Overview / Usage

Most of the effort is in trying to bridge the gap between group convolutions and standard spatial convolutions i.e. a new approach to convolution that will have the same number of flops and parameters as group convolutions but give higher accuracies approaching that achievable by standard convolutions. Once the new convolutional module is built, it can be used in any network, simply replacing the group convolutions in that network, keeping the sparsity same.

Methodology / Approach

The new module will be such that for every required output channel, only a subset of neighboring input channels will be used. This should help localise features in different parts of the network and increase accuracy, while using a limited number of input channels will keep the parameters and flops low. Current work includes testing on networks such as DenseNet, MobileNet etc. to see how well the idea works.

Technologies Used

The technologies used include PyTorch, Cuda and Tensor Comprehensions