def get_model(model_name, model_config, to_cuda):
""" Code chooses a model based on name"""
model = None
if model_name == 'Tacotron2':
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
model = WaveGlow(**model_config)
else:
raise NotImplementedError(model_name)
if to_cuda:
model = model.cuda()
return model
def get_model(model_name, model_config, cpu_run,
uniform_initialize_bn_weight=False, forward_is_infer=False):
""" Code chooses a model based on name"""
model = None
if model_name == 'Tacotron2':
if forward_is_infer:
class Tacotron2__forward_is_infer(Tacotron2):
def forward(self, inputs, input_lengths):
return self.infer(inputs, input_lengths)
model = Tacotron2__forward_is_infer(**model_config)
else:
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
if forward_is_infer:
class WaveGlow__forward_is_infer(WaveGlow):
def forward(self, spect, sigma=1.0):
return self.infer(spect, sigma)
model = WaveGlow__forward_is_infer(**model_config)
else:
model = WaveGlow(**model_config)
else:
raise NotImplementedError(model_name)
if uniform_initialize_bn_weight:
init_bn(model)
if not cpu_run:
model = model.cuda()
return model
def __init__(self, ckpt_file, device='cuda', use_fp16=False, use_denoiser=False):
self.ckpt_file = ckpt_file
self.device = device
self.use_fp16 = use_fp16
self.use_denoiser = use_denoiser
# model
# sys.path.append('waveglow')
from waveglow.arg_parser import parse_waveglow_args
parser = parser = argparse.ArgumentParser()
model_parser= parse_waveglow_args(parser)
args, _ = model_parser.parse_known_args()
model_config = dict(
n_mel_channels=args.n_mel_channels,
n_flows=args.flows,
n_group=args.groups,
n_early_every=args.early_every,
n_early_size=args.early_size,
WN_config=dict(
n_layers=args.wn_layers,
kernel_size=args.wn_kernel_size,
n_channels=args.wn_channels
)
)
self.model = WaveGlow(**model_config)
state_dict = torch.load(self.ckpt_file, map_location=self.device)['state_dict']
state_dict = unwrap_distributed(state_dict)
self.model.load_state_dict(state_dict)
self.model = to_device_async(self.model, self.device)
self.model = self.model.remove_weightnorm(self.model)
self.model.eval()
if self.use_fp16:
self.model = self.model.half()
self.model = self.model
if self.use_denoiser:
self.denoiser = Denoiser(self.model, device=device)
self.denoiser = to_device_async(self.denoiser, self.device)
tprint('Using WaveGlow denoiser.')
def get_model(model_name, model_config, to_cuda,
uniform_initialize_bn_weight=False):
""" Code chooses a model based on name"""
model = None
if model_name == 'Tacotron2':
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
model = WaveGlow(**model_config)
else:
raise NotImplementedError(model_name)
if uniform_initialize_bn_weight:
init_bn(model)
if to_cuda:
model = model.cuda()
return model
def load_waveglow(filename, waveglow_config):
class RenamingUnpickler(pickle.Unpickler):
def find_class(self, module, name):
if module == 'glow':
module = 'waveglow.model'
return super().find_class(module, name)
class RenamingPickleModule:
def load(self, f, *args, **kw_args):
return self.Unpickler(f, *args, **kw_args).load()
def Unpickler(self, f, **pickle_load_args):
return RenamingUnpickler(f, **pickle_load_args)
