Welcome to the DeepDriving for Tensorflow Project

Introduction

The DeepDriving project was originally created by the Princeton University. This original project can be found here. When I first read the corresponding paper I thought: "Hey this is the perfect project for my upcoming semester work on my university". Thus I started to reproduce the results from the paper and to port the AlexNet, used by this project, to Tensorflow.

As a very first step, I reproduced the results with a more recent caffe implementation. This implementation can be found here. In a second step, I created a Tensorflow based deep-learning framework which allows the training and evaluation of neural networks with Tensorflow without much overhead. This implementation is very generic and many different deep-learning tasks can be performed using this framework.

In the end, I was able to port the AlexNet, used by the original project, to a tenesorflow implementation based on my deep-learning framework. On the upcoming wiki-pages, I will describe, how to install the framework and how to use it for deep-learning tasks. Furthermore I will describe how to reproduce the results from the DeepDriving project and how to train an own network to drive on a computer game.

In contrast to the original implementation, I did not use TORCS as computer game, but SpeedDreams, which is a fork of TORCS. SpeedDreams comes with a very nice CMake-based build system, which makes it easy to compile and install the game on Windows. The following descriptions are for Windows and Ubuntu. A DeepDriving-ready implementation of SpeedDreams can be found here.

On YouTube, you can find some videos which shows this project in action. Have a lot of fun with driving!

Index

• Installation on Windows
  • Compile the C-Libraries
  • Setup your Python 3 Environment
• Installation on Ubuntu
  • Compile the C-Libraries
  • Setup your Python 3 Environment
• The Deep-Learning Framework
  • Basics
  • Setup and Train Cifar-10 Data
• The DeepDriving Project
  • Prepare the DeepDriving-Training-Data
  • Train the DeepDriving-Network
  • Evaluate the DeepDriving-Network
  • Drive in SpeedDreams
  • Create your own Training-Data

License

Since this project is derived from DeepDriving, the original license is still valid for all derived code parts. This is especially the code for the Situation-View (C-Library) and the Drive-Controller (C-Libraray). But also the code for normalizing and denormalizing the labels and output of the network. In general there should be a note inside the file, if the corresponding code is derived from the original project.

The remaining python code is under the MIT license and the C/C++ code is under the GNU library.

Additional Information

• Results