Convolutional-Neural-Networks-from-Scratch

Python Library for creating and training CNNs. Implemented from scratch.

This repo contains a project i did during my second year in college. I wanted to have a deeper understanding of how gradients were calculated and backpropagated through the Network and I felt like if I could implement backprop of a Convolutional Nerual Net that would help me immensely.

The project can be divided into 3 parts:

• Optimizers.py

• ConvNetModule.py

• Layers.py

---

Optimizers.py

Contains four Algorithms for optimization

• SGD
• Adam
• Momentum
• RMSProp

---

ConvNetModule.py

Controls the training process, interacts with the layers and sends the gradients to the optimizers.

---

Layers.py

The hardest part to implement. Contains forward and backward operations for all the layers.

---

Additional Libraries used: Numpy(matrix operations), tqdm(appearance) and Matplotlib(plotting).

---

There are still some things that need work and can be found here.

For help with understanding the Backward operation of Convolutions:
https://github.com/JeyrajK/convolutional-neural-networks/blob/master/Back%20propagation%20of%20cnn.ipynb
https://towardsdatascience.com/forward-and-backward-propagations-for-2d-convolutional-layers-ed970f8bf602
https://cs231n.github.io/
https://www.jefkine.com/general/2016/09/05/backpropagation-in-convolutional-neural-networks/
https://towardsdatascience.com/backpropagation-in-a-convolutional-layer-24c8d64d8509