This project is now deprecated and will not work with the current versions of the AWS Deepracer console or bundle. You should use DRFC instead.

---

DeepRacer local training (2020 version)

Heavily based off work by Crr0004, AlexSchultz, Richardfan1126 and LarsLL

Prerequisites

This project is designed to run on a linux system, ideally with an nvidia GPU. CPU training is possible but will be very slow. AMD GPUs are not currently supported. Ubuntu 18.04 has been extensively tested.

1. install nvidia cuda drivers and tools.

2. install docker and docker-compose

3. set docker-nvidia2 as default runtime in your /etc/docker/daemon.json

   {
    "default-runtime": "nvidia",
       "runtimes": {
           "nvidia": {
               "path": "nvidia-container-runtime",
               "runtimeArgs": []
           }
       }
   }
   

Visit the Community Knowledge Base for a list of detailed resources and checks to confirm your prerequisites are set up properly.

Configure training session

1. Edit the reward function in data/minio/bucket/custom_files/reward.py

2. Edit the action space, CNN layers and sensors in data/mini/bucket/custom_files/model_metadata.json

3. Edit the training params in config.env and data/minio/bucket/custom_files/training_params.yaml. Note that the track name MUST be the same in both files!

   Useful options include:

   option	description
   ENABLE_GPU_TRAINING	Enables GPU for SageMaker runtime: true (nvidia runtime) or false (CPU runtime). Default is GPU
   ENABLE_LOCAL_DESKTOP	Set to true if you have a local X-windows install (desktop machine) and want to automatically start the stream viewer and tail sagemaker and robomaker logs.
   ENABLE_TMUX	Enables tmux for automatic log tails in your existing terminal session (good for remote servers)
   ENABLE_GUI	Enables gazebo client. Access via vnc on localhost:8080
   WORLD_NAME	The track name. Tracks are contained within the robomaker container image, built from the deepracer-simapp community project (excluding the .world suffix)

   Many other options are available.

4. Edit hyperparameters in hyperparams.json inside src/rl_coach_2020_v2/hyperparams.json - a symlink has been created in the root directory.

   More information on configuring local training can be found at https://wiki.deepracing.io/Customise_Local_Training

Starting a training session

Run ./start-training.sh to start training.

The current model data dir (defaults to data/minio/bucket/current) must be empty.

To use a pretrained model as a base for a new training session rename data/minio/bucket/current to data/minio/bucket/rl-deepracer-pretrained and set "pretrained": "true" in hyperparams.json

The first run will likely take quite a while to start as it needs to pull over 10GB of all the docker images. You can avoid this delay by pulling the images in advance:

• docker pull awsdeepracercommunity/deepracer-sagemaker:<cpu or gpu>
• docker pull awsdeepracercommunity/deepracer-robomaker:<cpu or gpu>
• docker pull mattcamp/dr-coach
• docker pull minio/minio

Note that different flavours of CPU image are available, see https://github.com/aws-deepracer-community/deepracer-simapp for details. cpu-avx2 is the default.

Monitoring training

• Docker logs should open automatically in new terminal tabs if running with ENABLE_LOCAL_DESKTOP enabled, or via tmux in your existing terminal session if ENABLE_TMUX is enabled.
• Logs can be manually viewed using docker ps and docker logs robomaker or docker logs <sagemaker_container_id>
• The web video stream is available by default on port 8888. If running in desktop mode a browser window should open automatically, otherwise you can try opening a url such as http://127.0.0.1:8888/stream_viewer?topic=/racecar/deepracer/kvs_stream
• Kinesis video stream can also be enabled. See below for more details, however usually the web video stream just works better.
• if ENABLE_GUI is enabled then you can connect a vncviewer on port 8080 to view the gazebo client directly.

Stopping training

Run ./stop-training.sh to stop training.

If running, sagemaker will be stopped first and then after a 20s delay the rest of the containers will be stopped. This allows Robomaker to create a model.tar.gz file in the current model dir, ready to be loaded onto a physical DeepRacer car.

NOTE: Sagemaker should not be stopped during the policy training phase or things might get weird and corrupt. You should only stop training while the video stream status is "Training" and not "Evaluating" (or verify via sagemaker logs that policy training has completed for the current iteration)

Model management

• run ./delete_last_run.sh to clear out the buckets for a fresh run. For convenient version without sudo prompt check out utilites/delete-last.c.
• run ./local-copy.sh <model_backup_name> to backup current model files into user specified MODEL directory.
• run ./mk-model.sh <model_path> to create physical car uploadable .tar.gz file from your model. (Will be removed in a future update once file gets correctly generated after training)

Kinesis video stream:

Kinesis video currently only works via the real AWS Kinesis service and probably only makes sense if you are training on an EC2 instance.

To use Kinesis:

• create a real AWS user (with programmatic access keys) which has a policy attached that allows Kinesis access.
• Update the AWS keys in config.env (including the minio ones) to match the user you have created.
• Create a stream in Kinesis with a name to match the KINESIS_VIDEO_STREAM_NAME value (in config.vars) in region eu-west-1
• Set ENABLE_KINESIS to true in config.env

Kinesis video is a stream of approx 1.5Mbps so beware the impact on your AWS costs and your bandwidth.

Once working the stream should be visible in the Kinesis console.

Known issues:

• Sagemaker will occasionally fail to start with an error saying /opt/ml/input/config/resourceconfig.json is missing. This is currently proving hard to reproduce and fix but seems fairly rare.
• There is a possibility that stopping training at the wrong time could problem where sagemaker will crash next time when trying to load the 'best' model which may not exist properly. Work is continuing to make this more robust but to be safe you should only stop training while the status is "Training" and not "Evaluating".
• training_params.yaml must exist in the target bucket or robomaker will not start. The start-training.sh script will copy it over from custom_files if necessary.
• Scripts are not currently included to handle uploading to AWS Console or virtual league. https://github.com/cahya-wirawan/deepracer-tools includes such tools.
• Current Sagemaker and Robomaker GPU images are built for nvidia GPU only.
• The Sagemaker and Robomaker images are huge (~4.5GB) and can take a long time to download, especially on the first run.

Getting help

Join #dr-local-training-setup on the AWS Machine Learning Community Slack at https://deepracing.io