drfc_manager

A Pythonic workflow manager and wrapper for DeepRacer for Cloud (DRfC)

Table of Contents

1. Objective
2. Key Features
3. Installation
4. Usage
5. Advanced Usage
6. Troubleshooting
7. Idea behind
8. Contributing

Objective

The main purpose of this library is to provide a Pythonic, Jupyter-friendly interface to manage your workflow within the DeepRacer for Cloud (DRfC) environment.

This library allows users to optimize the training, evaluation, and management of Reinforcement Learning (RL) models by orchestrating the entire process from Python scripts or Jupyter Notebooks. It supports local, MinIO, and AWS S3 storage, and is designed for multi-user environments (e.g., JupyterHub).

Key Features

• Easy model configuration for training (hyperparameters, model metadata, reward function)
• Pipeline management for training, evaluation, cloning, stopping, and metrics
• Multi-user support: user-specific temp/log directories, safe for JupyterHub
• Local, MinIO, and AWS S3 storage support
• Automatic Docker Compose orchestration for all flows
• Jupyter Notebook integration: run, monitor, and stop jobs from notebooks
• Advanced logging: per-user, per-run logs for debugging and reproducibility
• Extensible: add new pipeline steps or customize existing ones
• Integrated viewer pipeline: Launches a real-time Streamlit-based viewer and video stream proxy for model evaluation and monitoring.

Installation

pip install drfc_manager
# or clone and install locally
# git clone https://github.com/joaocarvoli/drfc-manager.git
# cd drfc-manager && pip install .

Usage

1. Define configuration model data

from drfc_manager.types.hyperparameters import HyperParameters
from drfc_manager.types.model_metadata import ModelMetadata

model_name = 'rl-deepracer-sagemaker'
hyperparameters = HyperParameters() 
model_metadata = ModelMetadata()

2. Define the reward function

def reward_function(params):
    # Your custom reward logic here
    return float(...)

3. Run a training pipeline

from drfc_manager.pipelines import train_pipeline

train_pipeline(
    model_name=model_name,
    hyperparameters=hyperparameters,
    model_metadata=model_metadata,
    reward_function=reward_function,
    overwrite=True,
    quiet=False
)

4. Evaluate a model

from drfc_manager.pipelines import evaluate_pipeline

result = evaluate_pipeline(
    model_name=model_name,
    run_id=0,
    quiet=True,
    clone=False,
    save_mp4=True
)

5. Clone a model

from drfc_manager.pipelines import clone_pipeline

clone_pipeline(
    model_name=model_name,
    new_model_name='my-cloned-model',
    quiet=True
)

6. Stop a running pipeline

from drfc_manager.pipelines import stop_training_pipeline, stop_evaluation_pipeline

stop_training_pipeline(run_id=0)
stop_evaluation_pipeline(run_id=0)

7. Start/Stop metrics (Grafana/Prometheus)

from drfc_manager.pipelines import start_metrics_pipeline, stop_metrics_pipeline

start_metrics_pipeline(run_id=0)
stop_metrics_pipeline(run_id=0)

8. Start/Stop the viewer

from drfc_manager.pipelines import start_viewer_pipeline, stop_viewer_pipeline

# Start the viewer (for a given run_id)
viewer_result = start_viewer_pipeline(run_id=0, quiet=True)

# Stop the viewer
stop_viewer_pipeline(quiet=True)

Troubleshooting

• Docker Compose errors: Make sure Docker is running and your user has permission to run Docker commands.
• Multi-user issues: Each user gets their own temp/log directory. If you see permission errors, check directory ownership and permissions.

Idea behind

This lib is developed using the same ideas and implementation as the aws-deepracer-community/deepracer-for-cloud repo: "A quick and easy way to get up and running with a DeepRacer training environment using a cloud virtual machine or a local computer".

Contributing

We welcome contributions! Please see our CONTRIBUTING.md for guidelines and instructions.

---

For more examples and advanced configuration, see the examples/ directory.