https://github.com/vansh-py04/convolution-net-from-scratch
A minimal NumPy-based implementation of a 3-layer convolutional neural network (CNN) from scratch — including custom forward and backward passes for conv, ReLU, pooling, affine, and softmax layers. Perfect for learning how CNNs actually work under the hood.
https://github.com/vansh-py04/convolution-net-from-scratch
backpropagation cnn computer-vision convolutional-neural-networks cs231n cs231n-assignment deep-learning from-scratch from-scratch-in-python fully-connected-network machine-learning maxpool2d nueral-networks numpy softmax stanford-deep-learning
Last synced: 29 days ago
JSON representation
A minimal NumPy-based implementation of a 3-layer convolutional neural network (CNN) from scratch — including custom forward and backward passes for conv, ReLU, pooling, affine, and softmax layers. Perfect for learning how CNNs actually work under the hood.
- Host: GitHub
- URL: https://github.com/vansh-py04/convolution-net-from-scratch
- Owner: vansh-py04
- Created: 2025-05-10T12:09:26.000Z (about 1 year ago)
- Default Branch: main
- Last Pushed: 2025-05-20T05:38:12.000Z (about 1 year ago)
- Last Synced: 2025-06-18T08:48:47.359Z (11 months ago)
- Topics: backpropagation, cnn, computer-vision, convolutional-neural-networks, cs231n, cs231n-assignment, deep-learning, from-scratch, from-scratch-in-python, fully-connected-network, machine-learning, maxpool2d, nueral-networks, numpy, softmax, stanford-deep-learning
- Language: Jupyter Notebook
- Homepage:
- Size: 271 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
## 🧠 ThreeLayerConvNet
A from-scratch implementation of a three-layer convolutional neural network (CNN) in NumPy with custom forward and backward passes for convolution, pooling, activation, affine layers, and softmax loss.
## 🔧 Architecture
Input → Conv → ReLU → MaxPool(2x2) → Affine → ReLU → Affine → Softmax
---
This minimal CNN supports:
* Custom weight initialization
* Naive forward/backward implementations (no deep learning libraries used)
* Layer composability through modular design
* Debugging-friendly shape printing for each major block
---
## 📁 Files Overview
| File | Description |
| ------------------------ | ---------------------------------------------------------------------- |
| `CNN.py` | Defines the `ThreeLayerConvNet` class with full forward/backward logic |
| `Conv2d.py` | Naive 2D convolution forward and backward pass |
| `ReLU.py` | ReLU activation forward and backward pass |
| `Maxpool.py` | Naive 2x2 max pooling forward and backward |
| `Affine_layer.py` | Fully-connected (affine) layer logic |
| `Softmax_loss.py` | Softmax loss and gradient implementation |
| `Sandwich_layers.py` | Helper functions for chaining layers: `conv-relu-pool`, `affine-relu` |
| `Implementation.ipynb` | Notebook demonstrating usage and testing of the network |
| `Images` | Test Images for out Convolution Layers |
The model can classify images of shape (3, 32, 32) into 10 classes. You can also plug it into datasets like CIFAR-10 (after appropriate preprocessing).
---
## 📌 Features
NumPy-only: No TensorFlow/PyTorch/Keras used
Backpropagation from scratch: Every layer implements its own gradient logic
Debug output: Key print statements show internal shape flows
Extensible design: Each component is modular and reusable
This repository is designed not just to use a CNN — but to help you learn how it actually works under the hood. Every major layer of a convolutional neural network has been manually implemented from scratch using just NumPy, with readable and modular Python code.
Here’s how you can use this repo to gain a deep understanding of CNN fundamentals:
# Study Each Layer in Isolation
Each layer (e.g., convolution, ReLU, max-pooling, affine, softmax) is implemented in its own file:
* Conv2d.py: Learn how convolutions apply filters using nested loops and stride/padding logic.
* ReLU.py: Understand how non-linear activations shape model capacity.
* Maxpool.py: Observe how spatial downsampling is performed via max-pooling.
* Affine_layer.py: Explore fully-connected transformations using matrix reshaping and dot products.
* Softmax_loss.py: Learn how classification loss and gradients are computed for training.
## Clone the Repo
git clone https://github.com/vansh-py04/Convolution-Net-from-Scratch.git
cd ThreeLayerConvNet
## 🙋♂️ Credits
Built by [Yuganter Pratap](https://www.linkedin.com/in/yuganter-pratap-a3a719254/) — inspired by CS231n and DIY deep learning educational projects.