adept is a reinforcement learning framework designed to accelerate research by abstracting away engineering challenges associated with deep reinforcement learning. adept provides:

• multi-GPU training
• a modular interface for using custom networks, agents, and environments
• baseline reinforcement learning models and algorithms for PyTorch
• built-in tensorboard logging, model saving, reloading, evaluation, and rendering
• proven hyperparameter defaults

This code is early-access, expect rough edges. Interfaces subject to change. We're happy to accept feedback and contributions.

Read More

• Installation
• Quickstart
• Features
• Performance

Documentation

• Architecture Overview
• ModularNetwork Overview
• Resume training
• Evaluate a model
• Render environment

Examples

• Custom Network (stub | example)
• Custom SubModule (stub | example)
• Custom Agent (stub | example)
• Custom Environment (stub | example)

Installation

git clone https://github.com/heronsystems/adeptRL
cd adeptRL
pip install -e .[all]

From docker:

• docker instructions

Quickstart

Train an Agent Logs go to /tmp/adept_logs/ by default. The log directory contains the tensorboard file, saved models, and other metadata.

# Local Mode (A2C)
# We recommend 4GB+ GPU memory, 8GB+ RAM, 4+ Cores
python -m adept.app local --env BeamRiderNoFrameskip-v4

# Distributed Mode (A2C, requires NCCL)
# We recommend 2+ GPUs, 8GB+ GPU memory, 32GB+ RAM, 4+ Cores
python -m adept.app distrib --env BeamRiderNoFrameskip-v4

# IMPALA (requires ray, resource intensive)
# We recommend 2+ GPUs, 8GB+ GPU memory, 32GB+ RAM, 4+ Cores
python -m adept.app actorlearner --env BeamRiderNoFrameskip-v4

# To see a full list of options:
python -m adept.app -h
python -m adept.app help <command>

Use your own Agent, Environment, Network, or SubModule

"""
my_script.py

Train an agent on a single GPU.
"""
from adept.scripts.local import parse_args, main
from adept.network import NetworkModule, SubModule1D
from adept.agent import AgentModule
from adept.env import EnvModule


class MyAgent(AgentModule):
    pass  # Implement


class MyEnv(EnvModule):
    pass  # Implement


class MyNet(NetworkModule):
    pass  # Implement


class MySubModule1D(SubModule1D):
    pass  # Implement


if __name__ == '__main__':
    import adept
    adept.register_agent(MyAgent)
    adept.register_env(MyEnv)
    adept.register_network(MyNet)
    adept.register_submodule(MySubModule1D)
    main(parse_args())

• Call your script like this: python my_script.py --agent MyAgent --env env-id-1 --custom-network MyNet
• You can see all the args here or how to implement the stubs in the examples section above.

Features

Scripts

Local (Single-node, Single-GPU)

• Best place to start if you're trying to understand code.

Distributed (Multi-node, Multi-GPU)

• Uses NCCL backend to all-reduce gradients across GPUs without a parameter server or host process.
• Supports NVLINK and InfiniBand to reduce communication overhead
• InfiniBand untested since we do not have a setup to test on.

Importance Weighted Actor Learner Architectures, IMPALA (Single Node, Multi-GPU)

• Our implementation uses GPU workers rather than CPU workers for forward passes.
• On Atari we achieve ~4k SPS = ~16k FPS with two GPUs and an 8-core CPU.
• "Note that the shallow IMPALA experiment completes training over 200 million frames in less than one hour."
• IMPALA official experiments use 48 cores.
• Ours: 2000 frame / (second * # CPU core) DeepMind: 1157 frame / (second * # CPU core)
• Does not yet support multiple nodes or direct GPU memory transfers.

Agents

• Advantage Actor Critic, A2C (paper | code)
• Actor Critic Vtrace, IMPALA (paper | code)

Networks

• Modular Network Interface: supports arbitrary input and output shapes up to 4D via a SubModule API.
• Stateful networks (ie. LSTMs)
• Batch normalization (paper)

Environments

• OpenAI Gym Atari

Performance

• ~ 3,000 Steps/second = 12,000 FPS (Atari)
  • Local Mode
  • 64 environments
  • GeForce 2080 Ti
  • Ryzen 2700x 8-core
• Used to win a Doom competition (Ben Bell / Marv2in)
• Trained for 50M Steps / 200M Frames
• Up to 30 no-ops at start of each episode
• Evaluated on different seeds than trained on
• Architecture: Four Convs (F=32) followed by an LSTM (F=512)
• Reproduce with python -m adept.app local --logdir ~/local64_benchmark --eval -y --nb-step 50e6 --env <env-id>

Acknowledgements

We borrow pieces of OpenAI's gym and baselines code. We indicate where this is done.