Bidirectional Generative Modeling Using Adversarial Gradient Estimation

This repository contains the code for the paper Bidirectional Generative Modeling Using Adversarial Gradient Estimation.

Install prerequisites

pip install -r requirements.txt

Datasets

• Synthesized Mixture of Gaussians
• Stacked MNIST (same as that used in Pac-GAN)
• CelebA
• ImageNet

Run

• Train a bidirectional generative model (BGM) using AGES-ALL on Stacked MNIST:

sh run_mog_age_kl.sh

• Train a BGM using AGES-KL on MoG:

sh run_stack_mnist_age_all.sh

• Train a BGM using AGES-ALL on CelebA:

sh run_celeba_age_all.sh

• Train a BGM using AGES-ALL on ImageNet:

sh run_imagenet.sh

• Train a BGM using AGES-KL with scaling clipping on CelebA:

sh run_celeba_age_kl_sc.sh

• Train a unidirectional model using AGES-ALL on CelebA:

sh run_celeba_uni_age_all.sh

Output

This will create a directory ./results/<dataset>/<save_name> which will contain:

• model.sav: a Python distionary containing the generator, encoder, and discriminator.
• gen.png: generated images.
• recon.png: real images (odd columns) along with the reconstructions (even columns).
• log.txt: All losses computed during training.
• config.txt: training configurations.

Help

Important arguments:

Model elements:
  --latent_dim          	dimension of the latent variable
  --prior               	prior distribution p_z of the latent variable
  --enc_dist {deterministic, gaussian, implicit}
                        	distribution of the encoder p_e(z|x) (default: gaussian)
  --dec_dist {deterministic, gaussian, implicit}
                        	distribution of the generator p_g(x|z) (default: deterministic)
                     
Objective:
  --div {all, kl, js, hellinger, revkl}
                        	use which divergence as the objective of generative modeling
  --unigen               	whether to train a unidirectional generative model (defalt: False)
  --clip               		whether to use the scaling clipping technique (defalt: False)
  --scale_lower          	lower bound of the scaling factor (default: 0.5)         
  --scale_upper          	upper bound of the scaling factor (default: None; use 1/scale_lower as the upper bound) 	

Datasets:
  --dataset {celeba, cifar, imagenet, mnist, mnist_stack, mog}
                        	name of the dataset (default: celeba)
  --data_dir          		directory of the dataset                       
  --image_size          	resolution of the image (default: 64)

Acknowledgments

The code for SAGAN architectures is based on the PyTorch implementation of SAGAN from this repository.