aegis-vision 0.2.7

Cloud-native computer vision model training toolkit for Aegis AI

Aegis Vision

Cloud-native computer vision model training toolkit for Aegis AI

Overview

Aegis Vision is a streamlined toolkit for training computer vision models in cloud environments (Kaggle, Colab, etc.) with built-in support for:

• 🎯 YOLO Models (v8, v9, v10, v11) - Object detection training
• 📊 Wandb Integration - Experiment tracking and visualization
• 🔄 COCO Format - Dataset conversion and handling
• ☁️ Cloud-Optimized - Designed for Kaggle/Colab workflows
• 📦 Model Export - ONNX, CoreML, OpenVINO, TensorRT, TFLite

Installation

Standard Installation

# Basic installation
pip install aegis-vision

# With Kaggle support
pip install aegis-vision[kaggle]

# Development installation
pip install aegis-vision[dev]

# All features
pip install aegis-vision[all]

Nvidia DGX / High-Performance GPU Systems

For Nvidia DGX Spark or other systems with latest NVIDIA GPUs (Blackwell architecture), installation is the same:

# Standard installation with automatic environment checking
pip install aegis-vision

# Login and start (agent will auto-check and fix environment)
aegis-agent login
aegis-agent start

The agent automatically:

• Detects environment issues (NumPy, PyTorch compatibility)
• Explains what's wrong and why
• Offers one-click fixes
• Starts agent after fixes

See QUICKSTART_DGX.txt for detailed guide.

Quick Start

Training a YOLO Model

from aegis_vision import YOLOTrainer

# Initialize trainer
trainer = YOLOTrainer(
    model_variant="yolov11l",
    dataset_path="/kaggle/input/my-dataset",
    epochs=100,
    batch_size=16,
)

# Configure Wandb tracking (optional)
trainer.setup_wandb(
    project="my-project",
    entity="my-team",
    api_key="your-api-key"
)

# Train
results = trainer.train()

# Export to multiple formats
trainer.export(formats=["onnx", "coreml", "openvino"])

Converting COCO to YOLO Format

from aegis_vision import COCOConverter

# Convert dataset
converter = COCOConverter(
    annotations_file="annotations.json",
    images_dir="images/",
    output_dir="yolo_dataset/"
)

stats = converter.convert()
print(f"Converted {stats['total_annotations']} annotations")

Command-Line Interface

# Train a model
aegis-train \
    --model yolov11l \
    --data /path/to/dataset \
    --epochs 100 \
    --batch 16 \
    --wandb-project my-project

# Convert COCO to YOLO
aegis-train convert-coco \
    --annotations annotations.json \
    --images images/ \
    --output yolo_dataset/

Features

🎯 YOLO Training

• Multi-version support: YOLOv8, v9, v10, v11
• Fine-tuning & from-scratch training modes
• Automatic augmentation configuration
• Early stopping with patience
• Validation metrics: mAP50, mAP50-95, precision, recall

📊 Experiment Tracking

• Wandb integration for metrics, charts, and artifacts
• Automatic logging of hyperparameters, metrics, and model outputs
• Run resumption support

🔄 Dataset Handling

• COCO format support
• Auto-conversion to YOLO format
• Label filtering and validation
• Dataset statistics reporting

📦 Model Export

• ONNX - Cross-platform inference
• CoreML - iOS/macOS deployment
• OpenVINO - Intel hardware optimization
• TensorRT - NVIDIA GPU optimization
• TFLite - Mobile/edge deployment

☁️ Cloud Environment Support

• Kaggle - Kernel execution and dataset management
• Google Colab - Ready-to-use notebooks
• Environment detection - Auto-configuration for different platforms

Configuration

Training Configuration

config = {
    # Model settings
    "model_variant": "yolov11l",
    "training_mode": "fine_tune",  # or "from_scratch"
    
    # Training hyperparameters
    "epochs": 100,
    "batch_size": 16,
    "img_size": 640,
    "learning_rate": 0.01,
    "momentum": 0.937,
    "weight_decay": 0.0005,
    
    # Augmentation
    "augmentation": {
        "hsv_h": 0.015,
        "hsv_s": 0.7,
        "hsv_v": 0.4,
        "degrees": 0.0,
        "translate": 0.1,
        "scale": 0.5,
        "shear": 0.0,
        "perspective": 0.0,
        "flipud": 0.0,
        "fliplr": 0.5,
        "mosaic": 1.0,
        "mixup": 0.0,
    },
    
    # Early stopping
    "early_stopping": {
        "enabled": True,
        "patience": 50,
        "min_delta": 0.0001
    },
    
    # Wandb
    "wandb_enabled": True,
    "wandb_project": "my-project",
    "wandb_entity": "my-team",
    
    # Export
    "output_formats": ["onnx", "coreml", "openvino"],
}

trainer = YOLOTrainer(**config)

Examples

Kaggle Kernel

# In a Kaggle kernel
from aegis_vision import YOLOTrainer

trainer = YOLOTrainer(
    model_variant="yolov11l",
    dataset_path="/kaggle/input/my-dataset",
    epochs=100,
    wandb_api_key="/kaggle/input/secrets/wandb_api_key.txt"
)

results = trainer.train()
trainer.save_to_kaggle_output()

Custom Dataset

from aegis_vision import YOLOTrainer, COCOConverter

# 1. Convert your COCO dataset
converter = COCOConverter(
    annotations_file="my_annotations.json",
    images_dir="my_images/",
    output_dir="yolo_dataset/",
    labels_filter=["person", "car", "dog"]  # Optional filtering
)
converter.convert()

# 2. Train
trainer = YOLOTrainer(
    model_variant="yolov11m",
    dataset_path="yolo_dataset/",
    epochs=50,
)
results = trainer.train()

API Reference

YOLOTrainer

Main class for training YOLO models.

Methods:

• train() - Start training
• setup_wandb() - Configure Wandb tracking
• export() - Export trained model
• validate() - Run validation
• get_metrics() - Retrieve training metrics

COCOConverter

Convert COCO format datasets to YOLO format.

Methods:

• convert() - Perform conversion
• validate() - Check dataset integrity
• get_statistics() - Dataset statistics

Development

# Clone repository
git clone https://github.com/your-org/aegis-vision.git
cd aegis-vision

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Format code
black src/

# Lint
ruff src/

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Roadmap

• Support for additional YOLO architectures
• Integration with Hugging Face Hub
• Distributed training support
• Auto-hyperparameter tuning
• Model quantization utilities
• Segmentation and pose estimation models
• Real-time inference utilities

Citation

@software{aegis_vision,
  title = {Aegis Vision: Cloud-native Computer Vision Training Toolkit},
  author = {Aegis AI Team},
  year = {2025},
  url = {https://github.com/your-org/aegis-vision}
}

Support

• 📧 Email: [e-mail]
• 💬 Discord: Join our community
• 📚 Documentation: https://aegis-vision.readthedocs.io
• 🐛 Issues: GitHub Issues