dpct-run

Runtime-only package for running saved Deep Perceptual Control Theory (DPCT) individuals.

dpct-run is intended for external users who need to replay/evaluate trained DPCT individuals in Gymnasium environments without access to the full private DPCT evolution and optimization stack.

What is included

• Load saved DHPCTIndividual.config() / best_individual.json files
• Reconstruct the Keras runtime controller
• Run/evaluate individuals in Gymnasium environments
• Optional model summaries, network images, rollout history graphs, and videos
• Optional Comet best_individual.json download helper
• Optional legacy pct config loading/conversion support

What is deliberately not included

• Evolution (DHPCTEvolver)
• Optuna optimization (DHPCTOptimizer)
• Genetic mutation/mating APIs
• DEAP/Optuna dependencies
• Comet experiment logging for evolutionary runs

Install

From PyPI:

pip install dpct-run

Optional extras:

pip install "dpct-run[plots,video,comet,legacy]"

From GitHub:

pip install git+https://github.com/perceptualrobots/dpct-run.git

CLI examples

Show a saved config summary:

dpct-run-individual best_individual.json --show-config

Run a saved individual:

dpct-run-individual best_individual.json --run --steps 500 --seed 42

Save rollout history graphs:

dpct-run-individual best_individual.json --run --history-dir ./history --history-graphs all

Use the short alias:

dpct-run best_individual.json --run --steps 500

Python example

from dpct_run import DHPCTIndividual

config = DHPCTIndividual.load_config("best_individual.json")
individual = DHPCTIndividual.from_config(config)
fitness = individual.evaluate(steps=500, early_termination=True)
print(fitness, individual.success, individual.total_reward)

Compatibility notes

dpct-run follows the saved DPCT config schema produced by the full DPCT package. Treat best_individual.json as the primary interchange artifact.

This initial extraction targets classic Gymnasium-style environments and generic built-in fitness methods (cumulative_reward, evaluation_steps, rms, mae, adjusted RMS/MAE). Environment-specific scoring from dpct-env is optional and only used if that package is separately installed.