Prometheus

A machine learning library built from scratch in C++, with Python bindings (in progress).

Prometheus implements the core components of a modern deep learning framework — tensors, automatic differentiation, neural network layers, optimizers, and data utilities — without relying on any external ML dependencies.

---

Why

Most ML libraries are black boxes. Prometheus is an attempt to understand what's actually happening under the hood — every operation, every gradient, every weight update written by hand.

---

What's implemented

Tensor

• N-dimensional tensor with flat data storage
• Element-wise ops: add, subtract, multiply, divide, pow, sqrt, abs, exp, log, clip
• Matrix ops: matmul, transpose, reshape
• Reductions: sum, mean, max, min along an axis
• Factory methods: zeros, ones, randn

Autograd

• Reverse-mode automatic differentiation
• Computation graph built dynamically on forward pass
• backward() via topological sort
• requires_grad flag per tensor

Neural Network Layers

• Linear — fully connected layer
• ReLU, Sigmoid, Tanh, Softmax — activations
• Dropout — regularization
• Sequential — chain layers together

Loss Functions

• mse_loss — mean squared error
• mae_loss — mean absolute error
• binary_cross_entropy
• cross_entropy_loss

Optimizers

• SGD with optional momentum
• Adam
• RMSprop

Data Utilities

• DataLoader — batch iteration, shuffling, reshuffle per epoch
• data_split — train/val/test split with optional shuffle
• read_csv — load tabular data from CSV into tensors
• load_image — load PNG/JPG images as tensors [C, W, H]

Metrics

• accuracy — binary and multi-class

Model Utilities

• save / load — serialize model weights to binary file
• train — training loop helper

---

Project structure

include/ml/
├── tensor.hpp
├── ops.hpp
├── loss.hpp
├── autograd.hpp
├── nn/          — module, linear, activations, dropout, sequential
├── optim/       — optimizer, sgd, adam, rmsprop
├── data/        — dataloader, csv, image loader
├── metrics/     — accuracy, precision, recall etc.
└── utils/       — model_io, trainer

src/             — implementations
tests/           — one test file per component

---

Build

Requires CMake and a C++17 compiler.

mkdir build && cd build
cmake ..
cmake --build .

---

Roadmap

• Convolutional layers (Conv2D, MaxPool2D)
• Recurrent layers (RNN, LSTM, GRU)
• Attention and Transformers
• Python bindings via pybind11
• GPU support via CUDA