Setup¶

Install Dependencies¶

pip install ctoybox 
pip install matplotlib # used in this notebook; not necessary!

Import Packages¶

In [1]:

from ctoybox import Toybox, Input
import numpy as np
import json

from matplotlib.pyplot import imshow, subplots, tight_layout
%matplotlib inline

Inspect Breakout¶

We create an instance of Toybox's Breakout implementation here in a with-block. This block ensures that all memory gets released when finished (in case you don't want to wait for the object to be deleted by python!)

We use matplotlib's imshow to look at the color image of the very first frame. But nothing happens until we press the FIRE button to get a ball.

In [2]:

with Toybox("breakout") as tb:
    print('lives: {}, score: {}, game_over?: {}'.format(
        tb.get_lives(), tb.get_score(), tb.game_over()))
    imshow(tb.get_rgb_frame())

lives: 5, score: 0, game_over?: False

Explore Breakout over time:¶

First, we use the Input struct to create actions fire and noop to send to the game.

Then, we send fire once, and noop 60 times. We save a frame image every 20 frames.

Finally, we use imshow with np.hstack to show all images at once!

In [7]:

# Construct our two actions: fire & nothing
fire = Input()
fire.button1 = True
noop = Input()

# We will collect rendered frames in this variable.
frames = []
with Toybox('breakout') as tb:
    tb.apply_action(fire) # fire to get first ball!
    frames.append(tb.get_rgb_frame())
    
    # see what happens over time!
    for i in range(60):
        tb.apply_action(noop)
        if i % 20 == 0 and i > 0:
            frames.append(tb.get_rgb_frame())

# render the images we collected horizontally:
subplots(figsize=(len(frames)*3, 4))
imshow(np.hstack(frames))
tight_layout() # makes it a little bigger.

What next?¶

Basic interaction with the API allows for the sending of images, and inspection of lives, score, and game_over. This is enough to build a simple gym wrapper if you want to start training RL agents. This core ctoybox package does not provide one because it should not depend on gym.

For more code, checkout our experimentation / main repository: toybox-rs/Toybox which uses a fork of openai/baselines to run some experiments and tests.