def __init__(self, hparams, n_units=128):
super(MultiHeadAttention, self).__init__()
self.token_embedding_size = hparams.token_embedding_size
self.num_heads = hparams.num_heads
self.token_num = hparams.token_num
self.n_units = hparams.gstAtt_dim
self.split_size = n_units // self.num_heads
self.conv_Q = mm.Conv1d(self.token_embedding_size, n_units,
1) # in_channels, out_channels, kernel_size
self.conv_K = mm.Conv1d(self.token_embedding_size, n_units,
1) # in_channels, out_channels, kernel_size
self.fc_Q = nn.Sequential(
nn.Linear(n_units, n_units),
nn.Tanh(),
)
self.fc_K = nn.Sequential(
nn.Linear(n_units, n_units),
nn.Tanh(),
)
self.fc_V = nn.Sequential(
nn.Linear(self.token_embedding_size, self.split_size),
nn.Tanh(),
)
self.fc_A = nn.Sequential(
nn.Linear(n_units, self.token_num),
nn.Tanh(),
)
def __init__(self,
input_dim,
hidden_dim,
K=16,
n_highway=4,
bidirectional=True):
super(CBHG, self).__init__()
self.K = K
self.conv_bank = mm.Conv1dBank(input_dim,
hidden_dim,
K=self.K,
activation_fn=torch.relu)
self.max_pool = nn.MaxPool1d(2, stride=1, padding=1)
self.projection = nn.Sequential(
mm.Conv1d(self.K * hidden_dim,
hidden_dim,
3,
activation_fn=torch.relu,
bias=False,
bn=True),
mm.Conv1d(hidden_dim, input_dim, 3, bias=False, bn=True),
)
self.highway = nn.ModuleList(
[mm.Highway(input_dim) for _ in range(n_highway)])
self.gru = nn.GRU(
input_dim,
hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=bidirectional
) # if batch_first is True, (Batch, Sequence, Feature)
def __init__(self, mel_channels=80, z_dim=256): super(Encoder, self).__init__() self.encoder = nn.Sequential( mm.Conv1d(mel_channels, 64, kernel_size=4, stride=2, padding='same', bias=True, activation_fn=nn.ReLU), mm.Conv1d(64, 256, kernel_size=3, padding='same', bias=True), mm.Conv1d(256, 128, kernel_size=3, padding='same', bias=True), mm.Conv1dResBlock(128, 128, kernel_size=3, padding='same', bias=True, activation_fn=nn.ReLU), mm.Conv1dResBlock(128, 128, kernel_size=3, padding='same', bias=True, activation_fn=nn.ReLU), mm.Conv1d(128, z_dim, kernel_size=1, padding='same', bias=False) )
def __init__(self, in_channels, mel_channels=80): super(Decoder, self).__init__() self.res_blocks = nn.Sequential( mm.Conv1d(in_channels, 128, kernel_size=3, bias=False, padding='same'), mm.Conv1dResBlock(128, 128, kernel_size=3, bias=True, padding='same', activation_fn=nn.ReLU), mm.Conv1dResBlock(128, 128, kernel_size=3, bias=True, padding='same', activation_fn=nn.ReLU), mm.Upsample(scale_factor=2, mode='nearest'), mm.Conv1d(128, 256, kernel_size=2, bias=True, padding='same', activation_fn=nn.ReLU), mm.Linear(256, mel_channels) )
def __init__(self):
super(TextEncoder, self).__init__()
self.hc_blocks = nn.ModuleList([
norm(
mm.Conv1d(args.Ce,
args.Cx * 2,
1,
padding='same',
