navigation links

weighted_dataset package

This training module provides functions to train invertible flow models on datasets drawn from a known distribution, which does not correspond to the target of our flow model, which we know as a black box function

In other words, we have a set of (x,p(x),q(x)), such that - x lives in the target space - x ~ p(x) - q is a black box function - we want to train a flow model F such that F(z) ~ q(F(z)) when we sample z in the latent space

Note that typical cases where a flow is trained on an experimental dataset - q = p - p is unknown In which case one can maximize the likelihood of the dataset under the flow model.

Submodules

• dkl_training module
• generators module
• stateful_trainer module
• training_record module
• variance_training module
• weighted_dataset_trainer module