Deepracer-for-Cloud Reference

Environment Variables

The scripts assume that two files system.env containing constant configuration values and run.env with run specific values is populated with the required values. Which values go into which file is not really important.

Variable	Description
DR_RUN_ID	Used if you have multiple independent training jobs only a single DRfC instance. This is an advanced configuration and generally you should just leave this as the default 0.
DR_WORLD_NAME	Defines the track to be used.
DR_RACE_TYPE	Valid options are TIME_TRIAL, OBJECT_AVOIDANCE, and HEAD_TO_BOT.
DR_CAR_COLOR	Valid options are Black, Grey, Blue, Red, Orange, White, and Purple.
DR_CAR_NAME	Display name of car; shows in Deepracer Console when uploading.
DR_ENABLE_DOMAIN_RANDOMIZATION	If True, this cycles through different environment colors and lighting each episode. This is typically used to make your model more robust and generalized instead of tightly aligned with the simulator
DR_UPLOAD_S3_PREFIX	Prefix of the target location. (Typically starts with DeepRacer-SageMaker-RoboMaker-comm-
DR_EVAL_NUMBER_OF_TRIALS	How many laps to complete for evaluation simulations.
DR_EVAL_IS_CONTINUOUS	If False, your evaluation trial will end if you car goes off track or is in a collision. If True, your car will take the penalty times as configured in those parameters, but continue evaluating the trial.
DR_EVAL_OFF_TRACK_PENALTY	Number of seconds penalty time added for an off track during evaluation. Only takes effect if DR_EVAL_IS_CONTINUOUS is set to True.
DR_EVAL_COLLISION_PENALTY	Number of seconds penalty time added for a collision during evaluation. Only takes effect if DR_EVAL_IS_CONTINUOUS is set to True.
DR_EVAL_SAVE_MP4	Set to True to save MP4 of an evaluation run.
DR_EVAL_REVERSE_DIRECTION	Set to True to reverse the direction in which the car traverses the track.
DR_TRAIN_CHANGE_START_POSITION	Determines if the racer shall round-robin the starting position during training sessions. (Recommended to be True for initial training.)
DR_TRAIN_ALTERNATE_DRIVING_DIRECTION	True or False. If True, the car will alternate driving between clockwise and counter-clockwise each episode.
DR_TRAIN_START_POSITION_OFFSET	Used to control where to start the training from on first episode.
DR_TRAIN_ROUND_ROBIN_ADVANCE_DISTANCE	How far to progress each episode in round robin. 0.05 is 5% of the track. Generally best to try and keep this to even numbers that match with your total number of episodes to allow for even distribution around the track. For example, if 20 episodes per iternation, .05 or .10 or .20 would be good.
DR_TRAIN_MULTI_CONFIG	True or False. This is used if you want to use different run.env configurations for each worker in a multi worker training run. See multi config documentation for more details on how to set this up.
DR_TRAIN_MIN_EVAL_TRIALS	The minimum number of evaluation trials run between each training iteration. Evaluations will continue as long as policy training is occuring and may be more than this number. This establishes the minimum, and is generally useful if you want to speed up training especially when using gpu sagemaker containers.
DR_TRAIN_REVERSE_DIRECTION	Set to True to reverse the direction in which the car traverses the track.
DR_TRAIN_BEST_MODEL_METRIC	Can be used to control which model is kept as the "best" model. Set to progress to select the model with the highest evaluation completion percentage, set to reward to select the model with the highest evaluation reward.
DR_TRAIN_MAX_STEPS_PER_ITERATION	Can be used to control the max number of steps per iteration to use for learning, the excess steps will be discarded to avoid out-of-memory situations, default is 10000.
DR_LOCAL_S3_PRETRAINED	Determines if training or evaluation shall be based on the model created in a previous session, held in s3://{DR_LOCAL_S3_BUCKET}/{LOCAL_S3_PRETRAINED_PREFIX}, accessible by credentials held in profile {DR_LOCAL_S3_PROFILE}.
DR_LOCAL_S3_PRETRAINED_PREFIX	Prefix of pretrained model within S3 bucket.
DR_LOCAL_S3_MODEL_PREFIX	Prefix of model within S3 bucket.
DR_LOCAL_S3_BUCKET	Name of S3 bucket which will be used during the session.
DR_LOCAL_S3_CUSTOM_FILES_PREFIX	Prefix of configuration files within S3 bucket.
DR_LOCAL_S3_TRAINING_PARAMS_FILE	Name of YAML file that holds parameters sent to robomaker container for configuration during training. Filename is relative to s3://{DR_LOCAL_S3_BUCKET}/{LOCAL_S3_PRETRAINED_PREFIX}.
DR_LOCAL_S3_EVAL_PARAMS_FILE	Name of YAML file that holds parameters sent to robomaker container for configuration during evaluations. Filename is relative to s3://{DR_LOCAL_S3_BUCKET}/{LOCAL_S3_PRETRAINED_PREFIX}.
DR_LOCAL_S3_MODEL_METADATA_KEY	Location where the model_metadata.json file is stored.
DR_LOCAL_S3_HYPERPARAMETERS_KEY	Location where the hyperparameters.json file is stored.
DR_LOCAL_S3_REWARD_KEY	Location where the reward_function.py file is stored.
DR_LOCAL_S3_METRICS_PREFIX	Location where the metrics will be stored.
DR_OA_NUMBER_OF_OBSTACLES	For Object Avoidance, the number of obstacles on the track.
DR_OA_MIN_DISTANCE_BETWEEN_OBSTACLES	Minimum distance in meters between obstacles.
DR_OA_RANDOMIZE_OBSTACLE_LOCATIONS	If True, obstacle locations will randomly change after each episode.
DR_OA_IS_OBSTACLE_BOT_CAR	If True, obstacles will appear as a stationary car instead of a box.
DR_OA_OBJECT_POSITIONS	Positions of boxes on the track. Tuples consisting of progress (fraction [0..1]) and inside or outside lane (-1 or 1). Example: "0.23,-1;0.46,1"
DR_H2B_IS_LANE_CHANGE	If True, bot cars will change lanes based on configuration.
DR_H2B_LOWER_LANE_CHANGE_TIME	Minimum time in seconds before car will change lanes.
DR_H2B_UPPER_LANE_CHANGE_TIME	Maximum time in seconds before car will change langes.
DR_H2B_LANE_CHANGE_DISTANCE	Distance in meters how long it will take the car to change lanes.
DR_H2B_NUMBER_OF_BOT_CARS	Number of bot cars on the track.
DR_H2B_MIN_DISTANCE_BETWEEN_BOT_CARS	Minimum distance between bot cars.
DR_H2B_RANDOMIZE_BOT_CAR_LOCATIONS	If True, bot car locations will randomly change after each episode.
DR_H2B_BOT_CAR_SPEED	How fast the bot cars go in meters per second.
DR_CLOUD	Can be azure, aws, local or remote; determines how the storage will be configured.
DR_AWS_APP_REGION	(AWS only) Region for other AWS resources (e.g. Kinesis)
DR_UPLOAD_S3_PROFILE	AWS Cli profile to be used that holds the 'real' S3 credentials needed to upload a model into AWS DeepRacer.
DR_UPLOAD_S3_BUCKET	Name of the AWS DeepRacer bucket where models will be uploaded. (Typically starts with aws-deepracer-.)
DR_LOCAL_S3_PROFILE	Name of AWS profile with credentials to be used. Stored in ~/.aws/credentials unless AWS IAM Roles are used.
DR_GUI_ENABLE	Enable or disable the Gazebo GUI in Robomaker
DR_KINESIS_STREAM_NAME	Kinesis stream name. Used if you actually publish to the AWS KVS service. Leave blank if you do not want this.
DR_KINESIS_STREAM_ENABLE	Enable or disable 'Kinesis Stream', True both publishes to a AWS KVS stream (if name not None), and to the topic /racecar/deepracer/kvs_stream. Leave True if you want to watch the car racing.
DR_SAGEMAKER_IMAGE	Determines which sagemaker image will be used for training.
DR_ROBOMAKER_IMAGE	Determines which robomaker image will be used for training or evaluation.
DR_MINIO_IMAGE	Determines which Minio image will be used.
DR_COACH_IMAGE	Determines which coach image will be used for training.
DR_WORKERS	Number of Robomaker workers to be used for training. See additional documentation for more information about this feature.
DR_ROBOMAKER_MOUNT_LOGS	True to get logs mounted to $DR_DIR/data/logs/robomaker/$DR_LOCAL_S3_MODEL_PREFIX
DR_ROBOMAKER_MOUNT_SIMAPP_DIR	Path to the altered Robomaker bundle, e.g. /home/ubuntu/deepracer-simapp/bundle.
DR_CLOUD_WATCH_ENABLE	Send log files to AWS CloudWatch.
DR_CLOUD_WATCH_LOG_STREAM_PREFIX	Add a prefix to the CloudWatch log stream name.
DR_DOCKER_STYLE	Valid Options are Swarm and Compose. Use Compose for openGL optimized containers.
DR_HOST_X	Uses the host X-windows server, rather than starting one inside of Robomaker. Required for OpenGL images.
DR_WEBVIEWER_PORT	Port for the web-viewer proxy which enables the streaming of all robomaker workers at once.
CUDA_VISIBLE_DEVICES	Used in multi-GPU configurations. See additional documentation for more information about this feature.

