---

NN-ab-ovo (abovo) is a neural network engine written in C++ with Python bindings, designed to teach systems-level ML optimizations like threading, cache efficiency, hardware acceleration, and quantization.

---

Features

• C++ backend with modular layers and training pipeline
• Python API via pybind11 bindings (pip install abovo)
• Optimizations: SIMD, OpenMP multithreading, cache blocking
• Post-training quantization (PTQ) and quantization-aware training (QAT) (FP32 → INT8)
• Profiling support (Valgrind, cache misses, instruction counts)

Installation

pip install abovo

Requires a C++17-compatible compiler and OpenMP support.

Example (XOR)

from abovo import Sequential, DenseLayer, Matrix, ActivationType, LossType

X = Matrix(4, 2)
X[0, 0] = 0; X[0, 1] = 0
X[1, 0] = 0; X[1, 1] = 1
X[2, 0] = 1; X[2, 1] = 0
X[3, 0] = 1; X[3, 1] = 1

y = Matrix(4, 1)
y[0, 0] = 0
y[1, 0] = 1
y[2, 0] = 1
y[3, 0] = 0

model = Sequential()
model.add(DenseLayer(2, 4, ActivationType.RELU))
model.add(DenseLayer(4, 1, ActivationType.SIGMOID))
model.train(X, y, epochs=100, batch_size=1, learning_rate=0.1, loss_type=LossType.MSE)

Build (C++ Only)

You can either build natively or in Docker. Note the provided Dockerfile runs valgrind, so adjust as needed to run the correct binary. Recommended on Apple Silicon for x86 builds.

Native Build (Mac/Linux):

make
./NN-ab-ovo

Docker (x86_64 emulation):

docker build -t nn-ab-ovo .
docker run --rm nn-ab-ovo

Make sure the MNIST dataset files (train-images.idx3-ubyte, train-labels.idx1-ubyte, etc.) are in the project root or mounted into the Docker container.

Datasets

• XOR: Validates non-linear separability
• MNIST: Handwritten digit classification

Optimizations

Optimization experiments are documented in the GitHub repository under optimizations.md, including:

• Naive vs. blocked matrix multiplication
• Compiler flag benchmarking
• L1/L2 cache miss analysis (Valgrind)
• OpenMP and SIMD speedups
• Timing analysis with std::chrono

These experiments help evaluate system-level performance, guide improvements for training/inference, and validate optimizations available to the community.

Project Structure

The GitHub repository features the following project structure:

• Matrix.hpp / Matrix.cpp: Core matrix operations and linear algebra utilities.
• DenseLayer.hpp / DenseLayer.cpp: Fully connected layer with forward and backward pass.
• Activation.hpp / Activation.cpp: Support for activation functions (e.g., ReLU, LeakyReLU, Sigmoid).
• Loss.hpp: Interface for loss functions (e.g., MSE, CrossEntropy).
• Sequential.hpp / Sequential.cpp: High-level container for layer sequencing and model training.
• tests: Directory containing runnable code on specific datasets.

The engine is modular: activation functions, loss functions, and layers are easily swappable for flexibility and experimentation.

Documentation

Read the full docs at: https://nn-ab-ovo.readthedocs.io/.

Source Code

You can find the source code at: https://github.com/emirdur/abovo.

Future Work

• Switch Design Pattern for Activation + Loss
• Switch Matrix class to use size_t + Refactor
• More comprehensive Softmax implementation
• Continue with optimizations
• Add support for convolutional layers
• Implement GPU acceleration (Metal or CUDA)
• LLVMs?

Disclaimer

NN-ab-ovo (abovo) is an independent open-source project and is not affiliated with or endorsed by any company or organization.

License

MIT License