An Implementation of Dilated Convolutional Layer Based on Darknet Architecture

This implementation of dilated convolutional layer was accomplished during my internship in Zerotech Inc., Beijing. The need for dilated convolutional layer comes from training CSRNet for crowd detection.

About Darknet

Darknet is an open source neural network framework written in C and CUDA by Joseph Redmon. It is fast, easy to install, and supports CPU and GPU computation. For more information please refer to the Darknet project website.

Darknet Dependency

Darknet has only two dependencies, and they are optional:

* OpenCV - for video I/O and more image types support
* CUDA - for GPU acceleration

This dilated convolutional layer implementation on GPU was written with CUDA Toolkit 9.0

Darknet Installation

The Makefile is provided in the main folder, to install Darknet with dilated convolutional layer, type:

make

Dilated Convolutional Layer

The added layer was tested by comparing with the output of Caffe's implementation of dilated conv layer, and it was also tested by the training of CSRNet composed of it.

Configuring the Dilated Convolutional Layer

Darknet uses .cfg file as the configuration of Neural Network. In the .cfg file, the dilated convolutional layer is configured with:

[dilated_convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
dilate_rate=1

Here's a demonstration of dilated convolution and dilate rate: Dilated convolution expands its convolutional kernel according to dilate rate. In this demonstration, dilate rate = 1. If dilate rate = 0, dilated convolution is simply normal convolution.

CSRNet composed with this implementation training result

Yolov3-tiny with dilated conv layer trained from scratch

This configuration file is included in ./cfg/yolo-d-tiny.cfg