Commands

Command	Description
dr-update	Loads in all scripts and environment variables again.
dr-update-env	Loads in all environment variables from system.env and run.env.
dr-upload-custom-files	Uploads changed configuration files from custom_files/ into s3://{DR_LOCAL_S3_BUCKET}/custom_files.
dr-download-custom-files	Downloads changed configuration files from s3://{DR_LOCAL_S3_BUCKET}/custom_files into custom_files/.
dr-start-training	Starts a training session in the local VM based on current configuration.
dr-increment-training	Updates configuration, setting the current model prefix to pretrained, and incrementing a serial.
dr-stop-training	Stops the current local training session. Uploads log files.
dr-start-evaluation	Starts a evaluation session in the local VM based on current configuration.
dr-stop-evaluation	Stops the current local evaluation session. Uploads log files.
dr-start-loganalysis	Starts a Jupyter log-analysis container, available on port 8888.
dr-stop-loganalysis	Stops the Jupyter log-analysis container.
dr-start-viewer	Starts an NGINX proxy to stream all the robomaker streams; accessible remotly.
dr-stop-viewer	Stops the NGINX proxy.
dr-logs-sagemaker	Displays the logs from the running Sagemaker container.
dr-logs-robomaker	Displays the logs from the running Robomaker container.
dr-list-aws-models	Lists the models that are currently stored in your AWS DeepRacer S3 bucket.
dr-set-upload-model	Updates the run.env with the prefix and name of your selected model.
dr-upload-model	Uploads the model defined in DR_LOCAL_S3_MODEL_PREFIX to the AWS DeepRacer S3 prefix defined in DR_UPLOAD_S3_PREFIX
dr-download-model	Downloads a file from a 'real' S3 location into a local prefix of choice.