roadwork-v4

Fine-tuned ViT image classifier for the NATIX Street Vision Bittensor subnet (SN72). Predicts whether a street-level image contains roadwork.

• Base model: natix-network-org/roadwork
• Labels: {0: "None", 1: "Roadwork"}
• Training data: subset of natix-network-org/roadwork (validator's own corpus)
• Augmentation during training: True (validator's apply_augmentation_by_level)
• Train / Val: 4310 / 478
• Epochs: 5, LR: 2e-05

Evaluation

Metric	Value
eval_loss	0.2539
eval_accuracy	0.9059
eval_mcc	0.7614
eval_mcc_clamped	0.7614
eval_f1	0.9358
eval_precision	0.9535
eval_recall	0.9188
eval_sim_reward	0.5793
eval_runtime	13.1531
eval_samples_per_second	36.3410
eval_steps_per_second	0.6080
epoch	5.0000

Usage in the SN72 miner

# Option A: pin via env (no config edit needed)
export MINER_LOCAL_MODEL_DIR=/path/to/local/clone

# Option B: point hf_repo at this repo
# Edit base_miner/detectors/configs/ViT_roadwork.yaml:
#   hf_repo: 'vuongnguyen92/roadwork-v4'

Direct inference

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch

processor = AutoImageProcessor.from_pretrained("vuongnguyen92/roadwork-v4")
model = AutoModelForImageClassification.from_pretrained("vuongnguyen92/roadwork-v4").eval()

img = Image.open("street.jpg").convert("RGB")
inputs = processor(images=img, return_tensors="pt")
with torch.inference_mode():
    logits = model(**inputs).logits
prob_roadwork = torch.softmax(logits, dim=-1)[0, 1].item()
print(f"P(roadwork) = {prob_roadwork:.3f}")