activation_fn=torch.relu))
]) # filter up to split into K, V
self.hc_blocks.extend([
norm(
mm.Conv1d(args.Cx * 2,
args.Cx * 2,
1,
padding='same',
activation_fn=None))
])
self.hc_blocks.extend([
norm(
mm.HighwayConv1d(args.Cx * 2,
args.Cx * 2,
3,
dilation=3**i,
padding='same')) for _ in range(2)
for i in range(4)
])
self.hc_blocks.extend([
norm(
mm.HighwayConv1d(args.Cx * 2,
args.Cx * 2,
3,
dilation=1,
padding='same')) for i in range(2)
])
self.hc_blocks.extend([
norm(
mm.HighwayConv1d(args.Cx * 2,
args.Cx * 2,
1,
dilation=1,
padding='same')) for i in range(2)
])
def __init__(self):
super(SSRN, self).__init__()
self.name = 'SSRN'
# (N, n_mels, Ty/r) -> (N, Cs, Ty/r)
self.hc_blocks = nn.ModuleList([
norm(mm.Conv1d(args.n_mels, args.Cs, 1, activation_fn=torch.relu))
])
self.hc_blocks.extend([
norm(mm.HighwayConv1d(args.Cs, args.Cs, 3, dilation=3**i))
for i in range(2)
])
# (N, Cs, Ty/r*2) -> (N, Cs, Ty/r*2)
self.hc_blocks.extend([
norm(mm.ConvTranspose1d(args.Cs, args.Cs, 4, stride=2, padding=1))
])
self.hc_blocks.extend([
norm(mm.HighwayConv1d(args.Cs, args.Cs, 3, dilation=3**i))
for i in range(2)
])
# (N, Cs, Ty/r*2) -> (N, Cs, Ty/r*4==Ty)
self.hc_blocks.extend([
norm(mm.ConvTranspose1d(args.Cs, args.Cs, 4, stride=2, padding=1))
])
self.hc_blocks.extend([
norm(mm.HighwayConv1d(args.Cs, args.Cs, 3, dilation=3**i))
for i in range(2)
])
# (N, Cs, Ty) -> (N, Cs*2, Ty)
self.hc_blocks.extend([norm(mm.Conv1d(args.Cs, args.Cs * 2, 1))])
self.hc_blocks.extend([
norm(mm.HighwayConv1d(args.Cs * 2, args.Cs * 2, 3, dilation=1))
for i in range(2)
])
# (N, Cs*2, Ty) -> (N, n_mags, Ty)
self.hc_blocks.extend([norm(mm.Conv1d(args.Cs * 2, args.n_mags, 1))])
self.hc_blocks.extend([
norm(
mm.Conv1d(args.n_mags,
args.n_mags,
1,
activation_fn=torch.relu)) for i in range(2)
])
self.hc_blocks.extend([norm(mm.Conv1d(args.n_mags, args.n_mags, 1))])
def __init__(self, n_units=128, embed_size=128):
super(MultiHeadAttention, self).__init__()
self.split_size = n_units // args.n_heads
self.conv_Q = mm.Conv1d(embed_size, n_units, 1)
self.conv_K = mm.Conv1d(embed_size, n_units, 1)
self.fc_Q = nn.Sequential(
nn.Linear(n_units, n_units),
nn.Tanh(),
)
self.fc_K = nn.Sequential(
nn.Linear(n_units, n_units),
nn.Tanh(),
)
self.fc_V = nn.Sequential(
nn.Linear(embed_size, self.split_size),
nn.Tanh(),
)
self.fc_A = nn.Sequential(
nn.Linear(n_units, args.n_tokens),
nn.Tanh(),
)
def __init__(self, text_dims, style_dims):
super(TPSENet, self).__init__()
self.conv = nn.Sequential(
mm.Conv1d(text_dims,
style_dims,
3,
activation_fn=torch.relu,
bn=True,
bias=False),
# mm.Conv1d(style_dims, style_dims, 3, activation_fn=torch.relu, bn=True, bias=False)
)
self.gru = nn.GRU(style_dims,
style_dims,
batch_first=True,
bidirectional=True)
self.fc = nn.Linear(style_dims * 2, style_dims)