every mp3 is clip to the same length
(hp.training_files: ../clips.5s/ )
'''
vocal_dataset = AudioNpyLoader(hp.vocal_files)
linear_mixture_dataset = AudioNpyLoader(hp.linear_mixture_files)
accom_dataset = AudioNpyLoader(hp.accom_files)
vocal_iterator_tr = DataLoader(vocal_dataset,
batch_size=4,
num_workers=2,
shuffle=True,
drop_last=True,
pin_memory=True,
collate_fn=AudioCollate())
inf_iterator_voc_speech = make_inf_iterator(vocal_iterator_tr)
linear_iterator_tr = DataLoader(linear_mixture_dataset,
batch_size=4,
num_workers=2,
shuffle=True,
drop_last=True,
pin_memory=True,
collate_fn=AudioCollate())
inf_iterator_lin_speech = make_inf_iterator(linear_iterator_tr)
accom_iterator_tr = DataLoader(accom_dataset,
batch_size=4,
num_workers=2,
shuffle=True,
.
every mp3 is clip to the same length
(hp.training_files: ../clips.5s/ )
'''
speech_dataset = AudioNpyLoader(hp.speech_files)
singing_dataset = AudioNpyLoader(hp.singing_files)
sp_iterator_tr = DataLoader(
speech_dataset,
batch_size=4,
num_workers=2,
shuffle=True,
drop_last=True,
pin_memory=True, collate_fn = AudioCollate())
inf_iterator_tr_speech = make_inf_iterator(sp_iterator_tr)
si_iterator_tr = DataLoader(
singing_dataset,
batch_size=4,
num_workers=2,
shuffle=True,
drop_last=True,
pin_memory=True, collate_fn = AudioCollate())
inf_iterator_tr_sing = make_inf_iterator(si_iterator_tr)
##################################################################
# BEGAN parameters
if hp.loss == "BEGAN":
gamma = 1.0
data_dir = f'/home/ericwudayi/nas189/homes/kevinco27/dataset/LJSpeech-1.1/clips.5s.mel'
mean_fp = os.path.join(data_dir, f'mean.mel.melgan.npy')
std_fp = os.path.join(data_dir, f'std.mel.melgan.npy')
mean = torch.from_numpy(np.load(mean_fp)).float().cuda().view(1, 80, 1)
std = torch.from_numpy(np.load(std_fp)).float().cuda().view(1, 80, 1)
iterator_tr = DataLoader(dataset,
batch_size=4,
num_workers=2,
shuffle=True,
drop_last=True,
pin_memory=True)
inf_iterator_tr_speech = make_inf_iterator(iterator_tr)
##################################################################
# BEGAN parameters
if hp.loss == "BEGAN":
gamma = 1.0
lambda_k = 0.01
init_k = 0.0
recorder = BEGANRecorder(lambda_k, init_k, gamma)
k = recorder.k.item()
###################################################################
"""
Model Architecture from General Model
All Model is wrap by GeneralModel(nn.modules), but you can ignore