def load_data(args, pars):
# Data sets and data loaders
print('Data')
dataset_train = datain.DataSet(args.path_data,
pars.lchunk,
pars.stride,
split='train',
sampling_rate=pars.sr,
trim=pars.trim,
frame_energy_thres=pars.frame_energy_thres,
temp_jitter=pars.augment > 0,
seed=args.seed)
dataset_valid = datain.DataSet(args.path_data,
pars.lchunk,
pars.stride,
split='valid',
sampling_rate=pars.sr,
trim=pars.trim,
frame_energy_thres=pars.frame_energy_thres,
temp_jitter=False,
seed=args.seed)
dataset_test = datain.DataSet(args.path_data,
pars.lchunk,
pars.stride,
split='test',
sampling_rate=pars.sr,
trim=pars.trim,
frame_energy_thres=pars.frame_energy_thres,
temp_jitter=False,
seed=args.seed)
loader_train = torch.utils.data.DataLoader(dataset_train,
batch_size=args.sbatch,
shuffle=True,
drop_last=True,
num_workers=pars.nworkers)
loader_valid = torch.utils.data.DataLoader(dataset_valid,
batch_size=args.sbatch,
shuffle=False,
num_workers=pars.nworkers)
loader_test = torch.utils.data.DataLoader(dataset_test,
batch_size=1,
shuffle=False,
num_workers=pars.nworkers)
print('-' * 100)
return dataset_train, loader_train, dataset_valid, loader_valid, dataset_test, loader_test
def compute_speaker_averages(speakers, trim=5 * 60):
print('(' * 100)
print('[Averages]')
select_speaker = None
if args.force_source_speaker is not None and args.force_target_speaker is not None:
select_speaker = args.force_source_speaker + ',' + args.force_target_speaker
dataset = datain.DataSet(args.path_data,
args.lchunk,
args.lchunk,
sampling_rate=pars.sr,
split='train+valid',
trim=trim,
select_speaker=select_speaker,
seed=pars.seed)
loader = torch.utils.data.DataLoader(dataset,
batch_size=args.sbatch,
shuffle=False,
num_workers=0)
averages = {}
count = {}
with torch.no_grad():
for b, (x, idx) in enumerate(loader):
x = x.to(args.device)
s = idx[:, 3].to(args.device)
z = model.forward(x, s)[0]
for n in range(len(idx)):
i, j, last, ispk, ichap = idx[n]
spk, _ = dataset.filename_split(dataset.filenames[i])
spk = speakers[spk]
if spk not in averages:
averages[spk] = torch.zeros_like(z[n])
count[spk] = 0
averages[spk] += z[n]
count[spk] += 1
print('\r---> Speaker(s) average: {:5.1f}%'.format(
100 * (b * args.sbatch + x.size(0)) / len(dataset)),
end='')
print()
for spk in averages.keys():
averages[spk] = averages[spk] / count[spk]
print(')' * 100)
return averages
end='')
print()
for spk in averages.keys():
averages[spk] = averages[spk] / count[spk]
print(')' * 100)
return averages
########################################################################################################################
# Data
print('Load metadata')
dataset = datain.DataSet(args.path_data,
pars.lchunk,
pars.stride,
sampling_rate=pars.sr,
split='train+valid',
seed=pars.seed,
do_audio_load=False)
speakers = deepcopy(dataset.speakers)
lspeakers = list(speakers.keys())
if args.zavg:
averages = compute_speaker_averages(speakers)
# Input data
print('Load', args.split, 'audio')
dataset = datain.DataSet(args.path_data,
args.lchunk,
args.stride,
sampling_rate=pars.sr,
split=args.split,
)
window = torch.hann_window(args.lchunk)
args.synth_nonorm = False
window = window.view(1, -1)
print('-' * 100)
########################################################################################################################
# Data
print('Load VCTK_22kHz_adapt test split speakers')
adapt_data_path = os.path.join(args.path_data_root, 'VCTK_22kHz_10')
dataset = datain.DataSet(adapt_data_path,
pars.lchunk,
pars.stride,
sampling_rate=pars.sr,
split='test',
seed=pars.seed,
do_audio_load=False)
speakers = deepcopy(dataset.speakers)
lspeakers = list(speakers.keys())
print('Adapting to these speakers:')
print(lspeakers)
# Input data
print('Load VCTK_22kHz_train test split audio')
train_data_path = os.path.join(args.path_data_root, 'VCTK_22kHz_98')
dataset = datain.DataSet(train_data_path,
args.lchunk,
args.stride,
sampling_rate=pars.sr,
pars.load_existing=args.load_existing
pars.base_fn_out=args.base_fn_out
pars.multigpu=args.multigpu
pars.sbatch=args.sbatch
pars.optim=args.optim
pars.lr=args.lr
args=pars
print('New arguments')
utils.print_arguments(args)
########################################################################################################################
# Data sets and data loaders
print('Data')
dataset_train=datain.DataSet(args.path_data,args.lchunk,args.stride,split='train',sampling_rate=args.sr,
trim=args.trim,frame_energy_thres=args.frame_energy_thres,
temp_jitter=args.augment>0,
seed=args.seed)
dataset_valid=datain.DataSet(args.path_data,args.lchunk,args.stride,split='valid',sampling_rate=args.sr,
trim=args.trim,frame_energy_thres=args.frame_energy_thres,
temp_jitter=False,
seed=args.seed)
dataset_test=datain.DataSet(args.path_data,args.lchunk,args.stride,split='test',sampling_rate=args.sr,
trim=args.trim,frame_energy_thres=args.frame_energy_thres,
temp_jitter=False,
seed=args.seed)
loader_train=torch.utils.data.DataLoader(dataset_train,batch_size=args.sbatch,shuffle=True,drop_last=True,num_workers=args.nworkers)
loader_valid=torch.utils.data.DataLoader(dataset_valid,batch_size=args.sbatch,shuffle=False,num_workers=args.nworkers)
loader_test=torch.utils.data.DataLoader(dataset_test,batch_size=args.sbatch,shuffle=False,num_workers=args.nworkers)
print('-'*100)
########################################################################################################################
utils.print_arguments(args)
# Seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.device == 'cuda':
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.cuda.manual_seed(args.seed)
# Data sets and data loaders
print('Data')
dataset_train = datain.DataSet(args.path_data, args.lchunk, args.stride,
# select_speaker='p285,p303',
# split = 'train' or 'train-overfit'
split=args.train_split, sampling_rate=args.sr,
trim=args.trim,
frame_energy_thres=args.frame_energy_thres,
temp_jitter=args.augment > 0,
seed=args.seed)
dataset_valid = datain.DataSet(args.path_data, args.lchunk, args.stride,
# select_speaker='p285,p303',
split='valid', sampling_rate=args.sr,
trim=args.trim,
frame_energy_thres=args.frame_energy_thres,
temp_jitter=False,
seed=args.seed)
dataset_test = datain.DataSet(args.path_data, args.lchunk, args.stride,
# select_speaker='p285,p303',
split='test', sampling_rate=args.sr,
trim=args.trim,
frame_energy_thres=args.frame_energy_thres,