This repository contains adaptation of DeepDriving project for Tensorflow in addition to some supplementary scripts for training, evaluating and visualizing Convolutional Network used for driving.

The original training and test data can be downloaded from DeepDriving website.

wget http://deepdriving.cs.princeton.edu/TORCS_trainset.zip # 50GB
wget http://deepdriving.cs.princeton.edu/TORCS_baseline_testset.zip # 8GB

Original training data was recorded from 12 hours of human driving in TORCS video game and contains 484815 entries in LevelDB format. Each entry is a screenshot from the game and 14 numbers describing current road scene. (Original article mysteriously describes only 13 numbers)

To train a CNN in Tensorflow this data can be converted from LevelDB to TFRecords using leveldb2tfrecords.py script. This script depends on Caffe, so it could be launched inside Caffe docker.

The CNN given an image should predict 14 numbers, next steering command is computed based on those numbers . All targets are rescaled to [0..1], so that Euclidean L2 loss can be used for training. We trained AlexNet and MobileNet architectures with modified number of outputs:

Test MAE of AlexNet after 140k iterations:

To visualize, what CNN has learned, one can highlight parts of image, which influence the output of the network at most. For instance, one can compute the gradient of network output with respect to an input image. To generate the picture below, we've computed the gradient of the output node, which predicts distance to car ahead in the current lane, and highlighted each pixel in proportion to the magnitude of the gradient.

Jupyter notebooks for training and evaluating are gathered in notebooks subfolder:

net.ipynb - CNN training

evaluate.ipynb - CNN evaluation on test data

progress.ipynb - visualize current loss

attention.ipynb - visualize most influential parts of image

This code can be launched on a computer with NVidia graphics card and installed NVidia driver. You should download TORCS sources and modified tracks (torcs-1.3.6 and modified_tracks from DeepDriving) into this repository.

The code should be launched inside Docker container, so that you don't have to resolve dependencies. Dockerfile is provided, one can build an image with build_docker_image.sh. After image is built, container can be launched with run_docker.sh.

Following steps should be launched inside Docker container.

To build driver with Tensorflow, one have to download Tensorflow sources with ``get_TF.sh````.

After that, one should come into tensorflow folder, and build a driver, for instance, for 3 lane road (this may take a couple of hours)

bazel build --config=opt --config=cuda //tensorflow/cc/driver:drive_3lane

Next, one should launch torcs, configure players, start quick race and remove extra elements from the screen in accordance to README of the original DeepDriving.

Finally, driver can be launched with

./bazel-bin/tensorflow/cc/driver/drive_3lane

and started in accordance to README of the original DeepDriving.