Method __init__

architecture/autoencoder_kl_wan.py:38–53

(
        self,
        in_channels,
        out_channels,
        factor_t,
        factor_s=1,
    )

class AvgDown3D(nn.Module):
    def __init__(
        self,
        in_channels,
        out_channels,
        factor_t,
        factor_s=1,
    ):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.factor_t = factor_t
        self.factor_s = factor_s
        self.factor = self.factor_t * self.factor_s * self.factor_s

        assert in_channels * self.factor % out_channels == 0
        self.group_size = in_channels * self.factor // out_channels

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t