Skip to content

Files

Latest commit

 

History

History

custom-files

README.md

Hyperparameter Values

  • batch_size: 34, 64, 128, 256 or 512. Default is 64.
  • beta_entropy: Real number between 0 and 1 (inclusive). Default is 0.01.
  • discount_factor: Real number between 0 and 1 (inclusive). Default is 0.999.
  • e_greedy_value: Default is 0.05.
  • epsilon_steps: Default is 10000.
  • exploration_type: categorical (for clipped_ppo training algorithm) or additive_noise (for sac training algorithm)
  • loss_type: huber or mean squared error. Default is huber.
  • lr: Values between 0.00000001 and 0.001 (inclusive). Default is 0.0003.
  • num_episodes_between_training: Integer between 1 and 100 (inclusive). To avoid issues with the position from which evaluations are run ensure that ( num_episodes_between_training / DR_WORKERS) * DR_TRAIN_ROUND_ROBIN_ADVANCE_DIST = 1.0. As default DR_WORKERS is 2 and DR_TRAIN_ROUND_ROBIN_ADVANCE_DIST is 0.05 the default value is 40.
  • num_epochs: Values between 3 and 10 (inclusive). Default is 3.
  • stack_size: Default is 1.
  • term_cond_avg_score: Values between 35000.0 and 100000.0.
  • term_cond_max_episodes: Default is 100000.

Model Metadata Values

  • action_space_type: continuous or discrete.
  • action_space: See model_metadata_sac.json for a continuous example or model_metadata.json for a discrete example.
  • version: 5
  • training_algorithm: sac (for additive_noise exploration_type) or clipped_ppo (for categorical exploration_type)
  • neural_network: DEEP_CONVOLUTIONAL_NETWORK_SHALLOW or DEEP_CONVOLUTIONAL_NETWORK
  • sensor: a list containing FRONT_FACING_CAMERA or STEREO_CAMERAS and optionally LIDAR

reward function

  • Populate this file with your code examples can be found online, e.g. AWS examples

run.env values

  • Consult DeepRacer for Cloud Documentation
  • DR_REGULAR_UPLOAD: DeepRacer on the Spot specific var. Integer defining the number of minutes between regular uploads to your upload s3 location to get model checkpoints throughout your training. Default is 0 (disabled). WARNING - if you turn this setting on you'll be storing an additional ~75MB of model files in the upload folder at every internal of regular upload you define, e.g. 1.8GB over 24 hours with an interval of 60 minutes. You may want to delete / tidy up after training finishes and you've kept the optimal checkpoints.
  • DR_REGULAR_PHYSICAL_MODEL_UPLOAD: DeepRacer on the Spot specific var. Integer defining the number of minutes between regular uploads of the physical car model to your upload s3 location to get model checkpoints throughout your training. Default is 0 (disabled). WARNING - if you turn this setting on you'll be storing an additional ~20MB of model files in the upload folder at every internal of regular upload you define, e.g. ~0.5GB over 24 hours with an interval of 60 minutes. You may want to delete / tidy up after training finishes and you've kept the optimal checkpoints.
  • DR_CONTINUE_ON_SPOT_INTERRUPTION: DeepRacer on the Spot specific var. Boolean defining whether to continue training or not after a spot interruption. Only has effect when using the create-spot-instance.sh script.
  • DR_IMPORT_MODEL_ON_COMPLETION: DeepRacer on the Spot specific var. Boolean defining whether to upload the model to the DeepRacer Console at the end of training or not. Default is to upload it, set the variable to False to disable this.

system.env values

The following should be changed to enable OpenGL training with GPUs, to reduce CPU load: -

  • Set DR_GUI_ENABLE=True
  • Set DR_DOCKER_STYLE=compose
  • Set DR_HOST_X=True
  • Uncomment out DR_SAGEMAKER_CUDA_DEVICES=0
  • Uncomment out DR_ROBOMAKER_CUDA_DEVICES=0
  • Uncomment out DR_DISPLAY=:99

The following should be changed for SAC training_algorithm as it cannot use multiple workers: -

  • set DR_WORKERS=1