Skip to content

tmaila/autopilot

 
 

Repository files navigation

Training and Inference for End-to-End Self-Driving Car Deep Learning Models

Overview

Deep learning based self driving cars need to make steering and other control decisions at deterministic pace to be able to safely control the behavior of the vehicle in traffic. A car traveling 70 mph moves one feet every 10 ms. One feet can be the difference between successful corrective action and a fatal accident. In this project we analyze the inference execution time determinism and jitter. We compare Python, C++ and real-time optimized C++ deep learning inference implementations for self driving cars using TensorFlow. We expect some jitter and jitter of the order of 5-10 ms is probably acceptable for a self driving car but 50 ms delays probably would be too much for such a critical control system.

In this project we have created a TensorFlow implementation of the NVIDIA End-to-End Deep Learning for Self-Driving Cars self driving car model. Our implementation is based on the Python TensorFlow implementation by Sully Chen.

The project consists of a python based training script and a python and two C++ based inference implementations. At the end of the training the training script saves a graph definition where all weights are replaced with corresponding constant values. This saved graph is then loaded by the Python and both C++ inference implementations.

As the execution environment we used Ubuntu 16.10 with a custom Linux 4.9.13 kernel with a Preempt RT real-time patch. This patch adds hard real-time support to the operating system. The inference was performed on a NVIDIA GTX 1060 GPU.

Setting Up

Install Preempt RT Kernel

To be able to fully benefit from the real-time optimizations it is recommended to install the Preempt RT kernel patch on Linux. Building and installing the kernel patch is outside the scope of this readme. Everything should work on a regular linux as well but the timing statistics would be different.

Checkout TensorFlow Sources

To use the C++ inference, you will need to have TensorFlow source code and all dependencies to be able to build TensorFlow from sources.

~$ git clone https://github.com/tensorflow/tensorflow
~$ cd tensorflow
~/tensorflow$

You will need r1.0 or newer release of TensorFlow.

Build TensorFlow

Follow the instruction in TensorFlow Documentation on how to download and build TensorFlow from the sources.

Note: You don't have to install the TensorFlow pip package build from sources if you already have TensorFlow r1.0 or later installed that you'd prefer to use. However validating the build is difficult without installing the package so at the minimum test the pip package you build in a virtualenv. If the version of TensorFlow you have installed is different from the version of source code you checked out, you may run into problems.

Checkout This Source Code

After you have TensorFlow build from sources installed, check out this source code under tensorflow subfolder of the TensorFlow sources.

~/tensorflow$ cd tensorflow
~/tensorflow/tensorflow$ git clone https://github.com/tmaila/autopilot.git

Get Dataset

Download the driving dataset and extract into the repository folder. After the extraction you should have a driving_dataset subfolder under your repository folder.

Preparing Dataset

We have separated the data preprocessing to its own separate step to purely focus on running the data through the graph in the inference step.

Use python preprocess.py to prepare the dataset for training.

python preprocess.py

How To Build

The C++ executable for the project is being build using TensorFlow's bazel build system. Use bazel build -c opt ... to build the executable.

~/tensorflow/tensorflow$ cd autopilot
~/tensorflow/tensorflow/autopilot$ bazel build -c opt ...

You may want to pass optimization flags to bazel. These enable cuda gpu support if not enable by default and some cpu optimizations.

$ bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda ...

How To Use

Training the Model

Once everything is set up the model needs to be trained with the prerecorded driving data.

Use python train.py to train the model.

python train.py

Training the model should create a save folder that will contain a graph.pb and frozen_graph.pb files as well as some checkpoint files. The graph.pb contains our model without weight variables and the frozen_graph.pb contains the same model but with weight variables converted into constants. We use the latter for inference.

Running the Model Inference

After the model is trained we can run the inference on the model. We are going to run through all the data in the dataset to evaluate the inference execution time performance. Typically you don't want to use your test and validation sets for inference, however that is not a problem in our case as we are not evaluating the prediction performance of our model but purely focused on timing statistics. If you prefer to use some other driving data set for inference, feel free to do so.

Python

Use python run.py to run the model inference on the dataset using Python.

python run.py

Running the inference through the data will take a while. We are not logging any progress metrics to the screen as it may impact the timing statistics that we are interested in.

C++

Use the executable we built above to run the inference on the dataset using C++.

../../bazel-bin/tensorflow/autopilot/autopilot

Real-Time Optimized C++

Use the executable we built above to run the inference on the dataset using real-time optimized C++.

sudo ../../bazel-bin/tensorflow/autopilot/autopilot-rt

Analyzing the Results

Running the inferences using Python and C++ will create run log files under the results folder. To analyze the timing statistics and compare the inference implementations we have create a Jupyter notebook.

jupyter notebook autopilot.ipynb

To view the precomputed Jupyter notebook, open the autopilot.html file with your browser.

You can also create the plots shown in the notebook by running the analyze.py script without Jupyter. This is handy when you want to make changes to the inference implementation and view the outcome of your changes.

python analyze.py

Releases

No releases published

Packages

No packages published

Languages

  • HTML 66.4%
  • Jupyter Notebook 26.2%
  • C++ 4.5%
  • Python 2.9%