DeepSVG

• Library
  • DeepSVG has an excellent library doing SVG things that I’ll totally use! Simplification canonicalization etc. deepsvg/notebooks/svglib.ipynb at master · alexandre01/deepsvg
  • Convert them to Pytorch tensors and then they are gradient descend-able as well!
    • deepsvg/notebooks/svgtensor.ipynb at master · alexandre01/deepsvg
  • They have notebooks for latest space ops torch font generation etc.
  • Animation between user-drawn frames specifically has many nice demo functions based on their main pretrained model: deepsvg/notebooks/animation.ipynb at master · alexandre01/deepsvg
• #plan I can train a letter/not-letter classifier, vanilla RNN, based on different subsets of Unicode! (and put all derived letters in the same split!)
• Then a matrix of symbol/not-symbol similar-not-similar-to-existing

Sim

Shape similarity computation for SVG | Semantic Scholar

Tools

• svgpathtools · PyPI
• JohannesBuchner/imagehash: A Python Perceptual Image Hashing Module clusters on icon dataset

Plan

Basically

1. Font to lines w/ inkscape filter
2. ?? try diff generation options
3. ?? try diff filtration options (’letter-like’ things)
4. ?? try different similarity option (to existing letters)
5. PROFIT

Points

• Train an is-letter classifier w/ pytorch and DeepSVG tensors