README.md

fastrl

Warning: This is in alpha, and so uses latest torch and torchdata, very importantly torchdata. The base API, while at the point of semi-stability, might be changed in future versions, and so there will be no promises of backward compatiblity. For the time being, it is best to hard-pin versions of the library.

Warning: Even before fastrl==2.0.0, all Models should converge reasonably fast, however HRL models DADS and DIAYN will need re-balancing and some extra features that the respective authors used.

Overview

Fastai for computer vision and tabular learning has been amazing. One would wish that this would be the same for RL. The purpose of this repo is to have a framework that is as easy as possible to start, but also designed for testing new agents.

This version fo fastrl is basically a wrapper around torchdata.

It is built around 4 pipeline concepts (half is from fastai):

• DataLoading/DataBlock pipelines
• Agent pipelines
• Learner pipelines
• Logger plugins

Documentation is being served at https://josiahls.github.io/fastrl/ from documentation directly generated via nbdev in this repo.

Basic DQN example:

from fastrl.loggers.core import *
from fastrl.loggers.vscode_visualizers import  *
from fastrl.agents.dqn.basic import *
from fastrl.agents.dqn.target import *
from fastrl.data.block import *
from fastrl.envs.gym import *
import torch

# Setup Loggers
logger_base = ProgressBarLogger(epoch_on_pipe=EpocherCollector,
                 batch_on_pipe=BatchCollector)

# Setup up the core NN
torch.manual_seed(0)
model = DQN(4,2)
# Setup the Agent
agent = DQNAgent(model,[logger_base],max_steps=10000)
# Setup the DataBlock
block = DataBlock(
    GymTransformBlock(agent=agent,nsteps=2,nskips=2,firstlast=True), # We basically merge 2 steps into 1 and skip. 
    (GymTransformBlock(agent=agent,nsteps=2,nskips=2,firstlast=True,n=100,include_images=True),VSCodeTransformBlock())
)
dls = L(block.dataloaders(['CartPole-v1']*1))
# Setup the Learner
learner = DQNLearner(model,dls,logger_bases=[logger_base],bs=128,max_sz=20_000,nsteps=2,lr=0.001,
                     batches=1000,
                    dp_augmentation_fns=[
                        # Plugin TargetDQN code
                        TargetModelUpdater.insert_dp(),
                        TargetModelQCalc.replace_dp()
                    ])
learner.fit(10)
#learner.validate()

Whats new?

As we have learned how to support as many RL agents as possible, we found that fastrl==1.* was vastly limited in the models that it can support. fastrl==2.* will leverage the nbdev library for better documentation and more relevant testing, and torchdata is the base lib. We also will be building on the work of the ptan¹ library as a close reference for pytorch based reinforcement learning APIs.

¹ “Shmuma/Ptan”. Github, 2020, https://github.com/Shmuma/ptan. Accessed 13 June 2020.

Install

PyPI

Below will install the alpha build of fastrl.

Cuda Install

pip install fastrl==0.0.* --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu113

Cpu Install

pip install fastrl==0.0.* --pre --extra-index-url https://download.pytorch.org/whl/nightly/cpu

Docker (highly recommend)

Install: Nvidia-Docker

Install: docker-compose

docker-compose pull && docker-compose up

Contributing

After you clone this repository, please run nbdev_install_hooks in your terminal. This sets up git hooks, which clean up the notebooks to remove the extraneous stuff stored in the notebooks (e.g. which cells you ran) which causes unnecessary merge conflicts.

Before submitting a PR, check that the local library and notebooks match. The script nbdev_clean can let you know if there is a difference between the local library and the notebooks. * If you made a change to the notebooks in one of the exported cells, you can export it to the library with nbdev_build_lib or make fastai2. * If you made a change to the library, you can export it back to the notebooks with nbdev_update_lib.