Given a grayscale image of a single shorthand form, Gregg Vision can be used to generate its Grascii representation. When combined with Grascii Search, one can obtain possible English interpretations of the shorthand form.

How to Get Started with the Model

Use the code below to get started with the model.

python

from transformers import AutoModelForVision2Seq, AutoImageProcessor, AutoTokenizer
from PIL import Image
import numpy as np


model_id = "grascii/gregg-vision-v0.2.1"
model = AutoModelForVision2Seq.from_pretrained(model_id)
processor = AutoImageProcessor.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)


def generate_grascii(image: Image):
  # convert image to a single channel
  grayscale = image.convert("L")

  # prepare processor input
  images = np.array([grayscale])

  # preprocess image
  pixel_values = processor(images, return_tensors="pt").pixel_values

  # generate token ids
  ids = model.generate(pixel_values, max_new_tokens=12)[0]

  # decode ids and return grascii
  return tokenizer.decode(ids, skip_special_tokens=True)

Note: As of transformers v4.47.0, the model is incompatible with pipeline due to the model's single channel image input.

Technical Details

Model Architecture and Objective

Gregg Vision v0.2.1 is a transformer model with a ViT encoder and a Roberta decoder.

For training, the model was warm-started using vit-small-patch16-224-single-channel for the encoder and a randomly initialized Roberta network for the decoder.

Training Data

Gregg Vision v0.2.1 was trained on the gregg-preanniversary-words dataset.

Training Hardware

Gregg Vision v0.2.1 was trained using 1xT4.