is gaussian_blur_1d correct?

[thuliu-yt16]

I think we should view the discrete coordinate dimension as the "spatial" dimension in conv1d. Replacing

def gaussian_blur_1d(
    t: Tensor,
    *,
    sigma: float = 1.
) -> Tensor:

    _, channels, _, device, dtype = *t.shape, t.device, t.dtype

    width = int(ceil(sigma * 5))
    width += (width + 1) % 2
    half_width = width // 2

    distance = torch.arange(-half_width, half_width + 1, dtype = dtype, device = device)

    gaussian = torch.exp(-(distance ** 2) / (2 * sigma ** 2))
    gaussian = l1norm(gaussian)

    kernel = repeat(gaussian, 'n -> c 1 n', c = channels)
    return F.conv1d(t, kernel, padding = half_width, groups = channels)

with

def gaussian_blur_1d(
    t: Tensor,
    *,
    sigma: float = 1.
) -> Tensor:

    _, _, channels, device, dtype = *t.shape, t.device, t.dtype # change made here

    width = int(ceil(sigma * 5))
    width += (width + 1) % 2
    half_width = width // 2

    distance = torch.arange(-half_width, half_width + 1, dtype = dtype, device = device)

    gaussian = torch.exp(-(distance ** 2) / (2 * sigma ** 2))
    gaussian = l1norm(gaussian)

    kernel = repeat(gaussian, 'n -> c 1 n', c = channels)
    return F.conv1d(t.permute(0, 2, 1), kernel, padding = half_width, groups = channels).permute(0, 2, 1) # change made here

makes the blurring work.

I made a breakpoint to check the value of target_one_hot (ignore the peak location change). Before:

(Pdb) target_one_hot[0, :, 1]
tensor([0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0404,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0404, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.9192, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000], device='cuda:0')

After:

(Pdb) target_one_hot[0, :, 1]
tensor([0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0404, 0.9192, 0.0404, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000,
        0.0000, 0.0000], device='cuda:0')

[lucidrains]

🤦 indeed i got mixed up with what was the spatial dimension here

thank you!