Is there a way to realize 4d convolution using the ConvNd function

1

I found that conv1d, conv2d, conv3d use ConvNd = torch._C._functions.ConvNd for forward passing. I just want to know whether there is an efficient way to use ConvNd for 4dimension convolution, since my inputs are 4-dimension (6 dimension if count batchsize and channel.)

Thanks !!!

5 Likes
2

Not for the moment. We do not yet have nd convolution kernels, for n>3.

1 Like
3

What’s the blocking issue?

Since we have the ConvNd function, why can’t we implement a 4d convolution just like the others? Why not n-dimensional?

From the Python side, all the convolution layers ultimately delegate to that ConvNd function. Nothing seems special cased for any particular kernel shape.

I think the C++ side is implemented at torch/csrc/autograd/functions/convolution.{h,cpp}. The forward pass has a special case when n=3, but all it seems to do is reshape the kernel by adding/removing dummy dimensions. I’m not entirely sure why though. Perhaps it could be easily generalized? I’m a bit out of my element with these low-level details, so I may be missing something.

4

it’s mostly that we haven’t implemented the C bits going for ConvNd in the THNN/THCUNN libraries. (untimately autograd/functions/convolution.cpp dispatches to that or to CuDNN).

5

@smth could you please share if there are any developments on this front?

1 Like
6

@Edward_Hahn not really, because the demand for 4d convolution is not high. Your best bet is to create N Conv3D layers to do a Conv4D (there will be some boundary effects / issues, so you will have to pad edges appropriately)

7

Just wanted to chime in and say that I’m interested in this feature, it seems like the natural way to represent time series of multi-channel satellite images.

8

Wouldn’t this correspond to a 3D convolution, where the input would have the shape [batch_size, in_channels, seq, height, width] or is each “image” a volume?

9

Here you can find a fully functional convNd and nD transposed convolution based on conv3d as mentioned by smth.

But I’m also looking forward to a native 4D convolution, mainly due to speed concerns. I’m starting with the implementation, but indeed seems like a long way to go.

4 Likes
10

Thanks for your work!

1 Like
11

I’d also LOVE to see this implemented. In my case, it’s a natural way to represent a time series of 3D dimensional microscopy measurements. Hope the pytorch team will consider implementing it.

2 Likes
12

4D eats mad flops

JAX seems to have N-dim

3 Likes
13

I’ve been working on this package…

Which resamples 4 dimensional (T, Z, Y, X) atmospheric analysis along an isobaric log scale. Inversely scaling resolution and extent.

I’m currently working with 5 channels

B, C, T, Z, Y, X = 4, 5, 3, 6, 40, 80

I could also make use of this feature