This project is a small custom Autograd engine built to further my knowledge of neural networks and machine learning. This engine implements backpropagation over a dynamically built Directed Acyclic Graph (DAG). This project involved developing a small neural network’s library with a PyTorch-like API and included the capability for tracing and visualization using Graphviz visualizations.
The notebook demo.ipynb includes a full demo on training a multilayer neural network
(MLP) binary classifier. This is achieved by initializing a neural net from the src.nn module,
implementing a simple svm "max-margin" binary classification loss and using SGD for optimization.
The notebook trace_graph.ipynb produces graphvis visualizations.
The unit tests require the PyTorch dependency.
python -m pytest