Mamba Diffusion (IADB)

Mamba_Diffusion_IADB_Colab.ipynb

Code with the fixes

class Denoiser(nn.Module):
    def __init__(self, n_io=Config.channels, n_f=128, n_b=8):
        super().__init__()
        assert n_b % 8 == 0, f"Silly flipping logic breaks if n_b is not divisible by 8"
        self.enc = nn.Sequential(nn.Conv2d(n_io + 1, n_f, 1), nn.ReLU(), nn.Conv2d(n_f, n_f, 1, bias=False), nn.PixelUnshuffle(2))
        self.mid = nn.ModuleList(Block(n_f * 4, Mamba) for _ in range(n_b))
        self.dec = nn.Sequential(nn.Conv2d(n_f * 12, n_f * 4, 1), nn.ReLU(), nn.PixelShuffle(2), nn.Conv2d(n_f, n_io, 1))
        
    def transpose_xy(self, *args):
        # swap x/y axes of an N[XY]C tensor
        return [a.view(a.shape[0], int(a.shape[1]**0.5), int(a.shape[1]**0.5), a.shape[2]).transpose(1, 2).reshape(a.shape) for a in args]
    
    def flip_s(self, *args):
        # reverse sequence axis of an NSE tensor
        return [a.flip(1) for a in args]

    def forward(self, x_noisy, noise_level):
        x = self.enc(th.cat([x_noisy, noise_level.expand(x_noisy[:, :1].shape)], 1))
        y = x.flatten(2).transpose(-2, -1)
        z = None
        for i, mid in enumerate(self.mid):
            y, z = mid(y, z)
            # make mamba's 1d conv alternate axes (possible alternative: make mamba use a 2d conv...somehow...)
            y, z = self.transpose_xy(y, z)
            if (i + 1) % 4 == 0:
                # let the network also process the sequence of both directions, by reversing every 4 layers
                y, z = self.flip_s(y, z)
        y, z = y.transpose(-2, -1).view(x.shape), z.transpose(-2, -1).view(x.shape)
        out = self.dec(th.cat([x, y, z], 1))
        return Prediction(IADB.target_to_denoised(out, x_noisy, noise_level).detach(), out)

Samples after more training