VanillaFlow.md

Vanilla Flow

A Normalizing Flow model is designed to learn the underlying probability distribution of a given dataset. In a Normalizing Flow model, the input data is transformed through a series of invertible transformations, also known as flow steps. Each flow step consists of a deterministic function that maps the input data to a new representation, and its inverse function that maps the transformed data back to the original space. The key idea behind Normalizing Flow models is that by applying a sequence of invertible transformations, the model can learn a more complex and flexible distribution. This allows the model to capture intricate patterns and dependencies in the data. During training, the parameters of the flow steps are optimized to minimize the difference between the learned distribution and the true distribution of the data. This is typically done by maximizing the likelihood of the training data.

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	128
--n_epochs	Number of epochs	100
--lr	Learning rate	1e-3
--c_hidden	Hidden units in the first coupling layer	16
--multi_scale	Use multi scale	False
--vardeq	Use variational dequantization	False
--sample_and_save_freq	Sample and save frequency	5
--checkpoint	Checkpoint path	None
--outlier_detection	Outlier detection	False
--out_dataset	Outlier dataset name	fashionmnist	mnist, cifar10, cifar100, places365, dtd, fashionmnist, chestmnist, pneumoniamnist, tissuemnist, pneumoniamnist, svhn,tinyimagenet, imagenet
--n_layers	Number of layers	8
--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 VanFlow.py --help

Training

You can train this model with the following command:

python VanFlow.py --train --dataset pneumoniamnist

Sampling

To sample, please provide the checkpoint:

python VanFlow.py --sample --dataset pneumoniamnist --checkpoint ./../../models/VanillaFlow/VanFlow_pneumoniamnist.pt

Outlier Detection

Outlier Detection is performed by using the NLL scores generated by the model:

python VanFlow.py --outlier_detection --dataset pneumoniamnist --out_dataset mnist --checkpoint ./../../models/VanillaFlow/VanFlow_pneumoniamnist.pt