Method __init__

usr/diff/candidate_decoder.py:36–48

(self, hidden_size=None, num_layers=None, kernel_size=None, num_heads=None)

class FFT(FastspeechDecoder):
    def __init__(self, hidden_size=None, num_layers=None, kernel_size=None, num_heads=None):
        super().__init__(hidden_size, num_layers, kernel_size, num_heads=num_heads)
        dim = hparams['residual_channels']
        self.input_projection = Conv1d(hparams['audio_num_mel_bins'], dim, 1)
        self.diffusion_embedding = SinusoidalPosEmb(dim)
        self.mlp = nn.Sequential(
            nn.Linear(dim, dim * 4),
            Mish(),
            nn.Linear(dim * 4, dim)
        )
        self.get_mel_out = Linear(hparams['hidden_size'], 80, bias=True)
        self.get_decode_inp = Linear(hparams['hidden_size'] + dim + dim,
                                     hparams['hidden_size'])  # hs + dim + 80 -> hs

    def forward(self, spec, diffusion_step, cond, padding_mask=None, attn_mask=None, return_hiddens=False):
        """