def _generate_conditions(self,
batch,
cv_spkr_name=None,
use_cvfeats=False,
encoder=False):
# create lcf0, uv, h_onehot
if cv_spkr_name is not None:
spkr_num = self.spkrs[cv_spkr_name]
B, T, _ = batch["feats"].size()
lcf0 = torch.tensor(self._get_cvf0(batch,
cv_spkr_name)).to(self.device)
uv = batch["uv"]
h_onehot = torch.tensor(
create_one_hot(T, self.n_spkrs, spkr_num, B=B)).to(self.device)
else:
if use_cvfeats:
lcf0, uv, h_onehot = batch["cv_lcf0"], batch["uv"], batch[
"cv_h_onehot"]
else:
lcf0, uv, h_onehot = batch["lcf0"], batch["uv"], batch[
"org_h_onehot"]
# return conditions
if encoder:
if self.conf["encoder_f0"]:
return torch.cat([lcf0, uv], dim=-1).to(self.device)
else:
return None
else:
if self.conf["decoder_f0"]:
return torch.cat([lcf0, uv, h_onehot], dim=-1).to(self.device)
else:
return h_onehot.to(self.device)
def _get_spkr_conditions(self, batch, cv_spkr_name, use_cvfeats=False):
if cv_spkr_name is not None:
# use specified cv speaker
B, T, _ = batch["in_feats"].size()
spkr_num = self.spkrs[cv_spkr_name]
h_onehot_np = create_one_hot(T, self.n_spkrs, spkr_num, B=B)
h_onehot = torch.tensor(h_onehot_np).to(self.device)
h = (torch.ones(
(B, T)).long() * self.spkrs[cv_spkr_name]).to(self.device)
else:
if use_cvfeats:
# use randomly selected cv speaker by dataset
h = batch["cv_h"].clone()
h_onehot = batch["cv_h_onehot"]
else:
# use org speaker
h_onehot = batch["org_h_onehot"]
h = batch["org_h"].clone()
h[:, :] = h[:, 0:1] # remove ignore_index (i.e., -100)
return h, h_onehot
def _prepare_feats(self, batch, cv_spkr_name, use_cvfeats=False):
if cv_spkr_name is not None:
# use specified cv speaker
B, T, _ = batch["feats"].size()
spkr_num = self.spkrs[cv_spkr_name]
lcf0 = self._get_cvf0(batch, cv_spkr_name)
h_onehot_np = create_one_hot(T, self.n_spkrs, spkr_num, B=B)
h_onehot = torch.tensor(h_onehot_np).to(self.device)
h_scalar = torch.ones((B, T)).long() * self.spkrs[cv_spkr_name]
h_scalar = h_scalar.to(self.device)
else:
if use_cvfeats:
# use randomly selected cv speaker by dataset
lcf0 = batch["cv_lcf0"]
h_onehot = batch["cv_h_onehot"]
h_scalar = batch["cv_h_scalar"]
else:
# use org speaker
lcf0 = batch["lcf0"]
h_onehot = batch["org_h_onehot"]
h_scalar = batch["org_h_scalar"]
h_scalar[:, :] = h_scalar[:, 0:1] # remove ignore_index (i.e., -100)
return torch.cat([lcf0, batch["uv"]], axis=-1), h_onehot, h_scalar