GFNOrg/torchgfn

A modular, easy to extend GFlowNet library

What it solves

It provides a standardized framework for prototyping and implementing GFlowNets (Generative Flow Networks), allowing researchers and engineers to experiment with different sampling processes, environment definitions, and loss functions without having to build everything from scratch.

How it works

The library decouples three main components: the environment (which defines states and actions), the sampling process (how states are produced), and the function approximators (the neural networks or modules used to calculate GFN loss). It includes a variety of pre-built environments (discrete and continuous), multiple loss function parameterizations (such as TB, DB, and FM), and support for diverse policy types, including recurrent, non-autoregressive, and diffusion-based policies.

Who it’s for

Researchers and engineers working with GFlowNets who need a flexible tool for developing new algorithms or reproducing published results.

Highlights

  • Support for both discrete and continuous environments.
  • Implementation of multiple GFlowNet loss types (TB, DB, SubTB, FM, ZVar, and RTB).
  • Advanced training features including off-policy training, replay buffers, and expert data warm-starting.
  • Various exploration strategies like epsilon-greedy, temperature scaling, and variance annealing.
  • Support for conditional GFlowNets and diffusion-based sampling.

Related

  • Project
    alexhernandezgarcia/gflownet

    A PyTorch-based library for training Generative Flow Networks (GFlowNets) to sample diverse, high-reward objects, primarily for applications in scientific discovery.

  • Project
    apple/pfl-research

    A Python framework for simulating private federated learning experiments, supporting multiple ML frameworks and differential privacy mechanisms.

  • Project
    AIStream-Peelout/flow-forecast

    An open-source deep learning framework for time series forecasting, classification, and anomaly detection featuring a wide array of state-of-the-art models.

  • Project
    yuanchenyang/smalldiffusion

    A lightweight diffusion library for training and sampling from diffusion and flow models, featuring a minimal core for easy experimentation with U-Net and DiT architectures.

  • Project
    SheffieldML/GPy

    A Gaussian processes framework in Python that provides tools for implementing and optimizing probabilistic models for regression and classification.