim=0),
mu_quantization=self.wavenet.n_out_channels),
self.sample_rate)),
images=dict())
return summary
def infer_gpu(self, x, device=0):
self.cuda(device)
from .nv_wavenet.nv_wavenet import Impl
with torch.no_grad():
cond_input = self.get_cond_input(x.cuda(device))
audio = self.nv_wavenet.infer(cond_input, Impl.AUTO)
audio = mu_law_decode(audio, self.n_out_channels)
self.cpu()
if self.fading is not None:
assert self.fading in ['half', 'full']
pad_width = self.upsamp_window - self.upsamp_stride
if self.fading == 'half':
pad_width //= 2
audio = audio[..., pad_width:]
return audio
def infer_gpu(self, x):
self.cuda()
from .nv_wavenet.nv_wavenet import Impl
cond_input = self.get_cond_input(x.cuda())
audio = self.nv_wavenet.infer(cond_input, Impl.AUTO)
return mu_law_decode(audio, self.n_out_channels)