if not os.path.exists(args.check_path):
os.makedirs(args.check_path)
opt_kwargs = {
'lr': args.lr,
'lr_decay': args.lr_decay,
'weight_decay': args.weight_decay,
'dampening': args.dampening,
'momentum': args.momentum
}
l2_dist = nn.CosineSimilarity(
dim=1, eps=1e-12) if args.cos_sim else nn.PairwiseDistance(p=2)
if args.acoustic_feature == 'fbank':
transform = transforms.Compose([totensor()])
transform_T = transforms.Compose([
concateinputfromMFB(num_frames=c.NUM_FRAMES_SPECT,
input_per_file=args.test_input_per_file,
remove_vad=args.remove_vad),
])
transform_V = transforms.Compose([
varLengthFeat(remove_vad=args.remove_vad),
])
else:
transform = transforms.Compose([
truncatedinput(),
toMFB(),
totensor(),
# tonormal()
else:
l2_dist = PairwiseDistance(2)
voxceleb, voxceleb_dev = wav_list_reader(args.dataroot)
if args.makemfb:
# pbar = tqdm(voxceleb)
for datum in voxceleb:
mk_MFB(
(args.dataroot + '/voxceleb1_wav/' + datum['filename'] + '.wav'))
print("Complete convert")
if args.mfb:
transform = transforms.Compose([
concateinputfromMFB(),
# truncatedinputfromMFB(),
totensor()
])
transform_T = transforms.Compose([
concateinputfromMFB(input_per_file=args.test_input_per_file),
# truncatedinputfromMFB(input_per_file=args.test_input_per_file),
totensor()
])
file_loader = read_MFB
else:
transform = transforms.Compose([
truncatedinput(),
toMFB(),
totensor(),
# tonormal()
])
file_loader = read_audio
# create logger
logger = Logger(LOG_DIR)
kwargs = {'num_workers': 0, 'pin_memory': True} if args.cuda else {}
l2_dist = PairwiseDistance(2)
# voxceleb = read_my_voxceleb_structure(args.dataroot)
# if args.makemfb:
#pbar = tqdm(voxceleb)
# for datum in voxceleb:
# # print(datum['filename'])
# mk_MFB((args.dataroot +'/voxceleb1_wav/' + datum['filename']+'.wav'))
# print("Complete convert")
if args.mfb:
transform = transforms.Compose([truncatedinputfromMFB(), totensor()])
transform_T = transforms.Compose([
truncatedinputfromMFB(input_per_file=args.test_input_per_file),
totensor()
])
file_loader = read_MFB
else:
transform = transforms.Compose([
truncatedinput(),
toMFB(),
totensor(),
#tonormal()
])
file_loader = read_audio
# voxceleb_dev = [datum for datum in voxceleb if datum['subset']=='dev']
# print("Complete convert")
#
# if args.makespec:
# num_pro = 1.
# for datum in voxceleb:
# # Data/voxceleb1/
# # /data/voxceleb/voxceleb1_wav/
# GenerateSpect(wav_path='/data/voxceleb/voxceleb1_wav/' + datum['filename']+'.wav',
# write_path=args.dataroot +'/spectrogram/voxceleb1_wav/' + datum['filename']+'.npy')
# print('\rprocessed {:2f}% {}/{}.'.format(num_pro/len(voxceleb), num_pro, len(voxceleb)), end='\r')
# num_pro += 1
# print('\nComputing Spectrograms success!')
# exit(1)
if args.acoustic_feature == 'fbank':
transform = transforms.Compose([concateinputfromMFB(), totensor()])
transform_T = transforms.Compose([
concateinputfromMFB(input_per_file=args.test_input_per_file),
totensor()
])
file_loader = read_MFB
else:
transform = transforms.Compose([
truncatedinput(),
toMFB(),
totensor(),
# tonormal()
])
file_loader = read_audio
# concateinputfromMFB(num_frames=c.NUM_FRAMES_SPECT, remove_vad=True),
varLengthFeat(remove_vad=args.remove_vad),
to2tensor(),
tonormal()
])
transform_T = transforms.Compose([
# concateinputfromMFB(num_frames=c.NUM_FRAMES_SPECT, input_per_file=args.test_input_per_file, remove_vad=True),
varLengthFeat(remove_vad=args.remove_vad),
to2tensor(),
tonormal()
])
else:
transform = transforms.Compose(
[truncatedinput(), toMFB(),
totensor(), tonormal()])
file_loader = read_audio
# pdb.set_trace()
file_loader = read_mat
train_dir = ScriptTrainDataset(dir=args.train_dir,
samples_per_speaker=args.input_per_spks,
loader=file_loader,
transform=transform,
num_valid=args.num_valid)
test_dir = ScriptTestDataset(dir=args.test_dir,
loader=file_loader,
transform=transform_T)
if len(test_dir) < args.veri_pairs:
args.veri_pairs = len(test_dir)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.cuda:
cudnn.benchmark = True
# create logger
# Define visulaize SummaryWriter instance
kwargs = {'num_workers': 12, 'pin_memory': True} if args.cuda else {}
l2_dist = nn.CosineSimilarity(
dim=1, eps=1e-6) if args.cos_sim else PairwiseDistance(2)
if args.acoustic_feature == 'fbank':
transform = transforms.Compose([varLengthFeat(), totensor()])
transform_T = transforms.Compose([varLengthFeat(), totensor()])
file_loader = read_mat
else:
transform = transforms.Compose([
truncatedinput(),
toMFB(),
totensor(),
# tonormal()
])
file_loader = read_audio
# pdb.set_trace()
train_dir = KaldiExtractDataset(dir=args.train_dir,
loader=file_loader,
transform=transform)
'lr_decay': args.lr_decay,
'weight_decay': args.weight_decay,
'dampening': args.dampening,
'momentum': args.momentum
}
l2_dist = nn.CosineSimilarity(
dim=1, eps=1e-12) if args.cos_sim else nn.PairwiseDistance(p=2)
if args.acoustic_feature == 'fbank':
transform = transforms.Compose([
ConcateNumInput(input_per_file=args.input_per_spks,
num_frames=args.chunk_size,
feat_type=args.feat_format,
remove_vad=args.remove_vad),
totensor()
])
if args.test_input == 'var':
transform_V = transforms.Compose([
ConcateOrgInput(remove_vad=args.remove_vad,
feat_type=args.feat_format),
])
elif args.test_input == 'fix':
transform_V = transforms.Compose([
ConcateVarInput(remove_vad=args.remove_vad,
num_frames=args.chunk_size,
frame_shift=args.chunk_size,
feat_type=args.feat_format),
])
