Runet Inference Engine

Runet is a high-performance, lightweight C-based inference engine for Deep Learning models. By bypassing the overhead of heavy frameworks, Runet achieves 6.0x faster execution than TensorFlow on standard benchmarks (MNIST) with identical accuracy.

Features

• Pure C Backend: Optimized matrix operations with zero memory leaks (Valgrind-verified).
• Fast Inference: Specialized for Feed-Forward Neural Networks (Dense layers).
• Python Integration: Seamlessly export TensorFlow/Keras models to .bin files and run them via a Python wrapper.
• Lightweight: Minimal dependencies, perfect for resource-constrained environments.

Performance Benchmark

Framework	Task	Accuracy	Inference Speed (Relative)
TensorFlow	MNIST	98%	1.0x
Runet Engine	MNIST	98%	6.0x

Project Structure

• src/: C Source code for matrix math, layers, and network orchestration.
• include/: Header files and internal API definitions.
• scripts/: Python API (runet.py, c_wrapper.py).
• tests/: Extensive C unit test suite.
• libs/: Compiled shared library (librunet.so) and model binaries.

Quick Start (Google Colab/Linux)

1. Clone & Setup:

   git clone https://github.com/Karan2004xd/runet.git
   cd runet
   make clean && make librunet.so
   

2. Usage:

   from scripts.runet import RunetEngine
   
   engine = RunetEngine()
   
   # Load Pre-trained tensorflow model
   # The engine automatically exports it to .bin format
   engine.load_model(your_tf_model)
   
   # Predict
   result = engine.predict(input_data)

Limitations

• Training is not supported (only inference).
• CNNs, RNNs, Transformers are not supported.
• Models must be exported using Runet's exporter.
• Batch inference support is limited to feed-forward execution.

Supported Model Components

Models

• Sequential

Layers

• Dense (Fully Connected)

Activation Functions

• ReLU
• Softmax

Educational Purpose

This project was developed as a part of college coursework to explore,

• Neural network inference internals
• Systems-level performance optimization
• C and Python interoperability
• Custom binary formats
• Deterministic execution and memory management