def __init__(self, num_chars, hidden_channels, hidden_channels_ffn,
num_heads):
super().__init__()
self.embed = nn.Embedding(num_chars, hidden_channels)
self.pos_enc = PositionalEncoding(hidden_channels, dropout_p=0.1)
self.FFT = FFTransformerBlock(hidden_channels, num_heads,
hidden_channels_ffn, 2, 0.1)
self.out_layer = nn.Conv1d(hidden_channels, 1, 1)
def __init__(
self,
in_hidden_channels,
out_channels,
encoder_type="residual_conv_bn",
encoder_params={
"kernel_size": 4,
"dilations": 4 * [1, 2, 4] + [1],
"num_conv_blocks": 2,
"num_res_blocks": 13
},
c_in_channels=0,
):
super().__init__()
self.out_channels = out_channels
self.in_channels = in_hidden_channels
self.hidden_channels = in_hidden_channels
self.encoder_type = encoder_type
self.c_in_channels = c_in_channels
# init encoder
if encoder_type.lower() == "relative_position_transformer":
# text encoder
# pylint: disable=unexpected-keyword-arg
self.encoder = RelativePositionTransformerEncoder(
in_hidden_channels, out_channels, in_hidden_channels,
encoder_params)
elif encoder_type.lower() == "residual_conv_bn":
self.encoder = ResidualConv1dBNEncoder(in_hidden_channels,
out_channels,
in_hidden_channels,
encoder_params)
elif encoder_type.lower() == "fftransformer":
assert (
in_hidden_channels == out_channels
), "[!] must be `in_channels` == `out_channels` when encoder type is 'fftransformer'"
# pylint: disable=unexpected-keyword-arg
self.encoder = FFTransformerBlock(in_hidden_channels,
**encoder_params)
else:
raise NotImplementedError(" [!] unknown encoder type.")
def __init__(self, in_channels, out_channels, params):
super().__init__()
self.transformer_block = FFTransformerBlock(in_channels, **params)
self.postnet = nn.Conv1d(in_channels, out_channels, 1)