NeuralDecomposition

This is a PyTorch implementation of our AISTATS paper Neural Decomposition: Functional ANOVA with Variational Autoencoders (Märtens & Yau, 2020)

Example Colab notebook

See the Colab noteboook for an example on synthetic data:

• open in GitHub
• open in Colab

Implementation

• decoder.py impements the class for the decomposable CVAE decoder, for the special case of a one-dimensional latent variable z and one-dimensional covariate c.
• encoder.py implements the class for a standard CVAE encoder
• CVAE.py provides a wrapper to combine the decoder and encoder for training purposes
• Some comments on implementation details regarding integral penalties:
  • When initialising the decoder, we provide the input grids on which the integrals are evaluated using quadrature
  • The penalty terms c and lambda are initialised with lambda0
  • If one chooses to replace a fixed constant c with a sequence of c values (this can lead to faster convergence), this can be done during optimisation via specifying lambda_start and lambda_end values to CVAE optimize().
  • Feature-level sparsity can optionally be turned on or off via has_feature_level_sparsity