pickle_module = RenamingPickleModule()
blob = torch.load(filename, pickle_module=pickle_module)
if 'state_dict' in blob:
waveglow = WaveGlow(**waveglow_config).cuda()
state_dict = {}
for key, value in blob["state_dict"].items():
newKey = key
if key.startswith("module."):
newKey = key[len("module."):]
state_dict[newKey] = value
waveglow.load_state_dict(state_dict)
else:
waveglow = blob['model']
waveglow = split_cond_layers(waveglow)
waveglow = waveglow.remove_weightnorm(waveglow)
waveglow.cuda().eval()
return waveglow
def get_model(model_name, model_config, to_fp16, to_cuda, training=True):
""" Code chooses a model based on name"""
model = None
if model_name == 'Tacotron2':
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
model = WaveGlow(**model_config)
else:
raise NotImplementedError(model_name)
if to_fp16:
model = batchnorm_to_float(model.half())
model = lstmcell_to_float(model)
if model_name == "WaveGlow":
for k in model.convinv:
k.float()
if to_cuda:
model = model.cuda()
return model
def get_model(model_name, model_config, device,
uniform_initialize_bn_weight=False, forward_is_infer=False,
jitable=False):
""" Code chooses a model based on name"""
model = None
if model_name == 'Tacotron2':
if forward_is_infer:
class Tacotron2__forward_is_infer(Tacotron2):
def forward(self, inputs, input_lengths):
return self.infer(inputs, input_lengths)
model = Tacotron2__forward_is_infer(**model_config)
else:
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
if forward_is_infer:
class WaveGlow__forward_is_infer(WaveGlow):
def forward(self, spect, sigma=1.0):
return self.infer(spect, sigma)
model = WaveGlow__forward_is_infer(**model_config)
else:
model = WaveGlow(**model_config)
elif model_name == 'FastPitch':
if forward_is_infer:
if jitable:
class FastPitch__forward_is_infer(_FastPitchJIT):
def forward(self, inputs, input_lengths, pace: float = 1.0,
dur_tgt: Optional[torch.Tensor] = None,
pitch_tgt: Optional[torch.Tensor] = None,
pitch_transform: Optional[bool] = None):
return self.infer(inputs, input_lengths, pace=pace,
dur_tgt=dur_tgt, pitch_tgt=pitch_tgt)
else:
class FastPitch__forward_is_infer(_FastPitch):
def forward(self, inputs, input_lengths, pace: float = 1.0,
dur_tgt: Optional[torch.Tensor] = None,
pitch_tgt: Optional[torch.Tensor] = None,
pitch_transform=None):
return self.infer(inputs, input_lengths, pace=pace,
dur_tgt=dur_tgt, pitch_tgt=pitch_tgt,
pitch_transform=pitch_transform)
model = FastPitch__forward_is_infer(**model_config)
else:
model = _FastPitch(**model_config)
else:
raise NotImplementedError(model_name)
if uniform_initialize_bn_weight:
init_bn(model)
return model.to(device)
def get_model(model_name: str,
model_config: dict,
cpu_run: bool,
uniform_initialize_bn_weight: bool = False,
forward_is_infer: bool = False):
"""Return a model based on `model_name` defined by `model_config`.
Args:
model_name (str): One of the 'Tacotron2' or 'WaveGlow'.
model_config (dict): [description]
cpu_run (bool): Run on CPU (True) or GPU (False).
uniform_initialize_bn_weight (bool, optional): [description]. Defaults to False.
forward_is_infer (bool, optional): [description]. Defaults to False.
