def prepare_dataloaders(hparams):
# Get data, data loaders and collate function ready
#pids = [id2sp[hparams.speaker_A], id2sp[hparams.speaker_B]]
trainset = TextMelIDLoader(hparams.training_list,
hparams.mel_mean_std,
hparams.speaker_A,
hparams.speaker_B,
pids=None)
valset = TextMelIDLoader(hparams.validation_list,
hparams.mel_mean_std,
hparams.speaker_A,
hparams.speaker_B,
pids=None)
collate_fn = TextMelIDCollate(
lcm(hparams.n_frames_per_step_encoder,
hparams.n_frames_per_step_decoder))
train_sampler = DistributedSampler(trainset) \
if hparams.distributed_run else None
train_loader = DataLoader(trainset,
num_workers=1,
shuffle=True,
sampler=train_sampler,
batch_size=hparams.batch_size,
pin_memory=False,
drop_last=True,
collate_fn=collate_fn)
return train_loader, valset, collate_fn
def gen_embedding(speaker):
training_list = hparams.training_list
train_set_A = TextMelIDLoader(training_list,
hparams.mel_mean_std,
hparams.speaker_A,
hparams.speaker_B,
shuffle=False,
pids=[speaker])
collate_fn = TextMelIDCollate(
lcm(hparams.n_frames_per_step_encoder,
hparams.n_frames_per_step_decoder))
train_loader_A = DataLoader(train_set_A,
num_workers=1,
shuffle=False,
sampler=None,
batch_size=1,
pin_memory=False,
drop_last=True,
collate_fn=collate_fn)
with torch.no_grad():
speaker_embeddings = []
for i, batch in enumerate(train_loader_A):
#print i
x, y = model.parse_batch(batch)
text_input_padded, mel_padded, text_lengths, mel_lengths, speaker_id = x
speaker_id, speaker_embedding = model.speaker_encoder.inference(
mel_padded)
speaker_embedding = speaker_embedding.data.cpu().numpy()
#speaker_embedding = speaker_embedding * 0.1
speaker_embeddings.append(speaker_embedding)
speaker_embeddings = np.vstack(speaker_embeddings)
print(speaker_embeddings.shape)
if not os.path.exists('outdir/embeddings'):
os.makedirs('outdir/embeddings')
np.save('outdir/embeddings/%s.npy' % speaker, speaker_embeddings)
plot_data([speaker_embeddings], 'outdir/embeddings/%s.pdf' % speaker)
hparams.speaker_A,
hparams.speaker_B,
shuffle=False,
pids=[hparams.speaker_A])
test_set_B = TextMelIDLoader(test_list,
hparams.mel_mean_std,
hparams.speaker_A,
hparams.speaker_B,
shuffle=False,
pids=[hparams.speaker_B])
sample_list_A = test_set_A.file_path_list
sample_list_B = test_set_B.file_path_list
collate_fn = TextMelIDCollate(
lcm(hparams.n_frames_per_step_encoder, hparams.n_frames_per_step_decoder))
test_loader_A = DataLoader(test_set_A,
num_workers=1,
shuffle=False,
sampler=None,
batch_size=1,
pin_memory=False,
drop_last=False,
collate_fn=collate_fn)
test_loader_B = DataLoader(test_set_B,
num_workers=1,
shuffle=False,
sampler=None,
batch_size=1,