This repository was inspired by the pytorch-conv4d repository. It consists of an easy-to-use 4-dimensional convolution class (Conv4d) for PyTorch, in which, 4-dimensional convolution is disassembled into a number of official PyTorch 3-dimensional convolutions. It works by performing and stacking several 3D convolutions under proper conditions (see the original repository for a more detailed explanations).
However, the solution of pytorch-conv4d repository is restricted. Its stride and dilation are constrainedly set to 1. And its bias does not work perfectly. Moreover, the weight initialization does not achieve the uniform distribution in four dimensions. In this repository, some key code has been rewritten to support the arbitrary values for stride, dilation and bias. The code for weight initialization has also been rewritten to realize Kaiming Initialization.
To validate the correctness of this solution, we design 2D and 3D convolutions (validate_2d/validate_3d) equivalent to the official PyTorch version (torch.nn.Conv2d/Conv3d), and compare the outputs of ours and official ones. In our validation, all of values of kernel_size, padding, stride and dilation are set to reasonable random integers, and the error less than 1e-5 is considered qualified. You can directly run validate_2d.py and validate_3d.py to verify again.
An example of Conv4d in main.py
from Conv4d import Conv4d
import torch
input = torch.randn(2, 16, 5, 5, 32, 32).cuda()
net = Conv4d(16, 32,
kernel_size=(3, 1, 1, 1),
padding=(1, 1, 1, 1),
stride=(1, 1, 1, 1),
dilation=(1, 1, 1, 1),
bias=True).cuda()
out = net(input)
print(out.shape)This work is not perfect, but it is sufficient for most application for deep learning. The implementation of groups greater than 1 is still work in progress. And I hope that all padding_modes can be implemented in the future.
Questions or advice for improvements are very welcome!
If this work is useful to you, please consider starring it, so that it can reach a broader audience of like-minded people.