def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16):
mel_files = files_to_list(mel_files)
waveglow = torch.load(waveglow_path)['model']
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
if is_fp16:
waveglow.half()
for k in waveglow.convinv:
k.float()
for i, file_path in enumerate(mel_files):
stime = time.time()
file_name = os.path.splitext(os.path.basename(file_path))[0]
mel = torch.load(file_path)
mel = torch.autograd.Variable(mel.cuda())
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
stime2 = time.time()
with torch.no_grad():
audio = MAX_WAV_VALUE * waveglow.infer(mel, sigma=sigma)[0]
inf_time = time.time() - stime2
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
#print(audio_path)
len_audio = len(audio) / 22050.
print(
"{}: (audio length {:.2f} sec), (total computing time {:.2f} sec), (inference time: {:.2f} sec) "
.format(audio_path, len_audio,
time.time() - stime, inf_time))
def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16):
mel_files = files_to_list(mel_files)
waveglow = torch.load(waveglow_path)['model']
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
if is_fp16:
waveglow.half()
for k in waveglow.convinv:
k.float()
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
mel = torch.load(file_path)
mel = torch.autograd.Variable(mel.cuda())
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
#mel = torch.transpose(mel,1,2)
#print(mel.size())
with torch.no_grad():
audio = MAX_WAV_VALUE * waveglow.infer(mel, sigma=sigma)[0]
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16,
denoiser_strength):
mel_files = files_to_list(mel_files)
waveglow = torch.load(waveglow_path)['model']
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
if is_fp16:
from apex import amp
waveglow, _ = amp.initialize(waveglow, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(waveglow).cuda()
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
mel = torch.load(file_path)
mel = torch.autograd.Variable(mel.cuda())
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
with torch.no_grad():
audio = waveglow.infer(mel, sigma=sigma)
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16,
denoiser_strength):
mel_files = files_to_list(mel_files)
waveglow = torch.load(waveglow_path)['model']
for m in waveglow.modules():
if 'Conv' in str(type(m)):
setattr(m, 'padding_mode', 'zeros')
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
if is_fp16:
from apex import amp
waveglow, _ = amp.initialize(waveglow, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(waveglow).cuda()
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
#print(file_name)
mel = torch.load(file_path)
# print("mel",mel)
#print(mel.shape)
mel = torch.autograd.Variable(mel.cuda())
# print("mel",mel)
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
# print("mel",mel)
print(torch.min(mel),torch.max(mel))
with torch.no_grad():
audio = waveglow.infer(mel, sigma=sigma)
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
k.append(abs(audio).max().item())
#print(min(k),max(k))
#audio = audio*18000*abs(audio).max()/0.99
#print("audio",audio)
#print((audio).min().item(),(audio).max().item())
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(
output_dir, "{}_synthesis_sig0.7_d_0.1.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
def __init__(self, training_files, num_frame, filter_length, hop_length,
win_length, sampling_rate, mel_fmin, mel_fmax):
self.audio_files = ms.files_to_list(training_files)
self.all_length = torch.tensor(0, dtype=torch.long)
self.hop_length = hop_length
self.win_length = win_length
random.seed(4321)
random.shuffle(self.audio_files)
self.stft = ms.TacotronSTFT(filter_length=filter_length,
hop_length=self.hop_length,
win_length=self.win_length,
sampling_rate=sampling_rate,
mel_fmin=mel_fmin,
mel_fmax=mel_fmax)
self.segment = num_frame * self.hop_length + self.win_length
self.sampling_rate = sampling_rate
def main(mel_files, squeezewave_path, sigma, output_dir, sampling_rate, is_fp16,
denoiser_strength):
tic_prepare= time.time()
mel_files = files_to_list(mel_files)
squeezewave = torch.load(squeezewave_path)['model']
squeezewave = squeezewave.remove_weightnorm(squeezewave)
squeezewave.cuda().eval()
if is_fp16:
from apex import amp
squeezewave, _ = amp.initialize(squeezewave, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(squeezewave).cuda()
toc_prepare = time.time()
dur_prepare = toc_prepare - tic_prepare
print("prepare model {:3.2}sec".format(dur_prepare) )
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
mel = torch.load(file_path)
mel = torch.autograd.Variable(mel.cuda())
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
tic=time.time()
with torch.no_grad():
audio = squeezewave.infer(mel, sigma=sigma).float()
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
audio = audio * MAX_WAV_VALUE
toc=time.time()
dur = toc -tic
audio = audio.squeeze()
audio = audio.cpu().numpy()
len_wav = len(audio)
sec_wav = len_wav/sampling_rate
samples_sec = len_wav / dur
audio = audio.astype('int16')
audio_path = os.path.join(
output_dir, "{}_s{}.wav".format(file_name,sigma))
write(audio_path, sampling_rate, audio)
print("{} it took {:4.3f}sec for {:4.3f}sec {:4.2f}K sample 22Khz Audio files : RTF {:4.3f} {:4.3f}X {:4.2f}Ksamples/sec "
.format(audio_path, dur, sec_wav, len_wav/1000, dur/sec_wav, sec_wav/dur , samples_sec/1000 ) )
def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16,
denoiser_strength):
mel_files = files_to_list(mel_files) #测试集mel谱list
waveglow = torch.load(waveglow_path)['model'] #加载模型
waveglow = waveglow.remove_weightnorm(waveglow) #？移除权重归一化
waveglow.cuda().eval() #cuda()拷贝进gpu #？变成测试模式，dropout和BN在训练时和测不一样
#apex加速
if is_fp16:
from apex import amp
waveglow, _ = amp.initialize(waveglow, [], opt_level="O3")
