texrelenv 0.1.0

A Python package for generating data sets based on Hugh Perkins' TexRel (https://arxiv.org/abs/2105.12804)

texrelenv: TexRel-like environments for emergent language experiments

A Python package for generating data sets based on Hugh Perkins' TexRel.

Getting started

You can install this package from pypi using

pip install texrelenv

See example usage below for how to use the package to generate data.

Example usage

Here is an example of how to generate 100 images based on some config options specified in a config.toml file.

Let's say you have a config.toml file that looks like the below. "grid" refers to the overall canvas of the image while "things" refers to the coloured objects that appear in images. The config options shown below are exhausted and are explained in the (Config options)[#Config options] section below

[things]
distinct_shapes = 9
distinct_colours = 9

[environment]
grid_size = 16
thing_size = 4
things_per_image = 5
hard_boundary = true
objects_can_overlap = false

[split]
hold_out_things = 0.2
hold_out_images = 0.0

A dataset can be created from the above config in the following way

from texrelenv import DataSet

data = DataSet(config_file="config.toml")

You could also pass all the config arguments as arguments when you instantiate the DataSet (they are all named the same).

Having instantiated the data set, you can generate some train images like this:

data.sample(100, 'train')

Or some test images like this:

data.sample(100, 'test')

Config options

Option	Type	Meaning
grid_size	int	The size of the square images generated, expressed as the length of the side of the square in pixels. Default is 16, for an easy drop-in replacement for MNIST-like data sets, use 28.
hard_boundary	bool	If true, shapes added to an image will always be fully visible. If false, they may partially be out of frame.
objects_can_overlap	bool	Can objects go in front of or behind one another? Default is false.
thing_size	int	How big are the coloured objects added to images? Expressed as the length of one side of the square canvas used to create the templates for these objects, which can also be thought of as the maximum size of an object.
distinct_shapes	int
distinct_colours	int	blah
things_per_image	int	blah
rotate	bool	blah
flip	bool	blah
hold_out_things	float	blah
hold_out_images	float	blah

Contributing

To contribute to the project, please make sure you have poetry installed and before you start working on the code, set up a virtual environment, activate it and then run

poetry install

to install all dependencies including dependencies only required for development. Then update precommit hooks with

pre-commit install
pre-commit autoupdate

Pre-commit will help you to keep your code style in keeping with the rest of the project.

Please write tests for your code, store them in test/, and only commit code that passes the unit tests that already exist. You can run unit tests with

python -m unittest

(precommit will not run unit tests!)

Acknowledgements

We are hugely thankful to Hugh Perkins for coming up with the idea of TexRel in the first place, for creating the first version of PyTorch, and for doing various other cool stuff.