Release notes

Release notes#

These notes relate the changes between different JaxILI releases.

0.1.3#

  • Fix the save of checkpoints with the new orbax API.

  • Add functions to load the checkpoints when using the inference class NPE or NLE.

  • Fixed bugs related to issues when data points have zero or close to zero variance.

  • Moved classes Identity and Standardizer to compressor.py.

0.1.2#

  • Remove torch.utils.data.DataLoader causing kernel crash. Replaced with DataLoaders from jax_dataloader.

  • Add example notebooks to the API documentation.

  • Dependencies of JaxILI have been updated to avoid conflicts between jax and tensorflow with nvidia packages.

0.1.1#

  • Update of the API documentation.

0.1#

  • Added different classes to have user-friendly calls to the models and trainer implemented in the pre-release.

    • Classes in jaxili.inference allow to create an object to train a model for Neural Posterior Estimation (NPE) or Neural Likelihood Estimation (NLE).

    • Classes in jaxili.posterior provide the abstract class to sample from the posterior and evaluate the log-posterior of the trained model. Currently, the DirectPosterior for NPE is implemented and the MCMCPosterior for NLE uses numpyro to run gradient-based MCMC algorithms.

  • Future updates will add other samplers and add utilities to perform stacking of posteriors.

0.0.1 (Pre-release)#

  • First prerelease of JaxILI.

  • The package contains the module to create and train Normalizing Flows (e.g. Masked Autoregressive Flows and RealNVP). They are contained in jaxili.model and can be trained with the TrainerModule in jaxili.train.

  • JaxILI provides parent classes to use other types of Neural Density Estimator. Likewise, a class can inherit from the TrainerModule to define different loss functions, traning loop, etc…

Contents