README.md

Experimental Setup

Notation

Herein, the notation finopt referes to optimizing the variational parameters at train time and none refers to regular training of the models

Workflow

• Call python train.py with a variety of options -ds 200 (stochastic dimensions), -otype finopt (optimize variational parameters) and -ns 200 (with 200 updates using ADAM) to train models
• Checkpoints and model parameters will be saved to folders with format chkpt-<datasetname> where the save frequency may be specified using the flag as -sfreq 10

Code for setting up experiments

• setupExperiments.py
  • Use this with different options (see inside file) for setting up different experiments on different datasets
  • The pre-specified options are settings used in the paper
• evaluate_timing.py
  • Get the average per batch runtimes (saved in checkpoint files) for different datasets and optimization scheme
• evaluate_jac.py
  • Evaluate the Jacobian matrix for different models
• evaluate_finopt_table.py
  • Print the table comparing models trained with and without optimizing psi
• train.py
  • Main training script for DLGMs which accepts a variety of arguments specified in parse_args.py