DO-Conv: Depthwise Over-parameterized Convolutional Layer

TL;DR

DO-Conv introduces an over-parameterized convolutional layer that enhances CNN performance on vision tasks without increasing inference complexity by combining a depthwise convolution with a standard convolution.

Contribution

The paper proposes DO-Conv, a novel over-parameterized convolutional layer that improves CNN accuracy while maintaining inference efficiency through layer folding.

Findings

Boosts CNN performance on classification, detection, and segmentation tasks.

Increases training accuracy without additional inference cost.

Open-source implementation available in major deep learning frameworks.

Abstract

Convolutional layers are the core building blocks of Convolutional Neural Networks (CNNs). In this paper, we propose to augment a convolutional layer with an additional depthwise convolution, where each input channel is convolved with a different 2D kernel. The composition of the two convolutions constitutes an over-parameterization, since it adds learnable parameters, while the resulting linear operation can be expressed by a single convolution layer. We refer to this depthwise over-parameterized convolutional layer as DO-Conv. We show with extensive experiments that the mere replacement of conventional convolutional layers with DO-Conv layers boosts the performance of CNNs on many classical vision tasks, such as image classification, detection, and segmentation. Moreover, in the inference phase, the depthwise convolution is folded into the conventional convolution, reducing the computation to be exactly equivalent to that of a convolutional layer without over-parameterization. As DO-Conv introduces performance gains without incurring any computational complexity increase for inference, we advocate it as an alternative to the conventional convolutional layer. We open-source a reference implementation of DO-Conv in Tensorflow, PyTorch and GluonCV at https://github.com/yangyanli/DO-Conv.