HierarchicalVAE.md

Hierarchical Variational Autoencoders (Hierarchical VAEs)

NVAE introduced a deep hierarchical VAE designed for image generation through depth-wise separable convolutions and batch normalisation. NVAE has a residual parameterization for normal distributions and uses spectral regularisation to stabilise training.

Parameters

Argument	Description	Default	Choices
--train	Train model	False
--sample	Sample model	False
--dataset	Dataset name	mnist	mnist, cifar10, fashionmnist, chestmnist, octmnist, tissuemnist, pneumoniamnist, svhn, tinyimagenet, cifar100, places365, dtd, imagenet
--no_wandb	Disable Wandb	False
--batch_size	Batch size	256
--n_epochs	Number of epochs	100
--lr	Learning rate	0.01
--latent_dim	Latent dimension	512
--checkpoint	Checkpoint path	None
--sample_and_save_freq	Sample and save frequency	5
--num_workers	Number of workers for Dataloader	0

You can find out more about the parameters by checking util.py or by running the following command on the example script:

python HVAE.py --help

Training

HVAE can be trained similarly to other models in the Zoo:

python HVAE.py --train --dataset svhn

Sampling

For sampling you must provide the HVAE checkpoint:

python HVAE.py --sample --dataset svhn --checkpoint ./../../models/HierarchicalVAE/HVAE_svhn.pt