real_world

Experiments on Real-World Datasets

This folder contains the code implementation of the experiments presented in Sections 5.2, 5.3 and 5.4 of the paper Training Energy-Based Normalizing Flow withScore-Matching Objectives.

Setup

Installing the ebflow package allows you to run experiments with the ebflow command. Conduct the following instruction at the root of this directory to initiate the installation:

pip install -e .

Usage

Training Process

Use commands with the following format to train a model:

ebflow --config {$(1)} --loss {$(2)} --restore_path {$(3)}

• (1) config: training configuration (format: {dataset}_{architecture}).
• (2) loss: objective function (i.e., ml, sml, ssm, dsm, fdssm).
• (3) restore_path: the path to a checkpoint form which the training process resumes.

Examples:

• (Results in Table 2)

  • Train EBFlow on MNIST with the FC-based architecture and the SSM objective.

  ebflow --config 'mnist_fc' --loss 'ssm'
  

  • Train EBFlow on CIFAR-10 with the CNN-based architecture and the DSM objective.

  ebflow --config 'cifar_cnn' --loss 'dsm'
  

• (Results in Table 4) Train EBFlow on MNIST with the FC-based architecture and the SSM objective.

  ebflow --config 'mnist_fc' --loss 'ssm' --withoutMaP
  

• (Results in Figure 4)

  • Train EBFlow on MNIST with the Glow architecture and the SSM / DSM objective.

  ebflow --config 'mnist_glow' --loss 'ssm'
  ebflow --config 'mnist_glow' --loss 'dsm'
  

  • Generate samples using a pretrained Glow model.

  ebflow --config 'mnist_glow_inv' --restore_path 'results/mnist_glow_ssm/checkpoints/checkpoint_best.tar'
  ebflow --config 'mnist_glow_inv' --restore_path 'results/mnist_glow_dsm/checkpoints/checkpoint_best.tar'
  

  The Glow model can be trained more efficiently through DSM without compromising the performance (in comparison to SSM).

• (Results in Figure 5)

  • Train EBFlow on CelebA / STL-10 with the FC-based architecture and the DSM objective.

    ebflow --config 'celeb_fc' --loss 'dsm'
    ebflow --config 'stl_fc' --loss 'dsm'
    

  • Perform inplainting using a pretrained FC-based model.

    ebflow --config 'celeb_fc_mcmc' --restore_path 'results/celeb_fc_dsm/checkpoints/checkpoint_best.tar' --loss 'dsm'
    ebflow --config 'stl_fc_mcmc' --restore_path 'results/stl_fc_dsm/checkpoints/checkpoint_best.tar' --loss 'dsm'
    

  Use --restore_path {path_to_checkpoint} to resume the training from a prior checkpoint if the training accidentally terminates.

  Download CelebA dataset through [link].

Details about the Code Implementation

• Models are built using the FlowSequential module (i.e. ebflow/layers/flowsequential.py). Each FlowSequential module contains a number of flow layers (i.e., FlowLayer and LinearFlowLayer), and has the following functions: forward, reverse, log_prob, sample.
• All of the flow layers belong to either FlowLayer or LinearFlowLayer. Layers with constant Jacobians are categorized as LinearFlowLayer, while the others are categorized as FlowLayer.
• The Experiment class (i.e., ebflow/train/experiment.py) handles training, evaluation, sampling, and checkpoint saving.