Skip to content

Tiny Fashion VAE — Learning Latent Spaces

A completed educational study of convolutional autoencoders and variational autoencoders on Fashion-MNIST, exploring reconstruction, generation and latent-space trade-offs.

All projects
Tiny Fashion VAE — Learning Latent Spaces

Generated sample grid: Prashanna Tiwari · Author-provided project image · Resized and converted to WebP.

Created by Prashanna Tiwari.

Exploring small latent spaces

Tiny Fashion VAE implements convolutional autoencoders and variational autoencoders with PyTorch. Its training, evaluation and visualization tools explore reconstruction, prior sampling and interpolation between clothing images in Fashion-MNIST.

The completed study contains 14 experiment runs, including three training seeds for the core autoencoder, VAE and KL warm-up comparison. Additional experiments vary regularization strength and latent dimension. Saved metrics, reproducible figures, checkpoint hashes and a mathematical report document the work.

The core implementation passed 14 tests, all 14 study checkpoint hashes were verified, and CPU image generation was exercised during review. This is an educational empirical study on one dataset, with limited training seeds and some blurred generated samples.

The cover shows an unfiltered grid of model-generated decoder means. The source repository and checkpoints remain private.

Contributors

  • Prashanna Tiwari

Resources & links