fdavrs 0.2.0

A self-healing SDK that detects and fixes data drift in deployed computer vision models using federated learning

FDAVRS — Federated Drift-Aware Vision Reliability System

FDAVRS is a PyTorch SDK that makes computer vision models self-healing. When deployed models encounter real-world distribution shift — fog, lighting changes, sensor noise, blur — they fail silently. FDAVRS detects this automatically and fixes the model at runtime, without labels and without retraining.

Install

pip install fdavrs

The Problem It Solves

A ResNet trained on clean images loses 30–40% accuracy when deployed in fog. No error is raised. The model just starts predicting wrong — silently. FDAVRS wraps the model and fixes this automatically.

Quick Start

import torch
import torchvision
from fdavrs import FDAVRS

# 1. Load your existing model — nothing changes here
base_model = torchvision.models.resnet18(pretrained=True)

# 2. Wrap it with FDAVRS — one line
model = FDAVRS(
    client_model=base_model,
    feature_layer='avgpool',
    threshold=0.3
)

# 3. Calibrate once on clean data before deployment
model.fit(clean_data_loader)

# 4. Use exactly like a normal model during inference
for images in live_feed:
    predictions = model.predict(images)

    status = model.status()
    print(status['action'])
    # IDLE                  → model is healthy
    # LOCAL_ADAPTATION      → drift detected, fixing locally
    # GLOBAL_CURE_APPLIED   → fix retrieved from federated server

How It Works

When predict() is called, FDAVRS runs a 4-layer pipeline automatically:

1. Monitor Layer — A forward hook intercepts internal activations from the specified feature layer. It computes a Composite Drift Score: Score = (0.5 × Entropy) + (0.3 × Cosine Shift) + (0.2 × (1 - Confidence))

2. Decision Layer — Triage based on score:

• Score ≤ 0.3 → IDLE (model is reliable)
• 0.3 < Score ≤ 0.8 → LOCAL_ADAPTATION (fix locally via TTA)
• Score > 0.8 → REQUEST_SERVER_CURE (ask federated server)

3. Adaptation Layer — BatchNorm running statistics (μ and σ²) are recalculated on the current batch using Test-Time Adaptation. No labels required. No architecture changes. No retraining.

4. Knowledge Vault — Successful fixes are packaged as weight deltas and drift signatures, then uploaded to a federated server so other devices with the same drift pattern benefit instantly.

API Reference

FDAVRS(client_model, feature_layer, threshold=0.3)

Wraps your model with drift monitoring and self-healing.

Parameter	Type	Description
client_model	torch.nn.Module	Your pretrained PyTorch model
feature_layer	str	Layer name to monitor e.g. 'avgpool' for ResNet
threshold	float	Drift sensitivity. Default 0.3

fit(dataloader) / fit_baseline(dataloader)

Calibrates the baseline using clean data. Call once before deployment.

predict(images) / forward(images)

Drop-in replacement for model(images). Drift detection and adaptation happen automatically inside.

status()

Returns current SDK state:

{
    "action": "IDLE | LOCAL_ADAPTATION | GLOBAL_CURE_APPLIED | ...",
    "metrics": {
        "score": 0.42,
        "entropy": 0.81,
        "shift": 0.23,
        "confidence": 0.61
    }
}

Compatible Models

Tested with ResNet-18, YOLOv11, DeepLabV3, and custom CNNs. Any PyTorch model with BatchNorm layers is supported.

Requirements

• Python ≥ 3.8
• PyTorch ≥ 1.9.0
• NumPy ≥ 1.19.0
• SciPy ≥ 1.7.0