CNN-from-scratch

A minimal implementation of Convolutional Neural Network using only Numpy and SciPy

Installation

create a new directory and cd into it. start by cloning the repo:

git clone https://github.com/Niazi04/CNN-from-scratch/tree/main

Intall dependencies:

python install -r requiremenets.txt

Training Your First Model

Currently the code base and model creation is highly coupled. This will be fixed in future updates.

Follow the details to train your model in trainYourFirstModel.py The process is fully explained and limitations are mentioned

Future Improvements

• Decouple model creation and training loop (ASAP)
• Decouple Loss Functions and Optimizers into their own classes
• Fix pool size problem
• Clean up random variable naming conventions
• Add different related optimizers
• Add different related Loss Functions
• Implement batch normalizer (similar to DataLoader in Pytorch)