Unifying Variational Autoencoder (VAE) implementations in Pytorch (NeurIPS 2022)

An semi-supervised extension based on VAE for Regression, demonstrate its performance on two soft sensor benchmark problems.

Autocurator is a comprehensive benchmarking toolkit for evaluating synthetic tabular data. It measures fidelity, coverage, privacy, and utility through quantitative metrics, visual reports, and PCA/correlation diagnostics. Ideal for validating VAE, GAN, Copula, or Diffusion-generated datasets.

Benchmarking embedding methods (UMAP, VAE, PCA, FA, ICA, etc.) for survival prediction on omics data with TabNet, CatBoost and ridge models.

No description available

Research VQ-VAE with post-hoc geodesic quantization. Train a VAE, build a codebook via geodesic K-medoids using decoder-induced Riemannian distances, then model codes with a Transformer. Scripts/configs for reproducible pipelines and Euclidean vs geodesic benchmarks on MNIST/FashionMNIST/CIFAR-10.

No description available

Running benchmarks for vae latent space structure

VAE implementation and Using a DNN classifier to benchmark the generated VAE images for MNIST

This project aims to provide hands-on experience with three major classes of generative models — Generative Adversarial Networks , Variational Autoencoders , and Diffusion Models. Students will implement simplified versions of each model, train them on image datasets and compare their generated samples, training dynamics, and evaluation metrics.

Benchmarking suite for generative models. Currently supports VAE, GAN, and DDPM on the MNIST dataset (28×28).

VAE-based reconstruction of orderbook states (volumes and prices). Experiments with MLP, Conv1D, hybrid architectures, and ensemble methods. Includes detrending, log-cumulative volumes, and comprehensive benchmarks.

Implementation of benchmarks and deep learning frameworks (Me-VAE, Online Self-Learning, Toolbox) for EEG artifact removal. Includes MATLAB & Python codes, evaluation metrics, and reproducible pipelines from my MSc dissertation project.

A production-quality PyTorch implementation of a Beta-Variational Autoencoder (VAE) for unsupervised anomaly detection. Features explicit ELBO loss calculation, custom data generation, and rigorous benchmarking (AUC-ROC) against baseline Autoencoders

Design and prototype two advanced anomaly detection approaches for fraud detection — a Graph Neural Network (GNN) model and a Variational Autoencoder (VAE) model — and benchmark them against a traditional gradient-boosted baseline.

This project implements multiple deep learning models for anomaly detection, including CAE, VAE, GANomaly, PaDiM, and Perceptual Autoencoders. It provides training scripts, evaluation metrics for benchmarking model performance on datasets like MNIST and CIFAR.

In this notebook, a Variational Autoencoder (VAE) is created and used to generate synthetic data of the minority class of a credit loan dataset. A classification network is then trained based on both the original and synthetic data, and the performance is evaluated, using the AUC. Two benchmark models are implemented to assess the performance of the VAE.

A deep learning project for generating handwritten digits using Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). Trained on MNIST, this repository includes optimized architectures, KL annealing, mixed precision, and benchmarking insights. GAN achieves 88% image acceptability.

VAE Benchmark

To be used with benchmark_vae

Benchmarking multiple VAE methods

No description available

No description available

Benchmarking multiple method for property prediction optimisation in a VAE like architecture

No description available

No description available

No description available

No description available

MSc Dissertation Codes

F16 Benchmark Data Analysis