kamenbliznashki/normalizing_flows

Provides reference implementations for training normalizing flow models on both synthetic 2D datasets and real image/tabular data (CelebA, MNIST, UCI), with support for multi-GPU distributed training and gradient checkpointing for large models. Includes generative capabilities beyond density estimation, such as temperature-controlled sampling, latent space attribute manipulation, and conditional density estimation. Built entirely in PyTorch with modular architecture allowing interchange of flow components and straightforward integration into probabilistic inference pipelines.

637 stars. No commits in the last 6 months.