Imperssonator / fairplay

Fairplay

Fairplay is a library built with the goal of enabling fully automated extraction of data from plots. One of the primary components is a plot simulator which can be used to generate training images for various tasks.

Installation

Clone the repository. Then, from within fairplay/:

pip install -e ./

Dependencies will be installed automatically. Python >= 3.9 recommended.

Usage

python ./src/fairplay/gen/generate_random_scatter.py ./data/demo -n 20 -t 10

Arguments

• ./data/demo: directory to build image dataset and corresponding labeled images
• -n 20: 20 training images
• -t 10: 10 test images

Simulation Config

Simulation parameters are configured in data/plot_params. There are two key files:

• continuous.csv: parameters to define a truncated normal distribution from which to sample continuously-defined plotting arguments, like marker_alpha. Attributes are:
  • min: lowest allowable value
  • max: highest allowable value
  • mean: if specified, can place the mean of the truncated normal somewhere other than the midpoint of min and max, which is the default
  • n_stds: number of SD's of the normal distribution to "fit" between min and max. Default is 1, and if mean is also unspecified, this means the amplitude of the truncated normal PDF at min and max will be that of -1 and +1 SD. Higher values will result in more concentrated sampling near the mean and less at the edges.
• discrete.csv: literally-defined Lists on which to perform uniform sampling of discretely-defined plotting arguments, such as marker_style. To weight a member more heavily, simply add more copies of that member to the list. Very rudimentary.

Example Output

RGB values for class labels (e.g. x ticks, markers, background) are defined as label_colors in generate_random_scatter.py

Simulated	Labeled