def get_vocoder():
if VOCODER == 'MelGAN':
vocoder = torch.hub.load("descriptinc/melgan-neurips", "load_melgan",
"multi_speaker")
vocoder.mel2wav.eval()
vocoder.mel2wav.cuda()
elif VOCODER == "WaveGlow":
# waveglow_path = 'models/waveglow_256channels_universal_v4.pt'
# vocoder = torch.load(waveglow_path)['model'].cuda().eval()
vocoder = torch.hub.load('nvidia/DeepLearningExamples:torchhub',
'nvidia_waveglow')
vocoder = vocoder.remove_weightnorm(vocoder)
vocoder.eval()
for m in vocoder.modules():
if 'Conv' in str(type(m)):
setattr(m, 'padding_mode', 'zeros')
vocoder.cuda()
else:
with open("hifigan/config.json", "r") as f:
config = json.load(f)
config = hifigan.AttrDict(config)
vocoder = hifigan.Generator(config)
ckpt = torch.load("hifigan/generator_universal.pth.tar")
vocoder.load_state_dict(ckpt["generator"])
vocoder.eval()
vocoder.remove_weight_norm()
vocoder.cuda()
return vocoder
def get_vocoder(config, device):
name = config["vocoder"]["model"]
speaker = config["vocoder"]["speaker"]
if name == "MelGAN":
if speaker == "LJSpeech":
vocoder = torch.hub.load(
"descriptinc/melgan-neurips", "load_melgan", "linda_johnson"
)
elif speaker == "universal":
vocoder = torch.hub.load(
"descriptinc/melgan-neurips", "load_melgan", "multi_speaker"
)
vocoder.mel2wav.eval()
vocoder.mel2wav.to(device)
elif name == "HiFi-GAN":
with open("hifigan/config.json", "r") as f:
config = json.load(f)
config = hifigan.AttrDict(config)
vocoder = hifigan.Generator(config)
if speaker == "LJSpeech":
ckpt = torch.load("hifigan/generator_LJSpeech.pth.tar")
elif speaker == "universal":
ckpt = torch.load("hifigan/generator_universal.pth.tar")
vocoder.load_state_dict(ckpt["generator"])
vocoder.eval()
vocoder.remove_weight_norm()
vocoder.to(device)
return vocoder
def get_vocoder(config, device):
name = config["vocoder"]["model"]
speaker = config["vocoder"]["speaker"]
if name == "MelGAN":
if speaker == "LJSpeech":
vocoder = torch.hub.load("descriptinc/melgan-neurips",
"load_melgan", "linda_johnson")
elif speaker == "universal":
vocoder = torch.hub.load("descriptinc/melgan-neurips",
"load_melgan", "multi_speaker")
vocoder.mel2wav.eval()
vocoder.mel2wav.to(device)
elif name == "HiFi-GAN":
with open(
"/content/drive/MyDrive/checkpoint/hifigan/UNIVERSAL_V1/config.json",
"r") as f:
config = json.load(f)
config = hifigan.AttrDict(config)
vocoder = hifigan.Generator(config)
# if speaker == "LJSpeech":
# ckpt = torch.load("hifigan/generator_LJSpeech.pth.tar")
# elif speaker == "universal":
# ckpt = torch.load("hifigan/generator_universal.pth.tar")
ckpt = torch.load(os.path.join(
'/content/drive/MyDrive/checkpoint/hifigan/UNIVERSAL_V1',
'g_03175000'),
map_location=device)
vocoder.load_state_dict(ckpt["generator"])
vocoder.eval()
vocoder.remove_weight_norm()
vocoder.to(device)
return vocoder
def get_vocoder(name, device):
if name == "MelGAN":
vocoder = torch.hub.load("descriptinc/melgan-neurips", "load_melgan",
"multi_speaker")
vocoder.mel2wav.to(device)
vocoder.mel2wav.eval()
elif name == 'MelGAN-LJ':
vocoder = torch.hub.load("descriptinc/melgan-neurips", "load_melgan",
"linda_johnson")
vocoder.mel2wav.to(device)
vocoder.mel2wav.eval()
elif name == "HiFi-GAN":
with open("hifigan/config.json", "r") as f:
config = json.load(f)
config = hifigan.AttrDict(config)
vocoder = hifigan.Generator(config)
ckpt = torch.load("hifigan/generator_universal.pth.tar")
vocoder.load_state_dict(ckpt["generator"])
vocoder.remove_weight_norm()
vocoder.to(device)
vocoder.eval()
elif name == "HiFi-GAN-LJ":
with open("hifigan/config.json", "r") as f:
config = json.load(f)
config = hifigan.AttrDict(config)
vocoder = hifigan.Generator(config)
ckpt = torch.load("hifigan/generator_LJSpeech.pth.tar")
vocoder.load_state_dict(ckpt["generator"])
vocoder.remove_weight_norm()
vocoder.to(device)
vocoder.eval()
else: # WaveGlow
vocoder = torch.hub.load('nvidia/DeepLearningExamples:torchhub',
'nvidia_waveglow')
vocoder = vocoder.remove_weightnorm(vocoder)
vocoder = vocoder.to(device)
vocoder.eval()
return vocoder
def get_vocoder():
name = hp.vocoder
speaker = hp.vocoder_speaker
if name == "MelGAN":
if speaker == "LJSpeech":
vocoder = torch.hub.load("descriptinc/melgan-neurips",
"load_melgan", "linda_johnson")
elif speaker == "universal":
vocoder = torch.hub.load("descriptinc/melgan-neurips",
"load_melgan", "multi_speaker")
vocoder.mel2wav.eval()
vocoder.mel2wav.to(device)
elif name == "HiFi-GAN":
with open("hifigan/config.json", "r") as f:
config = json.load(f)
config = hifigan.AttrDict(config)
vocoder = hifigan.Generator(config)
if speaker == "LJSpeech":
ckpt = torch.load("hifigan/generator_LJSpeech.pth.tar")
elif speaker == "universal":
ckpt = torch.load("hifigan/generator_universal.pth.tar")
vocoder.load_state_dict(ckpt["generator"])
vocoder.eval()
vocoder.remove_weight_norm()
vocoder.to(device)
elif name == "WaveGlow":
vocoder = torch.hub.load('nvidia/DeepLearningExamples:torchhub',
'nvidia_waveglow')
vocoder = vocoder.remove_weightnorm(vocoder)
vocoder.eval()
for m in vocoder.modules():
if 'Conv' in str(type(m)):
setattr(m, 'padding_mode', 'zeros')
vocoder.to(device)
return vocoder