Loading...
Loading
Development Activity
121111 contributors0 commits/30dLast commit 134d ago
MITCI/CDDockerRequirements
Top Contributors
Loading
Powered by NATIX’s Internet of Cameras, StreetVision is advancing autonomous driving, Physical AI, and map-making.
Neuron registration trends
Construction site detection and roadwork identification require accurate computer vision models, but maintaining infrastructure to train, deploy, and evaluate these models is expensive and resource-intensive. Organizations either build proprietary models in-house—requiring significant ML expertise and GPU infrastructure—or rely on generic off-the-shelf solutions that lack domain specificity. There's no standardized way to compare model performance across diverse real-world conditions or incentivize continuous improvement.
A decentralized approach distributes this work across a network of participants. Miners develop and deploy their own classification models; validators challenge miners with diverse real-world and synthetic images, score predictions objectively, and rank models by accuracy. This shifts the burden from any single organization to a collaborative network where contributors are rewarded based on performance. The system maintains a public record of which models work best on which image types, reducing duplication and accelerating collective improvement.
StreetVision implements this on Bittensor, a decentralized machine learning network. Miners submit models to a public repository; validators continuously test them against construction site imagery. Performance rankings are broadcast on-chain, creating an incentive mechanism that rewards accuracy and penalizes stale models. Over time, the network converges toward high-quality detectors without requiring centralized infrastructure or funding.
Use cases were auto-generated from the GitHub repository and may not reflect the latest changes.
StreetVision is a decentralized Bittensor subnet for image classification and object detection that incentivizes miners to develop accurate models for detecting construction sites and roadwork in images. Validators challenge miners with a balanced mix of real and synthetic media, scoring their predictions and maintaining network integrity through a dynamic reward system that encourages continuous model improvement and submission.
The subnet operates as a distributed machine learning system where miners run inference models on images received via Bittensor's P2P layer (dendrite/axon calls) and return predictions as float values. Validators generate challenges by fetching images from a local cache (populated from HuggingFace datasets and S3), optionally applying augmentations, and sending them to miners via the ImageSynapse protocol. Validators score predictions against ground truth and rank miners accordingly.
Rankings are broadcast on-chain via the Bittensor metagraph using set_weights(). A background proxy service polls an external NATIX API for additional consensus tasks. The system uses modular challenge sources (real roadwork imagery, synthetic generation via diffusion models) to test miner robustness.
State is persisted in wandb and locally; miners load detector models from HuggingFace or local weights directories.
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository for the latest instructions.
Miners must submit at least one model to a public HuggingFace repository. For the first 45 days after submission, miners receive full rewards for correct predictions. After 45 days, rewards gradually decay toward zero, incentivizing submission of improved models.
Validators challenge miners with a balanced mix of organic tasks (real roadwork images) and synthetic challenges (AI-generated construction site images). Miner predictions (float 0-1, where >0.5 indicates roadwork present) are scored against ground truth. Validators earn dividends for maintaining network integrity.
The Yuma Consensus algorithm translates ranking and performance data into final incentive payouts.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository
Powered by NATIX’s Internet of Cameras, StreetVision is advancing autonomous driving, Physical AI, and map-making.
FAQs were auto-generated from the GitHub repository and may not reflect the latest changes.
Holder & total-staked metrics as of 5h ago (hourly snapshot)
Minimum compute spec includes GPU with 4GB VRAM, 2 cores CPU at 2.5 GHz. Recommended: 8GB VRAM, 4 cores at 3.5 GHz, 100GB SSD storage with 1000 IOPS.
Minimum compute spec includes GPU with 20GB VRAM, 4 cores CPU at 2.5 GHz, 32GB RAM. Recommended: 24GB VRAM (RTX A5000), 8 cores at 3.5 GHz, 64GB RAM, 500GB SSD storage with 5000 IOPS.
Clone the repository
git clone https://github.com/natixnetwork/natix-subnet.git && cd natix-subnet
Create and activate Python 3.11 virtual environment
python3.11 -m venv venv && source venv/bin/activate
Install miner dependencies (torch, timm, ultralytics, opencv, scikit-image)
pip install -e ".[miner]"
Pre-download training datasets (optional, only required for training)(optional)
python natix/miner/datasets/download_data.py
Generate miner.env configuration file with default values
bash ./setup_env.sh
Configure wallet, hotkey, miner port (default 8091), image detector, and device (cpu/cuda)
# Edit miner.env with wallet name, hotkey, port, and model configuration
Register hotkey on mainnet (netuid 72). Use netuid 323 and --subtensor.network test for testnet
btcli s register --netuid 72 --wallet.name [wallet_name] --wallet.hotkey [wallet.hotkey] --subtensor.network finney
Start the miner neuron
chmod +x ./start_miner.sh && ./start_miner.sh
Clone the repository
git clone https://github.com/natixnetwork/natix-subnet.git && cd natix-subnet
Create and activate Python 3.11 virtual environment
python3.11 -m venv venv && source venv/bin/activate
Install validator core dependencies (wandb, joblib, no heavy ML libraries)
pip install -r requirements.validator.txt
Install full validator with synthetic image generation support (adds torch, diffusers, transformers)(optional)
pip install -r requirements.validator-full.txt
Generate validator.env configuration file
bash ./setup_env.sh
Configure validator parameters including wallet, subnet ID, ports, and API credentials
# Edit validator.env with: WALLET_NAME, WALLET_HOTKEY, NETUID (72 mainnet/323 testnet), SUBTENSOR_NETWORK, VALIDATOR_AXON_PORT, VALIDATOR_PROXY_PORT, PROXY_CLIENT_URL, WANDB_API_KEY, HUGGING_FACE_TOKEN
Register validator hotkey on mainnet (netuid 72). Use netuid 323 and --subtensor.network test for testnet
btcli s register --netuid 72 --wallet.name [wallet_name] --wallet.hotkey [wallet.hotkey] --subtensor.network finney
Start validator using PM2 (launches validator neuron, cache updater, and synthetic generator)
pm2 start scripts/ecosystem.validator.config.js
