DiffConv2d

This is the code repository for the following paper:

K. Leng and J. Thiyagalingam, Padding-free Convolution based on
Preservation of Differential Characteristics of Kernels.

Installation

PyTorch is the only dependency for applications. If you need a TensorFlow version, please contact us.

Usage

All source files needed for applications are included in diff_conv2d/.

Part I. Image Filtering by Convolution

If you are using the following PyTorch API for image convolution:

output = torch.nn.functional.conv2d(input, weight, padding='same', bias=None,
                                    stride=1, dilation=1, groups=any_groups)

you can then switch to our method by

from diff_conv2d.functional import DiffConv2d

diff = DiffConv2d(kernel_size=3)  # kernel_size can be 3, 5, or 7
output = diff.conv2d(input, weight, groups=any_groups,
                     keep_img_grad_at_invalid=True,
                     edge_kernel=None, optimized_for='memory')

The additional arguments are explained below:

keep_img_grad_at_invalid: specifies whether to keep the gradient of input pixels repeatedly used for boundary handling. For example, when using padding by padding_mode=replicate with a normal Conv2d layer, one may consider turning off the gradient with respect to the replicated pixel values; however, PyTorch does not offer such an option. With our DiffConv2d.conv2d() method, one can do this by keep_img_grad_at_invalid=False.
edge_kernel: an extra kernel provided for Explict Boundary Handling (paper), which has the same shape as weight. It is usually None for forward image filtering, meaning that weight is used for both valid ( interior) and invalid (boundary) pixels.
optimized_for: specifies whether the implementation is optimized for 'speed' or 'memory'. Depending mainly on the image shape, 'speed' can make the code run faster by 20~40% at the cost of 50%~100% more GPU memory; the memory required by 'memory' is close to normal padding, so it is recommended at a development stage.

Limitation: though our method is compatible with stride and dilation, the current implementation assumes stride=1 and dilation=1. If you need this extension, please contact us.

Part II. Convolutional layers

If you are using the following PyTorch class in your CNNs,

layer = torch.nn.Conv2d(in_channels, out_channels, kernel_size,
                        stride=1, padding='same', dilation=1, groups=any_groups,
                        bias=any_bias, padding_mode=any_padding_mode)

you can then switch to our method by

from diff_conv2d.layers import DiffConv2dLayer

layer = DiffConv2dLayer(in_channels, out_channels, kernel_size,
                        groups=any_groups, bias=any_bias,
                        keep_img_grad_at_invalid=True, train_edge_kernel=False,
                        optimized_for='memory')

The additional arguments are explained below:

keep_img_grad_at_invalid: has the same meaning as above.
train_edge_kernel: specifies whether to activate Explict Boundary Handling (paper) by training an extra edge kernel. If train_edge_kernel=True, your layer will contain twice the weights as a normal torch.nn.Conv2d layer has, one for interior and the other for boundary pixels.
optimized_for: has the same meaning as above.

Limitation: again, the current implementation assumes stride=1 and dilation=1. If you need this extension, please contact us.

Experiments

If you wish to reproduce our experiments, please follow these steps:

Move all the files and folders from experiments/ to the directory including diff_conv2d/ (such as the root of this repo).
Install dependencies for experiments:

pip install -r exp_requirements.txt

Refer to exp_*_readme.md for detailed guidance on each experiment, including data downloading, training and plotting.

Acknowledgement

This work is supported by the EPSRC grant, Blueprinting for AI for Science at Exascale (BASE-II, EP/X019918/1), which is Phase II of the Benchmarking for AI for Science at Exascale (BASE) grant.