NanoNet

A tiny neural net built entirely in Numpy. No PyTorch, no Tensorflow, only raw code and math for learning

What's Inside

• MLP(784, 128, 10) from scratch
• Gets ~96% on MNIST
• Built to learn how a neural net actually works at its core: autograd engine, forward, backprop, gradient descent, Adam, softmax, cross-entropy; all in Numpy

Struture

• nanonet/autograd.py: Tensor engine with autograd
• nanonet/model.py: MLP + Adam, training uses cross-entropy loss
• nanonet/optim.py: Adam optimizer
• nanonet/manual.py: first version with manual backprop