Incorporate residual F0 prediction into acoustic model about nnsvs HOT 5 CLOSED

class ResConv1dResnet(BaseModel):
    def __init__(
        self,
        in_dim, hidden_dim, out_dim, num_layers=4,
        in_lf0_idx=300,
        in_lf0_min=5.3936276,
        in_lf0_max=6.491111,
        out_lf0_idx=180,
        out_lf0_mean=5.953093881972361,
        out_lf0_scale=0.23435173188961034,
    ):
        super().__init__()
        self.in_lf0_idx = in_lf0_idx
        self.in_lf0_min = in_lf0_min
        self.in_lf0_max = in_lf0_max
        self.out_lf0_idx = out_lf0_idx
        self.out_lf0_mean = out_lf0_mean
        self.out_lf0_scale = out_lf0_scale

        model = [
            nn.ReflectionPad1d(3),
            WNConv1d(in_dim, hidden_dim, kernel_size=7, padding=0),
        ]
        for n in range(num_layers):
            model.append(ResnetBlock(hidden_dim, dilation=2 ** n))
        model += [
            nn.LeakyReLU(0.2),
            nn.ReflectionPad1d(3),
            WNConv1d(hidden_dim, out_dim, kernel_size=7, padding=0),
        ]
        self.model = nn.Sequential(*model)

    def forward(self, x, lengths=None):
        out = self.model(x.transpose(1, 2)).transpose(1, 2)

        # denormalized lf0 from the input musical score
        lf0_score = x[:, :, self.in_lf0_idx].unsqueeze(-1)
        lf0_score_denorm = lf0_score * (self.in_lf0_max - self.in_lf0_min) + self.in_lf0_min


        # TODO: must be careful about dynamic features
        # Residual connection in denormalized f0
        lf0_res = out[:, :, self.out_lf0_idx].unsqueeze(-1)
        lf0_res = 0.693 * torch.tanh(lf0_res)
        lf0_pred_denorm = lf0_res + lf0_score_denorm
        # Back to normalized f0
        lf0_pred = (lf0_pred_denorm - self.out_lf0_mean) / self.out_lf0_scale

        out[:, :, self.out_lf0_idx] = lf0_pred.squeeze(-1)

        return out