leo-XUKANG/Darknet-Cross

Light-weighted & Multi-platform & Multi-type GPU support Darknet Framework

Darknet-Cross

Darknet-Cross is a lightweight deep learning framework based on Darknet and yolov2_light. It provides computation acceleration for multi-platform (Eg.Ubuntu/Android) and multiple types of GPUs (Eg. Nvidia GTX1070/ Adreno 630).

Darknet-Cross is part of my MSc diploma project in HKU CS.

The source code folder will be uploaded after organization.

Demo

The Darknet-Cross demo video has been uploaded to YouTube:

【Darknet-Cross User Guide】 includes all steps and specific tests in different versions.

【Darknet-Cross Performance】 includes detailed experiment results and comparation results.

Contents

I. Darknet-Cross Introduction

1. Framework Compare

1.1 Original Darknet

1.2 yolov2_light

1.3 Darknet-Cross

2. YOLO-V3 Model

3. Darknet-Cross Function

II. Darknet-Cross Version

1. Project Structure

1.1 Makefile

1.2 Source Code Folder

1.2.1 File/Folder in All Version

1.2.2 File/Folder in OpenCL Version

1.2.3 File/Folder in CUDA Version

1.3. 3rdParty Folder

1.4. Object Folder

1.5. Bin Folder

2. Workflow

2.1 Ubuntu Version

2.2 Android Version

III. Darknet-Cross Configuration

1. Environment Configuration in Ubuntu 16.04

1.1 Ubuntu-CPU Version

1.2 Ubuntu-CUDA Version

1.2.1 Deep Learning Environment Setting

1.2.2 OpenCV Installation

1.2.3 Ubuntu-CUDA Version Compile

1.3 Ubuntu-OpenCL Version

1.3.1 OpenCL Installation

1.3.2 CLBlast Installation

1.3.3 Ubuntu-OpenCL Version Compile

2. Cross-compiling Environment Configuration for Android 8.1.0

2.1 Cross-compiling Configuration

2.1.1 Hardware Setting

2.1.2 Install Cross-Compiling Toolchain

2.2 3rd Parties Compiling

2.2.1 OpenCL

2.2.2 CLBlast

2.2.3 OpenCV (Not used in this stage)

2.3 Android-CPU Version

2.4 Andorid-OpenCL Version

IV. Darknet-Cross Commands

1. Compiling

2. Test

2.1 Model Choose

2.2 Image Test

2.3 Video Test

2.4 Multi-Image Test

V. Darknet-Cross Performance

1. Foreword: Relationship of DNN model, DL Framework and Hardware Environment

1.1 DNN Model

1.2 DL Framework

1.3 Hardware Platform

1.4 Conclusion

2. Test Environment

2.1 Platform

2.2 Data

2.2.1 Single Image Data

2.2.2 Video Data

2.2.3 Multi-Image Data

3. Detailed Information for Experiments

4. Multi-framework Test (Darknet-Cross vs. Darknet)

4.1 Executable File Size Comparation

4.2 GPU Usage Rate Comparation

4.3 Process Speed Comparation

4.3.1 Test on Video

4.3.2 Test on Single Image

5. Multi-Platform Test (Different Verions Comparation in Darknet-Cross)

5.1 Foreword: Relationship of Prediction Time in Multi-Image and FPS in Video

5.1.1 Test on YOLO-V3-Tiny

5.1.2 Test on YOLO-V3

5.1.3 Conclusion

5.2 Multi-Platform Process Speed Comparation

5.2.1 Compare Group1 & Group2 (Single Image: Ubuntu vs. Android)

5.2.2 Compare Group3 & Group4 (Multi-Image: Ubuntu vs. Android)

5.2.3 Compare Group5 & Group7 and Group6 & Group8 (Verify OpenCL Acceleration)

5.2.4 Compare Latency in Different Version