⚡ FlashEdge

Edge AI & TinyML Toolkit for FlashVision

ONNX • TFLite • CoreML • OpenVINO • Quantization • Pruning • Profiling • Deployment

---

What is FlashEdge?

FlashEdge is a comprehensive Edge AI & TinyML toolkit for the FlashVision ecosystem. It provides everything you need to compress, convert, profile, and deploy deep learning models on edge devices, mobile platforms, and IoT hardware.

Key Features

Feature	Description
ONNX Export	PyTorch to ONNX conversion with graph optimization and simplification
TFLite Export	PyTorch to TensorFlow Lite for Android/embedded deployment
CoreML Export	PyTorch to CoreML for iOS/macOS deployment
OpenVINO Export	PyTorch to OpenVINO IR for Intel hardware
Quantization	INT8/INT4 post-training and quantization-aware training
Pruning	Structured channel/filter pruning for mobile-optimized models
Distillation	Knowledge distillation from large to compact models
NAS	Latency-constrained neural architecture search
Profiling	Latency, memory, FLOPs/MACs, and power consumption analysis
Runtime	Unified inference API for ONNX Runtime, TFLite, OpenVINO
Benchmarking	Cross-format accuracy, latency, throughput, and model size comparison

---

Installation

From PyPI (recommended)

pip install flashedge

From Source

git clone https://github.com/FlashVision/FlashEdge.git
cd FlashEdge
pip install -e ".[all]"

With Optional Dependencies

pip install flashedge[tflite]      # TFLite runtime support
pip install flashedge[openvino]    # OpenVINO support
pip install flashedge[coreml]      # CoreML support
pip install flashedge[analytics]   # Visualization & profiling
pip install flashedge[all]         # Everything
pip install flashedge[dev]         # Development tools

---

Quick Start

Export a Model to ONNX

from flashedge import FlashEdge, Exporter

model = FlashEdge("pretrained/model.pth")
exporter = Exporter(model)
exporter.export_onnx("model.onnx", input_shape=(1, 3, 224, 224), opset=17)

Profile Model Latency

from flashedge import DeviceProfiler

profiler = DeviceProfiler(device="cpu")
report = profiler.profile("model.onnx", input_shape=(1, 3, 224, 224), runs=100)
print(f"Latency: {report['mean_ms']:.2f} ms | Throughput: {report['fps']:.1f} FPS")

Quantize for Edge (INT8)

from flashedge import ModelOptimizer

optimizer = ModelOptimizer()
optimizer.quantize("model.onnx", output="model_int8.onnx", dtype="int8")

Benchmark Across Formats

from flashedge import Benchmark

bench = Benchmark()
results = bench.compare(
    models={"onnx": "model.onnx", "int8": "model_int8.onnx"},
    input_shape=(1, 3, 224, 224),
    runs=100,
)
bench.print_report(results)

---

CLI Usage

# Show version
flashedge version

# Check installation
flashedge check

# Export model to ONNX
flashedge export --config configs/flashedge_onnx_mobile.yaml

# Profile model latency
flashedge profile --model model.onnx --device cpu --runs 100

# Quantize model
flashedge optimize --config configs/flashedge_tflite.yaml

# Benchmark model
flashedge benchmark --model model.onnx --device cpu --runs 200

# Deploy to edge device
flashedge deploy --model model.onnx --target onnxruntime --output deployed/

---

Project Structure

FlashEdge/
├── configs/          # YAML configuration files
├── docker/           # Docker support
├── docs/             # Documentation
├── examples/         # Runnable example scripts
├── flashedge/        # Core library
│   ├── cfg/          # Configuration management
│   ├── data/         # Calibration & benchmark data
│   ├── engine/       # Training, validation, export
│   ├── export/       # Format exporters (ONNX, TFLite, CoreML, OpenVINO)
│   ├── optimization/ # Quantization, pruning, distillation, NAS
│   ├── runtime/      # Inference runtimes
│   ├── profiling/    # Latency, memory, FLOPs, power
│   ├── models/       # Model architectures
│   ├── solutions/    # High-level solutions
│   ├── analytics/    # Benchmarking & metrics
│   └── utils/        # Utilities
└── tests/            # Unit tests

---

Benchmarks

Model	Format	Size	Latency (CPU)	Throughput	Accuracy
MobileNetV3-S	PyTorch	10.2 MB	8.5 ms	118 FPS	67.4%
MobileNetV3-S	ONNX	9.8 MB	5.2 ms	192 FPS	67.4%
MobileNetV3-S	ONNX INT8	2.8 MB	2.1 ms	476 FPS	66.9%
MobileNetV3-S	TFLite	2.6 MB	3.1 ms	323 FPS	67.2%
EfficientNet-B0	ONNX	20.4 MB	12.3 ms	81 FPS	77.1%
EfficientNet-B0	ONNX INT8	5.4 MB	4.8 ms	208 FPS	76.5%

---

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

---

License

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

---

Acknowledgements

• PyTorch for the deep learning framework
• ONNX for model interoperability
• ONNX Runtime for optimized inference
• FlashVision ecosystem

---

Made with ❤️ by the FlashVision team