1D Gaussian Kernel

MUnique · September 6, 2022, 10:53pm

Hello,

Is there any equivalent for “scipy.ndimage.gaussian_filter1d” in Pytorch? I have a tensor of shape [T, H, W] that I would like to apply a 1d gaussian kernel on its first dimension (T). How can I do so?

Thanks

ptrblck · September 6, 2022, 11:17pm

You could create a gaussian kernel as described e.g. in this post or this one and apply it to your input via F.conv2d.

tom · September 7, 2022, 6:45am

To add to @ptrblck’s comments, here is what I like to do for images:

github.com

TorchDrift/TorchDrift/blob/3d590e8f25185e68d2749d16372d5c4715e543f5/torchdrift/data/functional/corruption_functions.py#L114-L123


      
          distance = torch.arange(
              -(width // 2), width // 2 + 1, dtype=torch.float, device=x.device
          )
          gaussian = torch.exp(
              -(distance[:, None] ** 2 + distance[None] ** 2) / (2 * sigma ** 2)
          )
          gaussian /= gaussian.sum()
          channels = x.size(1)
          kernel = gaussian[None, None].expand(channels, -1, -1, -1)
          return torch.nn.functional.conv2d(x, kernel, padding=width // 2, groups=channels)

To take this to 1d, you save a lot by not needing the funny None-indexing and broadcasting, but you will likely want to rearrange the input tensor to have the folding dimension last to use conv1d.

Best regards

Thomas