This folder contains the code and data used in the glyph-alignment pretraining stage. This codebase is developed based on OpenCLIP.
Note: Currently, we release a subset of fonts that we use (see assets/fonts) folder, containing 100 with free commercial use license.
sudo docker pull pytorch/pytorch:2.2.0-cuda12.1-cudnn8-devel
pip install -e .
Glyph-ByT5 requires 4xA100 GPUs for training. An example training script is provided in here.
Run the code:
bash scripts/train_glyph_byt5.sh