Skip to content

RUNTIME_ASSETS.md

Latest commit

 

History

History
168 lines (124 loc) · 5.8 KB

RUNTIME_ASSETS.md

File metadata and controls

168 lines (124 loc) · 5.8 KB
Raw

Runtime Assets And Training Artifacts

This repo is intended to run inference from a fresh clone, but not every local training artifact is checked in.

Included

These files are included because they are required for normal inference:

File Needed for Notes
auto_icon_vectorizer/runtime/nn-seg-results/best-gated-unet.pt stroke/outline icon mask inference Included model checkpoint.
auto_icon_vectorizer/runtime/nn-seg-results/best-filled-silhouette-unet.pt filled icon mask inference Included model checkpoint.
auto_icon_vectorizer/runtime/package.json and package-lock.json installing the Node Potrace wrapper Run python3 -m auto_icon_vectorizer.install_runtime after cloning.
auto_icon_vectorizer/runtime/*.py feature extraction, mask cleanup, inference, tracing Included runtime code.

Not Included

These are intentionally excluded:

Artifact Why it is not included How to regenerate
auto_icon_vectorizer/runtime/node_modules/ generated dependency folder python3 -m auto_icon_vectorizer.install_runtime
auto_icon_vectorizer/runtime/nn-seg-results/feature-cache-v2.npz large generated stroke training feature cache generated automatically by train_aux_fusion_icon_segmenter.py
auto_icon_vectorizer/runtime/nn-seg-results/filled-silhouette-feature-cache-v1.npz large generated filled training feature cache generated automatically by train_filled_silhouette_segmenter.py
auto_icon_vectorizer/runtime/truth-stress-eval/ local stroke-training corpus used during development generate a public synthetic replacement with scripts/generate_public_training_corpus.py
auto_icon_vectorizer/runtime/nn-training-corpus-v2/ optional extra local stroke-training corpus optional; use your own labeled reports if needed
diagnostic sheets and TSV summaries generated during local experiments rerun the training/eval scripts

The feature cache files are not needed for inference. They only speed up retraining by avoiding repeated feature extraction.

Fresh Clone Inference Setup

git clone https://github.com/jaydenbarnescs-tech/auto-icon-vectorizer.git
cd auto-icon-vectorizer
python3 -m pip install -e .
python3 -m auto_icon_vectorizer.install_runtime
python3 -m auto_icon_vectorizer.doctor

If the doctor command reports a missing native Cairo dependency, install Cairo for your platform:

# macOS
brew install cairo

# Debian/Ubuntu
sudo apt-get install libcairo2

Then run a smoke test:

auto-icon-vectorizer examples/sample-ai-icon-crop.png \
  --out-prefix out/sample-icon \
  --json out/sample-icon.json

The command prints SVG to stdout and writes out/sample-icon.svg, out/sample-icon-mask.png, out/sample-icon-source.png, and diagnostics JSON.

Regenerating Training Artifacts

Normal users do not need this. Use these commands only if you want to retrain or modify the neural networks.

Training data should match the output contract used by the vectorizer:

  • one cropped icon per source image
  • one dominant foreground color for the whole icon
  • one binary foreground mask in the truth alpha channel
  • varied backgrounds that represent the generated UI images, screenshots, or mockups where the model will run

The models learn foreground/background segmentation. They do not learn separate SVG layers for multicolor logos.

Stroke / Outline Branch

The stroke branch expects a report file at:

auto_icon_vectorizer/runtime/truth-stress-eval/latest-run.json

Each report points to:

  • sourceCrop: RGB crop image
  • truthIcon: RGBA image where alpha marks the icon mask
  • id, icon, backgroundMode: metadata used for splits and diagnostics

Generate a public synthetic corpus:

python3 scripts/generate_public_training_corpus.py --count 240

That generator draws outline icons with one foreground color over varied background families, matching the single-color icon assumption.

Then retrain:

cd auto_icon_vectorizer/runtime
python3 train_aux_fusion_icon_segmenter.py --force-features

This creates:

  • nn-seg-results/feature-cache-v2.npz
  • nn-seg-results/best-gated-unet.pt
  • nn-seg-results/training-history.json
  • diagnostic PNG/TSV files

Filled / Silhouette Branch

The filled branch has its own synthetic generator built into the training script. It does not require truth-stress-eval. It also draws one foreground color per generated filled icon, including cutout and same-color hybrid-style cases.

cd auto_icon_vectorizer/runtime
python3 train_filled_silhouette_segmenter.py --force-features

This creates:

  • nn-seg-results/filled-silhouette-feature-cache-v1.npz
  • nn-seg-results/best-filled-silhouette-unet.pt
  • nn-seg-results/filled-silhouette-training-history.json

If You Have Your Own Training Data

Create auto_icon_vectorizer/runtime/truth-stress-eval/latest-run.json with this shape:

{
  "reports": [
    {
      "id": "s01-000",
      "icon": "check-circle",
      "backgroundMode": "ai-noisy",
      "sourceCrop": "/absolute/path/to/source.png",
      "truthIcon": "/absolute/path/to/truth-rgba.png"
    }
  ]
}

truthIcon should be an RGBA image. The alpha channel is treated as the ground truth mask. Use one alpha mask for the whole icon foreground; do not encode separate color layers.

What An AI Agent Should Do First

  1. Run python3 -m auto_icon_vectorizer.doctor.
  2. If node_modules/potrace is missing, run python3 -m auto_icon_vectorizer.install_runtime.
  3. Do not look for .npz feature caches unless retraining is requested.
  4. Do not retrain unless the included .pt checkpoints are missing or the user explicitly wants a new model.
  5. For retraining, generate or provide truth-stress-eval/latest-run.json before running train_aux_fusion_icon_segmenter.py.