adept is a library designed to accelerate reinforcement learning research by providing:
- baseline reinforcement learning models and algorithms for PyTorch
- multi-GPU compute options
- access to various environments
- built-in tensorboard logging, model saving, reloading, evaluation, and rendering
- abstractions for building custom networks, agents, execution modes, and experiments
- proven hyperparameter defaults
This code is alpha, expect rough edges.
Agents / Networks
- Actor Critic with Generalized Advantage Estimation
- Stateful networks (ie. LSTMs)
- Batch norm
Execution Modes
- Local (Single-GPU, A2C)
- Towered (Multi-GPU, A3C-variant)
- Importance Weighted Actor Learner Architectures, IMPALA (Faster Multi-GPU)
Environments
- OpenAI Gym
- StarCraft 2 (alpha, impala mode does not work with SC2 yet)
We designed this library to be flexible and extensible. Plugging in novel research ideas should be doable.
- gym
- PyTorch 1.x
- Python 3.5+
- We use CUDA 10, pytorch 1.0, python 3.6
From docker:
From source:
- Follow instructions for PyTorch
- (Optional) Follow instructions for StarCraft 2
git clone https://github.com/heronsystems/adeptRL
cd adeptRL
# Remove mpi, sc2, profiler if you don't plan on using these features:
pip install .[mpi,sc2,profiler]
- Atari/SC2 scores pending
- ~ 3000 TFPS / 12000 FPS (Local Mode / 64 environments / GeForce 2080 Ti / Ryzen 2700x 8-core)
- Used to win a Doom competition (Ben Bell / Marv2in)
If you write your own scripts, you can provide your own agents or networks, but
we have some presets you can run out of the box. 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
# Towered Mode (A3C Variant, requires mpi4py)
# We recommend 2+ GPUs, 8GB+ GPU memory, 32GB+ RAM, 4+ Cores
python -m adept.app towered --env BeamRiderNoFrameskip-v4
# IMPALA (requires mpi4py and is resource intensive)
# We recommend 2+ GPUs, 8GB+ GPU memory, 32GB+ RAM, 4+ Cores
python -m adept.app impala --env BeamRiderNoFrameskip-v4
# StarCraft 2 (IMPALA not supported yet)
# Warning: much more resource intensive than Atari
python -m adept.app local --env CollectMineralShards
# To see a full list of options:
python -m adept.app -h
python -m adept.app help <command>
- Training Frame - An environment frame that is trained on
- Frame - A raw environment frame, including skipped frames. A skipped frame is a frame that is not trained on. By default, Atari only trains on every fourth frame.
- FPS - Frames per second.
- TFPS - Training frames per second. Typically research papers report raw environment frames. Our logs report training frames, since we are interested in sample efficiency.
An Agent acts on and observes the environment. Currently only ActorCritic is supported. Other agents, such as DQN or ACER may be added later.
Containers hold all of the application state. Each subprocess gets a container in Towered and IMPALA modes.
Environments run in subprocesses and send their observation, rewards, terminals, and infos to the host process. They work pretty much the same way as OpenAI's code.
An Experience Cache is a Rollout or Experience Replay that is written to after stepping and read before learning.
Modules are generally useful PyTorch modules used in Networks.
Networks are not PyTorch modules, they need to implement our abstract NetworkInterface or ModularNetwork classes. A ModularNetwork consists of a trunk, body, and head. The Trunk can consist of multiple networks for vision or discrete data. It flattens these into an embedding. The Body network operates on the flattened embedding and would typically be an LSTM, Linear layer, or a combination. The Head depends on the Environment and Agent and is created accordingly.
We borrow pieces of OpenAI's gym and baselines code. We indicate where this is done.