morpeN1/autolens-dinov3-safe-focal

AutoLens DINOv3 Safe Focal

Pre-calibration HF model card for the focal-loss DINOv3 candidate.

Model

• Architecture: DINOv3 ViT-Small/16 (timm: vit_small_patch16_dinov3)
• Classes: 8 vehicle body types
• Source run: dinov3_safe_focal_aug_20260510_111158
• Source checkpoint: checkpoints/dinov3_safe_focal_aug_20260510_111158/best-32-0.9537.ckpt
• Export format: model.safetensors plus metadata.json, with ONNX for deployment
• Calibration status: not applied yet; this repo is the pre-calibration package and will be updated later with temperature scaling

Classes

• SUV
• VAN
• STATION WAGON
• MICRO
• OPEN WHEEL / F1
• SEDAN
• HATCHBACK
• PICK UP

Preprocessing

• RGB input
• Resize: 256
• Center crop / model input: 256 x 256
• Mean: [0.4429, 0.4354, 0.437]
• Std: [0.2456, 0.2421, 0.2449]

Validation Results

Validation was used to select the checkpoint. The metrics below come from the best validation epoch in metrics.csv.

• Best epoch: 32
• Accuracy: 0.9658
• F1-macro: 0.9537
• F1-weighted: 0.9659
• Precision-macro: 0.9587
• Recall-macro: 0.9500
• Loss: 0.3394

Internal Test Results

Evaluated on the held-out internal test set (3170 samples, no data leakage).

• Accuracy: 0.9505
• F1-macro: 0.9302
• F1-weighted: 0.9502
• Precision-macro: 0.9358
• Recall-macro: 0.9269

Per-class Performance

Class	Precision	Recall	F1-score	Support
SUV	0.9171	0.9582	0.9372	670
VAN	1.0000	0.9670	0.9832	455
STATION WAGON	0.8133	0.9242	0.8652	66
MICRO	0.9524	0.9091	0.9302	22
OPEN WHEEL / F1	0.9983	0.9949	0.9966	586
SEDAN	0.9423	0.9761	0.9589	836
HATCHBACK	0.9156	0.8158	0.8628	266
PICK UP	0.9474	0.8699	0.9070	269

Training Curves

Internal Test Confusion Matrix

Artifact Sizes

• model.safetensors: 82.373 MB
• model.onnx: 82.572 MB

Files

• model.safetensors — weights-only model artifact
• metadata.json — architecture, classes, preprocessing, source run, and artifact metadata
• model.onnx — ONNX Runtime inference artifact
• metrics.csv — epoch-level training and validation history
• per_class_metrics.txt — internal-test classification report
• training_loss.png — training/validation loss curve
• training_accuracy.png — training/validation accuracy curve
• confusion_matrix.png — normalized internal-test confusion matrix
• size_check.json / size_report.json — artifact size and ONNX smoke evidence
• internal_test_results.json — structured internal-test summary

Intended use

Educational demo and report evidence for classifying uploaded vehicle images into the 8 AutoLens body-type classes.

Limitations

The dataset is assembled from public/open sources and may contain domain bias. Similar body styles such as hatchback, station wagon, and sedan can be ambiguous. This repo is intentionally pre-calibration; probability calibration will be added in the next upload after temperature scaling is fitted.