CNN Engine in Pure NumPy

Meme Motivation

After this project, I can say

---

About the Project

This mini-project is a minimalistic implementation of Convolutional Neural Networks (CNNs) built entirely from scratch using NumPy. I didn’t rely on any deep learning libraries like TensorFlow or PyTorch, the idea was to understand how CNNs works, including both forward and backward passes.

---

📌 What’s Inside

1) Zero Padding

Implemented a zero_pad() function to pad the input images symmetrically with zeros to preserve spatial dimensions before applying convolution.

2) Single-Step Convolution

conv_single_step() performs element-wise multiplication of a small filter on a region of the input, this is the core step behind convolution.

3) Full Convolution Forward Pass

conv_forward() handles a batch of inputs and filters to apply convolution operation over height, width, and channels. It also supports stride and padding hyperparameters.

4) Pooling Forward Pass

pool_forward() supports both max and average pooling, reducing dimensionality while preserving features.

5) Convolution Backward Pass

conv_backward() backpropagates through the convolution layer and calculates the gradients w.r.t. input, weights, and biases. Fully vectorized and compatible with batch inputs.

6) Max Masking for Backprop

create_mask_from_window() was used in the pooling backward pass to identify the max element during backpropagation.

7) Distribute Gradients for Average Pooling

distribute_value() helps in spreading the gradient equally during average pooling backward pass.

8) Pooling Backward Pass

pool_backward() completes the cycle by implementing backward propagation through both max and average pooling layers.

---

Tech Stack

• Python
• NumPy
• No frameworks 😎

---

💡 Usage

This project is great for:

• Learning the internals of CNNs
• Building intuition for backpropagation
• Interview prep
• Academic assignments

---