This repository contains the extracts from the AWS DeepRacer Robomaker simapp; including

• the environment and tracks
• the robotics agent code ('markov')
• Dockerfile

In most cases it will be sufficient to use one of our pre-built images which use optimized Tensorflow binaries. Depending on how modern processor you have available you can pick which instruction sets to use.

• AVX is available on most Intel Core processors from the second generation on - running DeepRacer on an older processor really does not make sense; hence we do not have pre-built images without this instruction set.
• AVX2 is available on more modern Intel Core processors; starting from the Haswell architecture (4th generation), and is probably the image most will use.
• AVX-512 is available on newer Xeon and Core i7-X processors. If you run DeepRacer on e.g. AWS EC2 this would be a natural choice.
• GPU uses Tensorflow-GPU instead of CPU. It is not compiled with special instruction sets for the CPU.

Available tags are:

The built image is available via docker pull awsdeepracercommunity/deepracer-robomaker:cpu. For other images replace cpu with the tag of your choice from the above table.

To build the CPU image run docker build -t awsdeepracercommunity/deepracer-robomaker:cpu -f docker/Dockerfile.cpu . --build-arg TENSORFLOW_VER='tensorflow==1.11.0

To build the GPU image run docker build -t awsdeepracercommunity/deepracer-robomaker:gpu -f docker/Dockerfile.gpu .

To add a custom tensorflow library run docker build -t awsdeepracercommunity/deepracer-robomaker:cpu -f docker/Dockerfile.cpu . --build-arg TENSORFLOW_VER=<http-url-to-file>

Tensorflow has been built with the following parameters:

• cpu-avx with -march=x86-64 -maes -mavx -mmmx -mpclmul -mpopcnt -msse -msse2 -msse3 -msse4.1 -msse4.2 -mssse3
• cpu-avx2 with -march=x86-64 -maes -mavx -mavx2 -mmmx -mpclmul -mpopcnt -msse -msse2 -msse3 -msse4.1 -msse4.2 -mssse3 -mf16c -mfma