Skip to content

README.md

Latest commit

 

History

History
106 lines (70 loc) · 3.19 KB

README.md

File metadata and controls

106 lines (70 loc) · 3.19 KB

Variational Autoencoder (VAE) on FashionMNIST

This project demonstrates a Variational Autoencoder (VAE) implemented from scratch using PyTorch to learn a compact, 2D latent representation of the FashionMNIST dataset. It highlights core deep learning principles including encoder-decoder architectures, the reparameterization trick, and latent space sampling — making it an ideal showcase of applied machine learning proficiency.

Objective

  • Develop and train a Variational Autoencoder (VAE) using PyTorch.
  • Learn and visualize compressed latent representations of image data.
  • Generate synthetic FashionMNIST images using decoder sampling.
  • Gain insights into model interpretability through visual analysis of the latent space.

Tools & Technologies

Purpose Tool/Library
Deep Learning PyTorch
Data Handling torchvision.datasets
Visualization Matplotlib, NumPy
Dataset FashionMNIST
IDE Google Colab / Jupyter

Core Components

  • Encoder: Maps flattened image (784D) to a 2D latent space (mean & log variance).
  • Reparameterization Trick: Enables backpropagation through stochastic sampling.
  • Decoder: Reconstructs original image from latent vector.
  • Loss Function: Combines reconstruction loss (binary cross-entropy) and KL divergence to regularize latent space.

Results

Training Loss Curve

The model demonstrates steady convergence over 10 epochs.

Loss Curve:

Latent Space Visualization

• Unlabeled Latent Space

Shows 2D clustering of 5000 samples in latent space.

• Colored by Class

Each cluster corresponds to a different FashionMNIST category.

Image Generation

• Grid Sampling from Latent Space

Samples from a 15x15 grid across 2D latent dimensions:

• Class-Averaged Latent Sampling

Each row shows slight variations around the latent mean vector for one category:

Key Learnings

  • Dimensionality Reduction: Compressing high-dimensional image space into 2D latent representations reveals structure in the data.
  • Generative Modeling: Once trained, the decoder can generate new, realistic-looking images conditioned on sampled latent vectors.
  • Model Interpretation: Latent space plots provide intuitive insights into how the model separates categories.

Challenges Faced

  • Balancing the trade-off between reconstruction loss and KL divergence.
  • Maintaining meaningful variance in a low-dimensional latent space (2D).
  • Avoiding mode collapse during training with small latent vectors.

How to Run

  1. Install Dependencies
pip install torch torchvision matplotlib numpy