reactor
This commit is contained in:
w-okada
parent
46ff05b04f
commit
9fe14f94e2
@ -74,6 +74,66 @@ class Pipeline(object):
|
||||
def setPitchExtractor(self, pitchExtractor: PitchExtractor):
|
||||
self.pitchExtractor = pitchExtractor
|
||||
|
||||
def extractPitch(self, audio_pad, if_f0, pitchf, f0_up_key, silence_front):
|
||||
try:
|
||||
if if_f0 == 1:
|
||||
pitch, pitchf = self.pitchExtractor.extract(
|
||||
audio_pad,
|
||||
pitchf,
|
||||
f0_up_key,
|
||||
self.sr,
|
||||
self.window,
|
||||
silence_front=silence_front,
|
||||
)
|
||||
# pitch = pitch[:p_len]
|
||||
# pitchf = pitchf[:p_len]
|
||||
pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
|
||||
pitchf = torch.tensor(pitchf, device=self.device, dtype=torch.float).unsqueeze(0)
|
||||
else:
|
||||
pitch = None
|
||||
pitchf = None
|
||||
except IndexError as e: # NOQA
|
||||
# print(e)
|
||||
# import traceback
|
||||
# traceback.print_exc()
|
||||
raise NotEnoughDataExtimateF0()
|
||||
return pitch, pitchf
|
||||
|
||||
def extractFeatures(self, feats, embOutputLayer, useFinalProj):
|
||||
with autocast(enabled=self.isHalf):
|
||||
try:
|
||||
feats = self.embedder.extractFeatures(feats, embOutputLayer, useFinalProj)
|
||||
if torch.isnan(feats).all():
|
||||
raise DeviceCannotSupportHalfPrecisionException()
|
||||
return feats
|
||||
except RuntimeError as e:
|
||||
if "HALF" in e.__str__().upper():
|
||||
raise HalfPrecisionChangingException()
|
||||
elif "same device" in e.__str__():
|
||||
raise DeviceChangingException()
|
||||
else:
|
||||
raise e
|
||||
|
||||
def infer(self, feats, p_len, pitch, pitchf, sid, out_size):
|
||||
try:
|
||||
with torch.no_grad():
|
||||
with autocast(enabled=self.isHalf):
|
||||
audio1 = (
|
||||
torch.clip(
|
||||
self.inferencer.infer(feats, p_len, pitch, pitchf, sid, out_size)[0][0, 0].to(dtype=torch.float32),
|
||||
-1.0,
|
||||
1.0,
|
||||
)
|
||||
* 32767.5
|
||||
).data.to(dtype=torch.int16)
|
||||
return audio1
|
||||
except RuntimeError as e:
|
||||
if "HALF" in e.__str__().upper():
|
||||
print("HalfPresicion Error:", e)
|
||||
raise HalfPrecisionChangingException()
|
||||
else:
|
||||
raise e
|
||||
|
||||
def exec(
|
||||
self,
|
||||
sid,
|
||||
@ -113,30 +173,6 @@ class Pipeline(object):
|
||||
pitchf = pitchf if repeat == 0 else np.zeros(p_len)
|
||||
out_size = out_size if repeat == 0 else None
|
||||
|
||||
# ピッチ検出
|
||||
try:
|
||||
if if_f0 == 1:
|
||||
pitch, pitchf = self.pitchExtractor.extract(
|
||||
audio_pad,
|
||||
pitchf,
|
||||
f0_up_key,
|
||||
self.sr,
|
||||
self.window,
|
||||
silence_front=silence_front,
|
||||
)
|
||||
# pitch = pitch[:p_len]
|
||||
# pitchf = pitchf[:p_len]
|
||||
pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
|
||||
pitchf = torch.tensor(pitchf, device=self.device, dtype=torch.float).unsqueeze(0)
|
||||
else:
|
||||
pitch = None
|
||||
pitchf = None
|
||||
except IndexError as e: # NOQA
|
||||
# print(e)
|
||||
# import traceback
|
||||
# traceback.print_exc()
|
||||
raise NotEnoughDataExtimateF0()
|
||||
|
||||
# tensor型調整
|
||||
feats = audio_pad
|
||||
if feats.dim() == 2: # double channels
|
||||
@ -144,21 +180,27 @@ class Pipeline(object):
|
||||
assert feats.dim() == 1, feats.dim()
|
||||
feats = feats.view(1, -1)
|
||||
|
||||
# ピッチ検出
|
||||
with Timer("main-process", True) as t:
|
||||
pitch, pitchf = self.extractPitch(audio_pad, if_f0, pitchf, f0_up_key, silence_front)
|
||||
print(f"[Perform(Pit)] {t.secs}")
|
||||
|
||||
# embedding
|
||||
with Timer("main-process", False) as te:
|
||||
with autocast(enabled=self.isHalf):
|
||||
try:
|
||||
feats = self.embedder.extractFeatures(feats, embOutputLayer, useFinalProj)
|
||||
if torch.isnan(feats).all():
|
||||
raise DeviceCannotSupportHalfPrecisionException()
|
||||
except RuntimeError as e:
|
||||
if "HALF" in e.__str__().upper():
|
||||
raise HalfPrecisionChangingException()
|
||||
elif "same device" in e.__str__():
|
||||
raise DeviceChangingException()
|
||||
else:
|
||||
raise e
|
||||
# print(f"[Embedding] {te.secs}")
|
||||
with Timer("main-process", True) as t:
|
||||
# with autocast(enabled=self.isHalf):
|
||||
# try:
|
||||
# feats = self.embedder.extractFeatures(feats, embOutputLayer, useFinalProj)
|
||||
# if torch.isnan(feats).all():
|
||||
# raise DeviceCannotSupportHalfPrecisionException()
|
||||
# except RuntimeError as e:
|
||||
# if "HALF" in e.__str__().upper():
|
||||
# raise HalfPrecisionChangingException()
|
||||
# elif "same device" in e.__str__():
|
||||
# raise DeviceChangingException()
|
||||
# else:
|
||||
# raise e
|
||||
feats = self.extractFeatures(feats, embOutputLayer, useFinalProj)
|
||||
print(f"[Perform(Emb)] {t.secs}")
|
||||
|
||||
# Index - feature抽出
|
||||
# if self.index is not None and self.feature is not None and index_rate != 0:
|
||||
@ -227,26 +269,11 @@ class Pipeline(object):
|
||||
p_len = torch.tensor([feats_len], device=self.device).long()
|
||||
|
||||
# 推論実行
|
||||
try:
|
||||
with torch.no_grad():
|
||||
with autocast(enabled=self.isHalf):
|
||||
audio1 = (
|
||||
torch.clip(
|
||||
self.inferencer.infer(feats, p_len, pitch, pitchf, sid, out_size)[0][0, 0].to(dtype=torch.float32),
|
||||
-1.0,
|
||||
1.0,
|
||||
)
|
||||
* 32767.5
|
||||
).data.to(dtype=torch.int16)
|
||||
except RuntimeError as e:
|
||||
if "HALF" in e.__str__().upper():
|
||||
print("11", e)
|
||||
raise HalfPrecisionChangingException()
|
||||
else:
|
||||
raise e
|
||||
with Timer("main-process", True) as t:
|
||||
audio1 = self.infer(feats, p_len, pitch, pitchf, sid, out_size)
|
||||
print(f"[Perform(Inf)] {t.secs}")
|
||||
|
||||
feats_buffer = feats.squeeze(0).detach().cpu()
|
||||
|
||||
if pitchf is not None:
|
||||
pitchf_buffer = pitchf.squeeze(0).detach().cpu()
|
||||
else:
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user