FGVC-IR8 (v9.5): Fine-Grained Iranian Vehicle Categorization & Open-Set Recognition Model

FGVC-IR8 (v9.5) is a standalone fine-grained vehicle classification and open-set recognition engine engineered for real-world Iranian automotive listings. It combines an offline DINOv2-ViT-B/14-reg4 backbone (with 28 mathematically fused LoRA adapters) with Competitive 8+2 Part-Slot Attention, a 2-Layer 5D Relational Graph Transformer, SubCenterArcFace ($C=8, K_s=4$), SetMILHead, and the v9.5 Two-Regime Geodesic & Structural OpenSetGuard.

Ecosystem Links (The Golden Triangle)

• Live Interactive Web Demo (Hugging Face Spaces): https://huggingface.co/spaces/ahmadzmvd/FGVC-IR8-Demo
• Source Code, Windows 1-Click App & 17-Stage Verification Suite (GitHub): https://github.com/ahmadZMVD/FGVC-IR8
• Full 18.8 GB Dataset, COCO Bounding Boxes & Shards (Kaggle): https://www.kaggle.com/datasets/ahmadzmvd/fgvc-ir8-iranian-vehicle-dataset

Checkpoint Inventory in This Repository

File	Size	Description
fgvc_ir8_laptop_standalone_fp16.pt	172.3 MB	Primary Standalone FP16 Bundle (Step 35,686): Fused DINOv2-ViT-B/14-reg4 + LoRA + 8-Slot 5D Relational Graph + SubCenterArcFace (266 tensors). Zero internet required at runtime.
best.pt	16.1 MB	Auxiliary Peak Validation Head (Step 28,039): Achieves 97.99% Balanced Validation Accuracy (BEST.json). Shared zero-copy with the primary backbone in engine.py v9.5.
fgvc_ir8_master_calibrated.pt	16.2 MB	Calibrated head weights at Step 35,686.
step_0035686.pt	56.5 MB	Full optimizer & trainable parameter state at Step 35,686 (MANIFEST.json).
engine.py	84 KB	Complete v9.5 inference engine with 17-stage --selftest verification suite.

Quick Usage in Python

from PIL import Image
from engine import FGVCInferenceEngine

# Initializes StandaloneFGVCModel + Dual-Head Ensemble + OpenSetGuard v9.5
engine = FGVCInferenceEngine(ckpt_dir="checkpoints", device_str="cpu")

img = Image.open("car_photo.jpg")
result = engine.predict([img], model_mode="pro_392")
print(result["status"], result["title_en"], f"{result['final_conf_pct']}%")