eduzhai > Applied Sciences > Engineering >

DO-Conv Depthwise Over-parameterized Convolutional Layer

  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Convolutional layers are the core building blocks of Convolutional NeuralNetworks (CNNs). In this paper, we propose to augment a convolutional layerwith an additional depthwise convolution, where each input channel is convolvedwith a different 2D kernel. The composition of the two convolutions constitutesan over-parameterization, since it adds learnable parameters, while theresulting linear operation can be expressed by a single convolution layer. Werefer to this depthwise over-parameterized convolutional layer as DO-Conv. Weshow with extensive experiments that the mere replacement of conventionalconvolutional layers with DO-Conv layers boosts the performance of CNNs on manyclassical vision tasks, such as image classification, detection, andsegmentation. Moreover, in the inference phase, the depthwise convolution isfolded into the conventional convolution, reducing the computation to beexactly equivalent to that of a convolutional layer withoutover-parameterization. As DO-Conv introduces performance gains withoutincurring any computational complexity increase for inference, we advocate itas an alternative to the conventional convolutional layer. We open-source areference implementation of DO-Conv in Tensorflow, PyTorch and GluonCV atthis https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×