TorchBeast

A PyTorch implementation of IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures by Espeholt, Soyer, Munos et al.

TorchBeast comes in two variants: MonoBeast and PolyBeast. While PolyBeast is more powerful (e.g. allowing training across machines), it's somewhat harder to install. MonoBeast requires only Python and PyTorch (we suggest using PyTorch version 1.2 or newer).

For further details, see our paper.

BibTeX

@article{torchbeast2019,
  title={{TorchBeast: A PyTorch Platform for Distributed RL}},
  author={Heinrich K\"{u}ttler and Nantas Nardelli and Thibaut Lavril and Marco Selvatici and Viswanath Sivakumar and Tim Rockt\"{a}schel and Edward Grefenstette},
  year={2019},
  journal={arXiv preprint arXiv:1910.03552},
  url={https://github.com/facebookresearch/torchbeast},
}

Getting started: MonoBeast

MonoBeast is a pure Python + PyTorch implementation of IMPALA.

To set it up, create a new conda environment and install MonoBeast's requirements:

$ conda create -n torchbeast
$ conda activate torchbeast
$ conda install pytorch -c pytorch
$ pip install -r requirements.txt

Then run MonoBeast, e.g. on the Pong Atari environment:

$ python -m torchbeast.monobeast --env PongNoFrameskip-v4

By default, MonoBeast uses only a few actors (each with their instance of the environment). Let's change the default settings (try this on a beefy machine!):

$ python -m torchbeast.monobeast \
     --env PongNoFrameskip-v4 \
     --num_actors 45 \
     --total_steps 30000000 \
     --learning_rate 0.0004 \
     --epsilon 0.01 \
     --entropy_cost 0.01 \
     --batch_size 4 \
     --unroll_length 80 \
     --num_buffers 60 \
     --num_threads 4 \
     --xpid example

Results are logged to ~/logs/torchbeast/latest and a checkpoint file is written to ~/logs/torchbeast/latest/model.tar.

Once training finished, we can test performance on a few episodes:

$ python -m torchbeast.monobeast \
     --env PongNoFrameskip-v4 \
     --mode test \
     --xpid example

MonoBeast is a simple, single-machine version of IMPALA. Each actor runs in a separate process with its dedicated instance of the environment and runs the PyTorch model on the CPU to create actions. The resulting rollout trajectories (environment-agent interactions) are sent to the learner. In the main process, the learner consumes these rollouts and uses them to update the model's weights.

Faster version: PolyBeast

PolyBeast provides a faster and more scalable implementation of IMPALA.

The easiest way to build and install all of PolyBeast's dependencies and run it is to use Docker:

$ docker build -t torchbeast .
$ docker run --name torchbeast torchbeast

To run PolyBeast directly on Linux or MacOS, follow this guide.

Installing PolyBeast

Linux

Create a new Conda environment, and install PolyBeast's requirements:

$ conda create -n torchbeast python=3.7
$ conda activate torchbeast
$ pip install -r requirements.txt

Install PyTorch either from source or as per its website (select Conda).

PolyBeast also requires gRPC and other third-party software, which can be installed by running:

$ git submodule update --init --recursive

Finally, let's compile the C++ parts of PolyBeast:

$ pip install nest/
$ python setup.py install

MacOS

Create a new Conda environment, and install PolyBeast's requirements:

$ conda create -n torchbeast
$ conda activate torchbeast
$ pip install -r requirements.txt

PyTorch can be installed as per its website (select Conda).

PolyBeast also requires gRPC and other third-party software, which can be installed by running:

$ git submodule update --init --recursive

Finally, let's compile the C++ parts of PolyBeast:

$ pip install nest/
$ python setup.py install

Running PolyBeast

To start both the environment servers and the learner process, run

$ python -m torchbeast.polybeast

The environment servers and the learner process can also be started separately:

python -m torchbeast.polybeast_env --num_servers 10

Start another terminal and run:

$ python3 -m torchbeast.polybeast_learner

(Very rough) overview of the system

|-----------------|     |-----------------|                  |-----------------|
|     ACTOR 1     |     |     ACTOR 2     |                  |     ACTOR n     |
|-------|         |     |-------|         |                  |-------|         |
|       |  .......|     |       |  .......|     .   .   .    |       |  .......|
|  Env  |<-.Model.|     |  Env  |<-.Model.|                  |  Env  |<-.Model.|
|       |->.......|     |       |->.......|                  |       |->.......|
|-----------------|     |-----------------|                  |-----------------|
   ^     I                 ^     I                              ^     I
   |     I                 |     I                              |     I Actors
   |     I rollout         |     I rollout               weights|     I send
   |     I                 |     I                     /--------/     I rollouts
   |     I          weights|     I                     |              I (frames,
   |     I                 |     I                     |              I  actions
   |     I                 |     v                     |              I  etc)
   |     L=======>|--------------------------------------|<===========J
   |              |.........      LEARNER                |
   \--------------|..Model.. Consumes rollouts, updates  |
     Learner      |.........       model weights         |
      sends       |--------------------------------------|
     weights

The system has two main components, actors and a learner.

Actors generate rollouts (tensors from a number of steps of environment-agent interactions, including environment frames, agent actions and policy logits, and other data).

The learner consumes that experience, computes a loss and updates the weights. The new weights are then propagated to the actors.

Learning curves on Atari

We ran TorchBeast on Atari, using the same hyperparamaters and neural network as in the IMPALA paper. For comparison, we also ran the open source TensorFlow implementation of IMPALA, using the same environment preprocessing. The results are equivalent; see our paper for details.

Repository contents

libtorchbeast: C++ library that allows efficient learner-actor communication via queueing and batching mechanisms. Some functions are exported to Python using pybind11. For PolyBeast only.

nest: C++ library that allows to manipulate complex nested structures. Some functions are exported to Python using pybind11.

tests: Collection of python tests.

third_party: Collection of third-party dependencies as Git submodules. Includes gRPC.

torchbeast: Contains monobeast.py, and polybeast.py, polybeast_learner.py and polybeast_env.py.

Hyperparamaters

Both MonoBeast and PolyBeast have flags and hyperparameters. To describe a few of them:

num_actors: The number of actors (and environment instances). The optimal number of actors depends on the capabilities of the machine (e.g. you would not have 100 actors on your laptop). In default PolyBeast this should match the number of servers started.
batch_size: Determines the size of the learner inputs.
unroll_length: Length of a rollout (i.e., number of steps that an actor has to be perform before sending its experience to the learner). Note that every batch will have dimensions [unroll_length, batch_size, ...].

Contributing

We would love to have you contribute to TorchBeast or use it for your research. See the CONTRIBUTING.md file for how to help out.

License

TorchBeast is released under the Apache 2.0 license.

tesslerc/torchbeast