yawning_titan.yawning_titan_run.YawningTitanRun#

class yawning_titan.yawning_titan_run.YawningTitanRun(network=None, game_mode=None, red_agent_class=<class 'yawning_titan.envs.generic.core.red_interface.RedInterface'>, blue_agent_class=<class 'yawning_titan.envs.generic.core.blue_interface.BlueInterface'>, print_metrics=False, show_metrics_every=1, collect_additional_per_ts_data=False, eval_freq=10000, total_timesteps=200000, training_runs=1, n_eval_episodes=1, deterministic=False, warn=True, render=False, verbose=1, logger=None, output_dir=None, auto=True, **kwargs)[source]#

Bases: object

The YawningTitanRun class is the run class for training YT agents from a given set of parameters.

The YawningTitanRun class can be used ‘straight out of the box’, as all params have default values.

yt_run = YawningTitanRun()

The YawningTitanRun class can also be used manually by setting auto=False.

yt_run = YawningTitanRun(auto=False)
yt_run.setup()
yt_run.train()
yt_run.evaluate()

Trained agents can be saved by calling .save(). If no path is provided, a path is generated using the AGENTS_DIR, today’s date, and the uuid of the instance of YawningTitanRun.

yt_run = YawningTitanRun()
yt_run.save()

Todo

  • Build a reporting functionality that captures all logs and eval and generates a PDF report.

  • Add multiple training runs functionality for the same agent.

  • Add the ability to load a saved agent and continue training it.

The YawningTitanRun constructor.

# TODO: Add proper Sphinx mapping for classes/methods.

Parameters:

  • network – An instance of Network.

  • game_mode – An instance of GameMode.

  • red_agent_class – The agent/action set class used for the red agent.

  • blue_agent_class – The agent/action set class used for the blue agent.

  • print_metrics – Print the metrics if True. Default value = True.

  • show_metrics_every – Prints the metrics every show_metrics_every time steps. Default value = 10.

  • collect_additional_per_ts_data – Collects additional per-timestep data if True.Default value = False.

  • eval_freq – Evaluate the agent every eval_freq call of the callback. Default value = 10,000.

  • total_timesteps – The number of samples (env steps) to train on. Default value = 200000.

  • training_runs – The number of times the agent is trained.

  • n_eval_episodes – The number of episodes to evaluate the agent. Default value = 1.

  • deterministic – Whether the evaluation should use stochastic or deterministic actions. Default value = False.

  • warn – Output additional warnings mainly related to the interaction with stable_baselines if True. Default value = True.

  • render – Renders the environment during evaluation if True. Default value = False.

  • verbose – Verbosity level: 0 for no output, 1 for info messages (such as device or wrappers used), 2 for debug messages. Default value = 1.

  • logger – An optional custom logger to override the use of the default module logger.

  • output_dir – An optional output path for eval output and saved agent zip file. If none is provided, a path is generated using the yawning_titan.AGENTS_DIR, today’s date, and the uuid of the instance of YawningTitanRun.

  • auto – If True, setup(), train(), and evaluate() are called automatically.

Methods

evaluate

Evaluates the trained agent.

export

Export the YawningTitanRun as a zip.

import_from_export

Import and return an exported YawningTitanRun.

load

Load and return a saved YawningTitanRun.

save

Saves the trained agent using the stable_baselines3 save as zip functionality.

setup

Performs a setup of the NetworkInterface, GenericNetworkEnv, PPO algorithm.

train

Trains the agent.
setup(new=True, ppo_zip_path=None)[source]#

Performs a setup of the NetworkInterface, GenericNetworkEnv, PPO algorithm.

The setup needs to be performed before training can occur.

Parameters:

  • new – If True, a new instance of PPO is generated. If False, a ppo_zip_path must be passed tooo.

  • ppo_zip_path – Optional path to a saved ppo.zip file. Required if new = False.

Raises:

AttributeError – When new=False and ppo_zip_path hasn’t been provided.

train()[source]#

Trains the agent.

Returns:

The trained instance of stable_baselines3.ppo.ppo.PPO.

evaluate()[source]#

Evaluates the trained agent.

Returns:

Mean reward per episode, std of reward per episode.

save()[source]#

Saves the trained agent using the stable_baselines3 save as zip functionality.

The instance of PPO is saved to ppo.zip.

The YawningTitanRun args are saved to args.json.

The YawningTitanRun.uuid is saved to UUID.

Returns:

The path the agent has been saved to.

export()[source]#

Export the YawningTitanRun as a zip.

The contents of output_dir is archived to the agents_dir exported dir.

Included is an INVENTORY file that contains all files and their sizes. This is used for file verification when an exported YawningTitanRun is imported.

Returns:

The exported filepath as a str.

classmethod load(path)[source]#

Load and return a saved YawningTitanRun.

YawningTitanRun’s that have auto=True will not be automatically ran on load.

Parameters:

path – A saved YawningTitanRun path.

Returns:

An instance of YawningTitanRun.

classmethod import_from_export(exported_zip_file_path, overwrite_existing=False)[source]#

Import and return an exported YawningTitanRun.

YawningTitanRun’s that have auto=True will not be automatically ran on import.

Parameters:

  • exported_zip_file_path – The path of an exported YawningTitanRun.

  • overwrite_existing – If True, if the uuid of the imported agent already exists in the trainer agents dir it is overwritten.

Returns:

The imported instance of YawningTitanRun.

Raises:

YawningTitanRunError – When the INVENTORY file fails its verification.