Skip to content
Learning Versatile Filters for Efficient Convolutional Neural Networks - NeurIPS 2018
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
README.md

README.md

Versatile Filters

Code for paper: Learning Versatile Filters for Efficient Convolutional Neural Networks (NeurIPS 2018)

We introduce versatile filters to construct efficient convolutional neural network. Considering the demands of efficient deep learning techniques running on cost-effective hardware, a number of methods have been developed to learn compact neural networks. Most of these works aim to slim down filters in different ways, e.g. investigating small, sparse or binarized filters. In contrast, we treat filters from an additive perspective. A series of secondary filters can be derived from a primary filter. These secondary filters all inherit in the primary filter without occupying more storage, but once been unfolded in computation they could significantly enhance the capability of the filter by integrating information extracted from different receptive fields. Besides spatial versatile filters, we additionally investigate versatile filters from the channel perspective. The new techniques are general to upgrade filters in existing CNNs. Experimental results on benchmark datasets and neural networks demonstrate that CNNs constructed with our versatile filters are able to achieve comparable accuracy as that of original filters, but require less memory and FLOPs.

Fig. An illustration of the proposed versatile convolution filter. The input data is first divided into several areas according to the size and parameters of the a 5 x 5 convolution filter, the proposed versatile convolution filter is then applied three times according to its secondary filters (i.e., 5 x 5 blue, 3 x 3 green, and 1 x 1 red) to generate multiple features.

Experimental results on ImageNet dataset. Details in our paper Learning Versatile Filters for Efficient Convolutional Neural Networks (NeurIPS 2018)

Files description

Platform: Pytorch 0.4

vcnn.py is the implementation of Versatile Convolution (an example of VGG-16). The VConv2d class can be used to replace the nn.Conv2d in any CNN.

imagenet-vcnn.py is the script for training ImageNet on Cloud DLS.

config.png is a config example on Cloud DLS.

Hyper-paprameters

In VConv2d:

  • delta: (c-\hat{c}) in Eq.(6)
  • g: g in Eq.(6)

Performance

backbone method top1 acc top5 acc
VGG-16 baseline 71.5 90.1
spatial versatile 72.2 91.1
spatial+channel versatile 70.4 89.6
You can’t perform that action at this time.