Zuko 1.1.0

✨ What's new

• New VAE tutorial using the MNIST dataset (8812e04)
• Add support for unconditional univariate flows (#34)
• New Bernstein polynomial flow (#32 and #33) by @oduerr and @MArpogaus

Full Changelog: 1.0.0...1.1.0

Zuko 1.0.0

💥 Breaking news

Zuko's repository has been transferred to the probabilists organization.

♻️ Refactor

To facilitate the maintenance of existing flows and the addition of new flows, the zuko.flows module has been refactored into many submodules (#24). As part of this refactor, the DistributionModule and TransformModule have been renamed to LazyDistribution and LazyTransform to clarify the purpose of these classes and avoid confusion with Pyro components.

✨ What's new

• New general coupling transformation (#23) by @simonschnake
• New gaussianization flow (f7e4f85)
• New tolerance parameters for odeint (bc4323f)
• Use independent networks in neural transformations (c600dad)
• New tutorials in the documentation (#27)

Full Changelog: 0.2.0...1.0.0

Zuko 0.2.0

💥 Breaking news

We decided to drop the support for older PyTorch versions (1.11 and below) as it made it much more difficult to implement some features, especially ones relying on automatic differentiation. PyTorch 1.12 has been around for about a year and is compatible with all CUDA drivers since 10.2.

✨ What's new

• Add rsample_and_log_prob method to generate samples with their log-probability (#18, #19)
• Add LU-factorized linear transformation inspired by nflows's LULinear (9c74233)

🐛 Bug fixes

• Fix _call not implemented error call_and_ladj (#14) by @felixdivo

Full Changelog: 0.1.4...0.2.0

Zuko 0.1.4

First official release of Zuko 🥳

✨ What's new

• Refactor NormalizingFlow class to improve log_prob efficiency (a755a69, 3e9c68d)
• Documentation hosted on Read the Docs (882b821)
• New continuous normalizing flow (CNF) (a755a69)
• New neural circular spline flow (NCSF) (#12)
• New Gaussian mixture model (GMM) (995efc2)

Full Changelog: 0.0.6...0.1.4