Font Detection - Minimal Quickstart

Detect the font of rendered text images. Generate synthetic data with Playwright + Google Fonts, then train a ResNet classifier.

TL;DR

# setup
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m playwright install chromium
# install torch/vision for your platform
pip install torch torchvision  # or CUDA wheels from pytorch.org

# generate dataset into ./data/
python render_phrases.py

# open and run a notebook (choose one)
# - train.ipynb   (ResNet-18, fast)
# - train2.ipynb  (ResNet-50, longer, early stop)

Data layout (expected)

data/
  Inter/            # class folder
  Roboto/
  ...

render_phrases.py creates this automatically using Google Fonts.

Inference (CLI)

# single image
python inference.py --ckpt runs/font_densenet/best.ckpt.pt --image path/to/image.png

# multiple images
python inference.py --ckpt runs/font_densenet/best.ckpt.pt \
  --image img1.png --image img2.png --image img3.png

# options
python inference.py --ckpt ... --image ... --topk 5 --img-size 224 --device auto|cpu|cuda|mps

Notes

• First run downloads a Chromium runtime via Playwright.
• render_phrases.py clears ./data before writing; comment it out to append.
• On restart, load your checkpoint before evaluating (or re-run training).
• ResNet head is fc (not classifier). Keep fc trainable during warmup.
• CUDA AMP warning: use torch.amp.GradScaler('cuda', enabled=...).

License

MIT