Autoencoder from Scratch

This project implements basic autoencoder neural networks from scratch using PyTorch, designed to learn and compress image representations in an unsupervised manner. The models are trained on the MNIST dataset of handwritten digits.

A practical project to understand the working principles of autoencoders, how they compress input data, and how reconstruction loss drives the training process.

---

Overview

An autoencoder is a type of neural network that learns to encode data into a compressed latent representation and then decode it back to the original input form. It has two main components:

• Encoder: Compresses the input into a lower-dimensional latent space.
• Decoder: Reconstructs the input from the latent representation.

This project includes:

• Data loading using torchvision (MNIST dataset)
• Two custom autoencoder architectures built with PyTorch (nn.Sequential)
• Training loops optimizing reconstruction loss (mean squared error)
• Visualization of original and reconstructed digits

---

Model Architectures

1. Fully Connected (Dense) Autoencoder

Encoder:

• Flatten input images (28×28 → 784 features)
• Linear(784 → 128), ReLU
• Linear(128 → 64), ReLU
• Linear(64 → 12), ReLU
• Linear(12 → 3) (Latent representation: 3-dimensional)

Decoder:

• Linear(3 → 12), ReLU
• Linear(12 → 64), ReLU
• Linear(64 → 128), ReLU
• Linear(128 → 784), Sigmoid
• Reshape output to image size (784 → 28×28)

This architecture provides a deep encoder and decoder, compressing the original input down to a compact 3-dimensional latent space before reconstructing the images.

---

2. Convolutional Autoencoder

Encoder:

• Conv2d: 1 → 16, kernel size=3, stride=2, padding=1 → ReLU
• Conv2d: 16 → 32, kernel size=3, stride=2, padding=1 → ReLU
• Conv2d: 32 → 64, kernel size=7 (output: 1×1×64 latent representation)

Decoder:

• ConvTranspose2d: 64 → 32, kernel size=7 → ReLU
• ConvTranspose2d: 32 → 16, kernel size=3, stride=2, padding=1, output_padding=1 → ReLU
• ConvTranspose2d: 16 → 1, kernel size=3, stride=2, padding=1, output_padding=1 → Sigmoid

This convolutional architecture preserves spatial relationships and typically yields sharper reconstructions compared to fully connected variants.

---

Credits

This project is inspired by and based on the YouTube tutorial:

"Autoencoders - EXPLAINED!" by Aladdin Persson
https://www.youtube.com/watch?v=zp8clK9yCro

Full credit goes to Aladdin Persson for the clear explanations and foundational code walkthroughs.