Click anywhere to close

Retro on Colab

Google Colab

I have previously written about Google Colab, but if you don’t know about it yet: Google Colab is a Jupyter Notebook hosting and colaboration platform provided by Google. It supports Google Docs style interactive sharing, and they provide free shared access to a Nvidia K80 GPU. This makes it an excelent choice for Machine Learning Practitioners on a budget. If you haven’t checked it out I highly suggest doing so.

Gym Retro

OpenAI recently released Gym Retro: a platform that allows you to play over 1,000 games on several different emulators. This allows you to easily test RL Agents accross a wide variaty of games. The biggest problem with RL is the intense amount of compute required to create meaninful agent. This is where Google Colabs free K80 use comes in.

Gym Retro

Google Colab + OpenAI Gym Retro

There are a few steps to get started running retro on Google Colab. The code blocks below are intended to be run in their own jupyter notebook cells.

I have created a public Colab notebook that follows along with these steps. Feel free to clone and play around with it.

Step 1: Install the required dependencies

Colab allows you to run shell commands by prepending them with a !. This allows you to install the required dependencies and packages through apt-get and pip.

First install some apt-get packages that you need to install and run retro.

!apt-get install pkg-config lua5.1 build-essential libav-tools git

Then install gym-retro and some other useful pip dependencies. The second line is installing my personal AI/ML monorepo dumbrain which has some useful tools for running gym-retro.

!pip install tqdm retrowrapper gym-retro
!pip install -U git+git://github.com/frenchie4111/dumbrain.git
Step 2: Install your games

The trickiest part of getting Colab running is getting the games installed for the emulator. In my dumbrain package I have created two different ways to install games. If you wish to use pre-installed retro games, then feel free to skip this step.

The first way is to use colab’s built in file upload tool.

from dumbrain.rl.retro_contest.install_games import colabInstallGames
colabInstallGames()

Upload game zip

This should allow you to upload and install games via a zip-file.

The second way is to download the rom zipfile from a server. I use this to download my roms from my personal fileserver. (If you don’t own the games, please don’t download them from my server)

!python -m dumbrain.rl.retro_contest.install_games http://aiml.mikelyons.org/datasets/sonic/Sonic%20Roms.zip

This will download the zipfile at that url and install any roms found in the package.

Once the games are installed you should be able to list them using:

import retro
retro.list_games()

This list should now include your installed games.

Step 3: Create an env using retrowrapper

Now that we have our games installed we can create a environment that runs the game. I suggest using the retrowrapper package to create your environments. This stops you from getting errors due to accidentally creating multiple environments in the same process.

import retrowrapper

env = retrowrapper.RetroWrapper(
    game='SonicTheHedgehog2-Genesis',
    state='MetropolisZone.Act1'
)

For more information about retrowrapper, see my post about using RetroWrapper to run multiple retro environments

Step 4: Play the game

Once you have your env you can begin stepping through your environment as usual. If you want to visualize the game you can use plt.imshow.

%matplotlib inline
import retrowrapper
import matplotlib.pyplot as plt

observation = env.reset()
for i in range( 100 ): # Take random actions for 100 steps
    random_action = env.action_space.sample()
    observation, reward, done, info = env.step( random_action )
    if done: # If the env is done make sure you reset it
        observation = env.reset()

plt.imshow( observation )

Plt Imshow Output

What Next?

Now that you have retro running you can begin to play around with the current state of the art reinforcement learning algorithms. I suggest taking a look at the retro-baselines provided by OpenAI.

If you are competing in the OpenAI contest consider reading my post about using retrowrapper with custom make functions.

If you are training a model, and want graphical outuput, take a look at my post about Colab + Losswise. Losswise is a great (and free!) platform for visualizing your model’s training.

Recent Posts

rough draft 2: falling in love by being radically honest processing my broken state trying to understand why a past relationship fucked me up so badly
Personal
posted: August 29, 2024
Farewell to Arkansas St a farewell letter to the best apartment I have ever had
Personal
posted: July 01, 2024
A quick teardown of the NES Zapper Lightgun A quick teardown of the NES Zapper Lightgun, mostly as a love letter to the geniuses who built this and a girl I will always love
Technology
posted: April 26, 2024
shouldibreaknocontact.com I have too much money, and I like to buy novelty domains
Personal
posted: March 28, 2024
How to add Currently Playing to your homepage A code snippet to add a last.fm based currently playing section to your personal site
Technology
posted: February 05, 2024