This repository contains pytorch impelementation of the architectures proposed in the paper "Deep Anchored Convolutional Neural Networks". (accepted to CVPR2019 workshops, oral)
All model files can be directly imported into pytorch training codes.
DACNN is a network stacked with a single convolution kernel across layers, incorperating 2 other weight sharing techniques coined "Mixed Architecture and Regulators".
As a network compression technique, the architecture achieved similar performances on CIFAR & SVHN dataset compared to some popular models (VGG, ResNet, etc.) while using much less parameters.