Raises:
NotImplementedError: [description]
Returns:
[type]: [description]
"""
model = None
if model_name == 'Tacotron2':
if forward_is_infer:
class Tacotron2__forward_is_infer(Tacotron2):
def forward(self, inputs, input_lengths):
print(input_lengths) #FIXME
return self.infer(inputs, input_lengths)
model = Tacotron2__forward_is_infer(**model_config)
else:
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
if forward_is_infer:
class WaveGlow__forward_is_infer(WaveGlow):
def forward(self, spect, sigma=1.0):
return self.infer(spect, sigma)
model = WaveGlow__forward_is_infer(**model_config)
else:
model = WaveGlow(**model_config)
else:
raise NotImplementedError(model_name)
if uniform_initialize_bn_weight:
init_bn(model)
if cpu_run == False:
model = model.cuda()
return model
def get_model(model_name, model_config, to_cuda,
uniform_initialize_bn_weight=False, rename=False):
""" Code chooses a model based on name"""
model = None
if model_name == 'Tacotron2':
if rename:
class Tacotron2_extra(Tacotron2):
def forward(self, inputs, input_lengths):
return self.infer(inputs, input_lengths)
model = Tacotron2_extra(**model_config)
else:
model = Tacotron2(**model_config)
elif model_name == 'WaveGlow':
model = WaveGlow(**model_config)
else:
raise NotImplementedError(model_name)
if uniform_initialize_bn_weight:
init_bn(model)
if to_cuda:
model = model.cuda()
return model
def build_model(params):
upsample_net = UpsampleNet(
upsample_factor=params['upsample_net']['upsample_factor'],
upsample_method=params['upsample_net']['upsample_method'],
squeeze_factor=params['waveglow']['squeeze_factor'])
input_channels = params['upsample_net']['input_channels']
local_condition_channels = input_channels * params['waveglow']['squeeze_factor']
model = WaveGlow(
squeeze_factor=params['waveglow']['squeeze_factor'],
num_layers=params['waveglow']['num_layers'],
wn_filter_width=params['waveglow']['wn_filter_width'],
wn_dilation_layers=params['waveglow']['wn_dilation_layers'],
wn_residual_channels=params['waveglow']['wn_residual_channels'],
wn_dilation_channels=params['waveglow']['wn_dilation_channels'],
wn_skip_channels=params['waveglow']['wn_skip_channels'],
local_condition_channels=local_condition_channels)
return upsample_net, model
def get_model(model_name, model_config, device,
uniform_initialize_bn_weight=False, forward_is_infer=False,
jitable=False):
if model_name == 'WaveGlow':
model = WaveGlow(**model_config)
elif model_name == 'FastPitch':
if jitable:
model = FastPitchJIT(**model_config)
else:
model = FastPitch(**model_config)
else:
raise NotImplementedError(model_name)
if forward_is_infer:
model.forward = model.infer
if uniform_initialize_bn_weight:
init_bn(model)
return model.to(device)
class WaveGlowInferencer(object):
def __init__(self, ckpt_file, device='cuda', use_fp16=False, use_denoiser=False):
self.ckpt_file = ckpt_file
self.device = device
self.use_fp16 = use_fp16
self.use_denoiser = use_denoiser
# model
# sys.path.append('waveglow')
from waveglow.arg_parser import parse_waveglow_args
parser = parser = argparse.ArgumentParser()
model_parser= parse_waveglow_args(parser)
args, _ = model_parser.parse_known_args()
model_config = dict(
n_mel_channels=args.n_mel_channels,
n_flows=args.flows,
n_group=args.groups,
n_early_every=args.early_every,
n_early_size=args.early_size,
WN_config=dict(
n_layers=args.wn_layers,
kernel_size=args.wn_kernel_size,
n_channels=args.wn_channels
)
)
self.model = WaveGlow(**model_config)
state_dict = torch.load(self.ckpt_file, map_location=self.device)['state_dict']
state_dict = unwrap_distributed(state_dict)
self.model.load_state_dict(state_dict)
self.model = to_device_async(self.model, self.device)
self.model = self.model.remove_weightnorm(self.model)
self.model.eval()
if self.use_fp16:
self.model = self.model.half()
self.model = self.model
if self.use_denoiser:
self.denoiser = Denoiser(self.model, device=device)
self.denoiser = to_device_async(self.denoiser, self.device)
tprint('Using WaveGlow denoiser.')
def __enter__(self):
pass
def __exit__(self, exception_type, exception_value, traceback):
pass
def infer(self, mels):
if self.use_fp16:
mels = mels.half()
mels = to_device_async(mels, self.device)
wavs = self.model.infer(mels, sigma=0.6)
if self.use_denoiser:
wavs = self.denoiser(wavs, strength=0.01)
return wavs.float()