Deepdrive
Last updated
Was this helpful?
Last updated
Was this helpful?
The easiest way to experiment with self-driving AI
Linux
Python 3.6+
10GB disk space
8GB RAM
CUDA capable GPU (tested and developed on 970, 1070, and 1060's)
1.7 <= Tensorflow < 2.0
Clone
Optional - Activate the Python conda env or virtualenv with Tensorflow installed, then
Note: If you use Anaconda (we recommend Miniconda) - only use
pip install
in your deepdrive conda environment, neverconda install
Install
Note that we recorded the baseline dataset in sync mode which is much slower than async mode. Async mode probably is fine to record in, we just haven't got around to trying it out for v3.
Optional: Convert to HDF5 files to tfrecords (for training MNET2)
Where <your-deepdrive-home>
below is by default in $HOME/Deepdrive
and can be configured in $HOME/.deepdrive/deepdrive_dir
Esc - Pause (Quit in Unreal Editor)
Enter - Pause with no menu
P - Pause in Unreal Editor
1 - Chase cam
2 - Orbit (side) cam
3 - Hood cam
4 - Free cam (use WASD to fly)
Space - Handbrake
Alt+Tab - Control other windows / Show mouse
` - Unreal console - do things like stat FPS
M - Drive the car with the keyboard WASD - be sure sync is off - Also known issue: Only works in path-follower mode right now
Ctrl-number - Change sun position - works for 1 => 7
B - Show vehicle bounding boxes
N - Show vehicle collision boxes
Page Up - Next vehicle
Page Down - Prev vehicle
All values returned in the observation keep Unreal conventions, specifically
All distances are in centimeters per Unreal's default data type
All rotations / angular values are in the order of roll, pitch, yaw in degrees
x,y,z is forward, right, up
100GB (8.2 hours of driving) of camera, depth, steering, throttle, and brake of an 'oracle' path following agent. We rotate between three different cameras: normal, wide, and semi-truck - with random camera intrisic/extrinsic perturbations at the beginning of each episode (lap). This boosted performance on the benchmark by 3x. We also use DAgger to collect course correction data as in previous versions of Deepdrive.
Ensure you have 104GB of free space
Download our dataset of mixed Windows (Unreal PIE + Unreal packaged) and Linux + variable camera and corrective action recordings
(generated with --record
)
or for the legacy HDF5 files for training AlexNet
On Windows, use standard (non-CUDA packaged) display drivers which meet the min required. When installing CUDA, do a custom install and uncheck the display driver install.
glxinfo | grep OpenGL
should return something like:
You may need to disable secure boot in your BIOS in order for NVIDIA’s OpenGL and tools like nvidia-smi to work. This is not Deepdrive specific, but rather a general requirement of Ubuntu’s NVIDIA drivers.
To run tests in PyCharm, go to File | Settings | Tools | Python Integrated Tools and change the default test runner to pytest
.
- operates over the network using the
Grab the
Additional observation data can be exposed without compiling C++ or Blueprints by accessing the Unreal API with .
Agents are automatically graded via
Get the
If you'd like to check out our Tensorboard training session, you can download the 1GB , unzip, and run
and checkout , which graphs wall time.
If you experience low frame rates on Linux, you may need to install NVIDIA’s display drivers including their OpenGL drivers. We recommend installing these with CUDA which bundles the version you will need to run the baseline agent. Also, make sure to . If CUDA is installed, skip to testing .
Make sure to install the CUDA / cuDNN major and minor version the Tensorflow instructions specify. i.e. CUDA 9.0 / cuDNN 7.3 for Tensorflow 1.12.0. These will likely be older than the latest version NVIDIA offers. You can see all .
Use the packaged install, i.e. deb[local] on Ubuntu, referred to in
If you are feeling dangerous and use the runfile method, be sure to follow on how to disable the Nouveau drivers if you're on Ubuntu.
Also, disable SciView per .