sneo/v802-classifier

v802 classifier

Multi-label image classifier that predicts which of six characters appear in an image: Applejack, Fluttershy, Pinkie Pie, Rainbow Dash, Rarity, Twilight Sparkle.

• Backbone (frozen, not included): DINOv3 ViT-L/16, facebook/dinov3-vitl16-pretrain-lvd1689m
• Head (assets/v802/classifier.joblib): ensemble of a MIL head on patch tokens from layers 16/20/24 (weight 0.6) and a per-class logistic-regression probe (weight 0.4), followed by a label-set decoder with a label-set prior and a cardinality model. The output is a set of labels plus per-class probabilities.
• Input: RGB resized to 640x352, inferred on the original and its horizontal flip. Runs on CPU.

Files

Path	Contents
assets/v802/classifier.joblib	Trained head bundle (classes, components, config, decoder)
src/	Inference, training and experiment code
requirements.txt	Pinned dependency versions used in production

classifier.joblib is a pickle. Loading it executes code, so only load it from a source you trust.

Setup

1. Request access to facebook/dinov3-vitl16-pretrain-lvd1689m on Hugging Face (gated under the DINOv3 license), then download it:

   hf download facebook/dinov3-vitl16-pretrain-lvd1689m --local-dir models/dinov3-vitl16
   

2. Install dependencies (torch is the CPU build):

   pip install --extra-index-url https://download.pytorch.org/whl/cpu -r requirements.txt
   

Inference

python src/predict_v802.py \
    --bundle assets/v802/classifier.joblib \
    --model models/dinov3-vitl16 \
    image1.jpg images_dir/

Options: --threads N (default 4), --fp32 (disable bf16 autocast), --output out.jsonl. Each line of output is JSON: {"file": ..., "labels": [...], "probabilities": {class: p, ...}}.

src/serve_v802.py takes the same --bundle / --model arguments and serves a minimal HTTP endpoint (--port, default 8802).

Notes

• Training and experiment scripts (train_*, experiment_*, extract_features_*, build_v802.py) expect an artifacts/ tree of datasets, splits and features that is not included, so they will not run as-is.
• Training data: 736 labeled dev images plus two 32-image holdout sets (not included).