Accidental Classifier

A small CNN that classifies musical accidentals — sharp, flat, and natural — from cropped grayscale images of engraved scores.

Trained on crops extracted from the PrIMuS dataset and achieves 100% accuracy on a held-out test set (750 samples).

Confusion matrix (rows=true, cols=predicted):
               flat  natural    sharp
  flat          250        0        0
  natural         0      250        0
  sharp           0        0      250

Architecture

Input	1 × 40 × 40 grayscale
Backbone	3 conv blocks (16 → 32 → 64 channels), each with BatchNorm + ReLU + MaxPool
Head	Dropout → FC(1600, 64) → ReLU → Dropout → FC(64, 3)
Parameters	~115 k
Output	3-class logits (flat / natural / sharp)

A saved ONNX export is also produced by train.py for inference via OpenCV's dnn module.

Repository layout

train.py                  Training pipeline (CNN + augmentation + weighted sampling)
eval.py                   Evaluate saved checkpoint on the held-out test set
extract_accidentals.py    Extract accidental crops from PrIMuS MEI files via Verovio
extract_fast.py           Faster single-font variant of the extractor
explore_dataset.py        Explore PrIMuS agnostic encodings and image statistics
segment_test.py           Connected-component symbol segmentation experiment
model/
  accidental_classifier.pt   Saved PyTorch weights (best validation accuracy)

Pipeline

1. Extract crops

Render each PrIMuS MEI file with Verovio, locate accidental glyphs (SMuFL codepoints E260/E261/E262) in the SVG, rasterize with cairosvg, and crop each symbol into crops/{flat,natural,sharp}/.

python extract_fast.py

2. Train

Loads all crops into RAM, applies a stratified train/val/test split (seed 42), trains with data augmentation (random affine + erasing) and class-balanced sampling, and saves the best checkpoint.

python train.py

3. Evaluate

Reproduces the identical test split and reports accuracy, per-class metrics, the confusion matrix, and sample misclassifications.

python eval.py

Dependencies

• Python 3.10+
• PyTorch
• torchvision
• NumPy
• Pillow
• OpenCV (cv2) — for ONNX verification and segmentation experiments
• Verovio (Python bindings) — for crop extraction only
• cairosvg — for crop extraction only

Data

Training data is derived from PrIMuS (Calvo-Zaragoza & Rizo, 2018). The extracted crops and raw dataset are not included in this repository due to size.

License

This project is released into the public domain.