CNN from scratch using only Numpy

I developed this project because I wanted to deeply understand the mathematics behind the learning process of a neural network.

The architecture is based on a VGG block, and the goal is to perform image classification.

The model has been trained both on CIFAR10 and MNIST.

In the config.py it is possible to set variables that allow to change some model parameters (e.g.: dataset, optimizer, convolution type, ...).

To run the project just start the main.py file.

TODO

• Add bias to convolutional layers
• complete the documentation by adding examples for better understanding the math
• remove some configuration flags (used during for the project presentation)
• remove math sample from comments code (used during for the project presentation)