# denoiser_strength=0
if denoiser_strength > 0:
denoiser = Denoiser(waveglow).cuda()
for i, file_path in enumerate(mel_files):
#file_name-对应的wav
file_name = os.path.splitext(os.path.basename(file_path))[0]
#加载MFCC特征，80个滤波器
mel = torch.load(file_path)
#mel={key:mel[key].cuda() for key in mel}
#封装数据
mel = torch.autograd.Variable(mel.cuda())
#80，375 -> 1*80*375
mel = torch.unsqueeze(mel, 0)
#变成fp16数据以便apex加速
mel = mel.half() if is_fp16 else mel
#反向传播不会自动求导
with torch.no_grad():
#生成1*96000Tensor数据,x为原始音频，z为mel谱
audio = waveglow.infer(mel, sigma=sigma)
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
#为了转成wav？
audio = audio * MAX_WAV_VALUE
#变成1维数据
audio = audio.squeeze()
#在cpu中转成numpy
audio = audio.cpu().numpy()
#改变类型
audio = audio.astype('int16')
#生成数据存储位置
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
#写入音频
print(audio_path)
def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16,
denoiser_strength):
mel_files = files_to_list(mel_files)
waveglow = torch.load(waveglow_path)['model']
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
if is_fp16:
from apex import amp
waveglow, _ = amp.initialize(waveglow, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(waveglow).cuda()
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
print('Loading file: ', file_path)
if file_path.find('.pt') != -1:
print('load by torch')
mel = torch.load(file_path)
elif file_path.find('.npy') != -1:
print('load by numpy')
mel = np.load(file_path)
mel = torch.from_numpy(mel)
print(f"original mel shape: {mel.shape}")
mel = torch.autograd.Variable(mel.cuda())
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
print(f"mel shape right before using waveglow: {mel.shape}")
with torch.no_grad():
audio = waveglow.infer(mel, sigma=sigma)
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
audio = audio.cpu().numpy()
audio = audio.astype('int16')
if not os.path.exists(output_dir):
os.mkdir(output_dir)
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
def main(mel_files, waveglow_path, sigma, output_dir, sampling_rate, is_fp16,
denoiser_strength):
mel_files = files_to_list(mel_files)
waveglow = torch.load(waveglow_path)['model']
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
if is_fp16:
from apex import amp
waveglow, _ = amp.initialize(waveglow, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(waveglow).cuda()
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
if True:
# Processing for generic mel files
shape = tuple(np.fromfile(file_path, count=2, dtype=np.int32))
mel = np.memmap(file_path, offset=8, dtype=np.float32, shape=shape)
# mel = mel[1:1000,:]
mel = mel.transpose()
mel = torch.from_numpy(mel)
else:
mel = torch.load(file_path)
mel = torch.autograd.Variable(mel.cuda())
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
with torch.no_grad():
audio = waveglow.infer(mel, sigma=sigma)
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(output_dir, "{}.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
def main(mel_files, squeezewave_path, sigma, output_dir, sampling_rate,
is_fp16, denoiser_strength):
mel_files = files_to_list(mel_files)
#device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
device = torch.device('cpu')
squeezewave = torch.load(squeezewave_path, map_location=device)['model']
squeezewave = squeezewave.remove_weightnorm(squeezewave)
squeezewave.eval()
if is_fp16:
from apex import amp
squeezewave, _ = amp.initialize(squeezewave, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(squeezewave)
start = time.time()
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
mel = torch.load(file_path, map_location=device)
mel = torch.autograd.Variable(mel)
mel = mel.half()
with torch.no_grad():
audio = squeezewave.infer(mel, sigma=sigma).float()
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
end = time.time()
print("Squeezewave vocoder time")
print(end - start)
def main(mel_files, squeezewave_path, sigma, output_dir, sampling_rate,
is_fp16, denoiser_strength, device):
mel_files = files_to_list(mel_files)
squeezewave = torch.load(squeezewave_path, map_location=device)['model']
squeezewave.device = device # hack for loading model trained on gpu to cpu
squeezewave = squeezewave.remove_weightnorm(squeezewave)
squeezewave.to(device=device).eval()
if is_fp16:
from apex import amp
squeezewave, _ = amp.initialize(squeezewave, [], opt_level="O3")
if denoiser_strength > 0:
denoiser = Denoiser(squeezewave).to(device=device)
for i, file_path in enumerate(mel_files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
mel = torch.load(file_path)
if len(mel.shape) > 2:
mel = mel.squeeze()
print(f"squeezed to {mel.shape}")
assert len(mel.shape) == 2
mel = torch.autograd.Variable(mel.to(device=device))
mel = torch.unsqueeze(mel, 0)
mel = mel.half() if is_fp16 else mel
with torch.no_grad():
audio = squeezewave.infer(mel, sigma=sigma).float()
if denoiser_strength > 0:
audio = denoiser(audio, denoiser_strength)
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
audio = audio.cpu().numpy()
audio = audio.astype('int16')
audio_path = os.path.join(output_dir,
"{}_synthesis.wav".format(file_name))
write(audio_path, sampling_rate, audio)
print(audio_path)
import argparse
import os
import subprocess
from mel2samp import files_to_list
if __name__ == "__main__":
# Get defaults so it can work with no Sacred
parser = argparse.ArgumentParser()
parser.add_argument('-f', "--filelist_path", required=True)
parser.add_argument('-o',
'--output_dir',
type=str,
help='Output directory')
args = parser.parse_args()
filepaths = files_to_list(args.filelist_path)
for filepath in filepaths:
source_wav = filepath
filename = os.path.basename(filepath)
dest_wav = f'{args.output_dir}/{filename}'
command = f'sox {source_wav} {dest_wav} remix 1,2 rate 22050'
subprocess.run(command, shell=True)
