Neural-Net-Framework

I custom library I made for training neural networks from scratch, using numpy and scipy

Imformation ℹ

I've been working on many neural network projects for the past year now, with my earliest project finished in April 2024

And I've learned a lot through much trial and error in the Machine Learning field and my end goal is to create a boudning box regression model similiar to yolo-v1.

Purpose:

I noticed my old code was really limiting and innefecient:
- https://github.com/TheonlyIcebear/Image-Recognition-AI
- https://github.com/TheonlyIcebear/Tic-Tac-Toe-Machine-Learnin
I decided to do a complete rewrite allowing my code to be more modular much like keras' Sequential class allowing me to quickly add features like Batch Normalization

Image classification problem training loss:

Features ⚙

The model is completely sequential, meaning the output from one layer will only go to a single next layer
Performs a 2d convolution, good for feature extraction in images

It includes, Convolutional Layers, Dense Layers, Batch Normalization and Max Pooling

Notes

This is NOT a replacement for actual Neural Network libraries such as Tensorflow or PyTorch, this is simply a library i made because I want to understand neural networks on a deep level.
If you value your time, please don't use this on serious projects lol
For some reason (aka im too dumb to fix it) I cant get the custom striding to work but I decided to leave it in anyways
This is CPU bound only meaning it is really slow compared to other GPU based models, I plan to add modules such as CuPy once I get a Nividia GPU so I can use and test CUDA, but for now multiprocessing it is.

TheonlyIcebear/Neural-Net-Framework

Neural-Net-Framework

Imformation ℹ

Purpose:

Features ⚙

Notes

Sources 🔌