v3.1.3

• Changed jnp.trapz (which has been removed) to jax.scipy.integrate.trapezoid.
• Removed upper-limits on version numbers for most dependencies
• Increased python version to a minimum of 3.9

v3.1.2

Changed jnp.NINF to -jnp.inf as the former has been removed from Jax.

v3.1.1

• Fixed typo in info for CentBeta13 which prevented saving and then loading a flow with this distribution

v3.1.0

• Added validation loss tracking to the training method
• After training, rather than saving the parameters from the final epoch of training, the parameters from the epoch with the best loss are saved. If a validation set is provided, this reference loss is the validation loss. If not, the reference loss is the training loss.

v3.0.3

• Fixed overflow bug on Uniform distribution log_prob when number of dimensions is greater than 10
• Updated dependencies

v3.0.2

Added -> None return signatures to functions and methods that don't return an object.

v3.0.0 / v3.0.1

Release of PZFlow 3.0.0 (the 0.0.1 was just to fix the README on PyPI and doesn't reflect any changes to the code)

Changes

• Made Uniform the default distribution
• Made all distributions symmetric about zero
• Moved to Optax for optimization
• Added default bijectors to Flow and FlowEnsemble.

Features

• Added early stopping to train methods
• Added progress bar to train methods
• Added CentBeta13 distribution, which looks like a Guassian with hard cutoffs.

Bugs

• Fixed bug where FlowEnsemble metadata was not available from the ensemble itself, only from the individual flows in the esemble

v2.0.7

• Made some fixes so that pzflow works with the latest version of jax
• I had to change the Uniform Dequantizer so that you always have to specify which columns to dequantize

v2.0.6

Changes for last few bug fixes:
2.0.4 - fixed conditional sampling for FlowEnsembles so that it is obvious which sample belongs to which condition.
2.0.5 - fixed conversion of samples to pandas DataFrames, so that the dtype is float rather than object
2.0.6 - fixed error in CentBeta distribution where output shape was accidentally printed whenever calling log_prob

v2.0.3

Fixed error where marginalization yields NaNs when combined with error convolution.