Latte: Cross-framework Python Package for Evaluation of Latent-Based Generative Models

Disentanglement learning is often used with encoder-decoder architectures to produce latent representations in the form of latent vectors or tensors in the bottleneck layer, such that each latent dimension has an approximately exclusive mapping to a semantic attribute of interest. These disentangled latent representations are particularly useful in the generative models that aim to produce samples with specific and controllable semantic attributes [3, 4]. With the growth of the fields comes the need for a reliable and consistent method of evaluation that allows for the comparison of different systems across a variety of metrics.