Read the full writeup on https://kuenzi.dev/nanograd/
This Github project implements a neural network pipeline from scratch without using any machine learning libraries including NumPy. The project is inspired by Andrej Karpathy's video series on building neural networks from scratch. It aims to implement a forward and backward pass, a simple classification task, a multi-layer perceptron, and RMSE as the loss function.
With PyTorch a simple digits classifier is trained and then used to generate images with the help of backpropagation and gradient descent.