Stimuli

Single images

from corread import generate_image

img = generate_image('house', 'DejaVuSans', size=60, xshift=-20, upper=True)   # a PIL image
img.save('HOUSE.png')

The text is drawn in black on a white canvas (500 × 500 by default), centred and then shifted by (xshift, yshift) pixels. mirror flips the image left to right and invert reverses the order of the letters.

Word sets

A word set holds num_train training and num_val validation images of each word, in ImageNet-style folders train/<word>/ and val/<word>/:

import corread

corread.generate_training_set(
    'english',                       # its name: the folder in the wordsets folder
    'english',                       # the wordlist, wordlists/english.txt
    train_fonts=['DejaVuSans', 'LiberationSerif'],
    val_fonts=['NotoSans'],          # a font the training images never use
    num_words=1000, num_train=1000, num_val=50,
    seed=1, workers=16,
)

Each image draws, uniformly at random, its font (from the fonts of its split), its size (40, 50, 60, 70 or 80), whether it is uppercase, and a shift of up to ±50 px horizontally and ±30 px vertically.

Reproducibility

All random choices are drawn up front, from one seed, in a fixed order, so a set depends only on its seed, never on the number of worker processes. Next to the images, the set records its recipe:

File	Holds
metadata.json	the arguments, the seed used, the rendering settings, the words, the sha256 of the wordlist and of every font, and the versions of CORread, Python and Pillow
images.csv	the font, size, case and shift drawn for every image
checksums.json	the sha256 digest of every word folder, and one over the whole set

regenerate_training_set draws a set again from its recipe. It checks the fonts and wordlist against the recorded hashes, the draws against images.csv before drawing anything, and every word folder against checksums.json afterwards:

corread.regenerate_training_set('english', recipe='path/to/english', wordsetsdir='elsewhere')

The images are byte-identical as long as Pillow is the same version: Pillow bundles the FreeType and libjpeg that render the text and encode the JPEGs. With another version the check names the word folders that differ; allow_mismatch=True keeps the images anyway and records verified: false in regeneration.json.

Trained models record the sha256 of their word set's metadata.json and the digest of its images, so every model can be traced back to the exact images it saw. Recipes can be shared and regenerated through the Hugging Face Hub.