vectra-sdk 0.1.0

A high-level inference SDK for Vectra Engine few-shot models.

Vectra SDK

A high-level Python library for efficient and user-friendly inference using few-shot models generated by the Vectra Engine.

Features

• Multi-Format Input: Seamlessly predict using file paths, PIL images, or NumPy arrays (OpenCV frames).
• Batch Inference: Process multiple images in a single call.
• Live Stream Support: Simple wrapper for real-time camera inference with visual overlays.
• Unknown Category Rejection: Natively supports out-of-distribution detection.
• Confidence Scores: Returns similarity-based confidence for every prediction.
• Optimized: Automatic GPU acceleration where available.

Installation

Install via pip:

pip install vectra-sdk

Quick Start

1. Basic Inference

from vectra.inference import VectraInference

# Initialize with your trained .pt model
sdk = VectraInference("my_model.pt")

# Predict using a file path
result = sdk.predict("image.jpg")
print(f"Prediction: {result['label']} (Confidence: {result['confidence']:.2f})")

# Predict using a PIL Image
from PIL import Image
img = Image.open("photo.png")
result = sdk.predict(img)

2. Live Camera Feed

from vectra.inference import VectraInference
from vectra.utils.vision import LiveStreamInference

sdk = VectraInference("my_model.pt")
live = LiveStreamInference(sdk)

# Start real-time inference on default camera (index 0)
live.start(camera_index=0)

3. Batch Processing

results = sdk.predict_batch(["img1.jpg", "img2.png", "img3.jpg"])
for res in results:
    print(res['label'])

Requirements

• Python 3.8+
• PyTorch & Torchvision
• OpenCV
• Pillow
• NumPy

License

This project is licensed under a Custom Research and Educational Use License. Commercial use is prohibited without prior permission.