FlowPlusPlus.md

Flow++

Flow++ is a generative model that aims to learn the underlying probability distribution of a given dataset. This work improves upon three limiting design choices employed by flow-based models in prior work: the use of uniform noise for dequantization, the use of inexpressive affine flows, and the use of purely convolutional conditioning networks in coupling layers.

Parameters

Argument	Description	Default	Choices
--train	Train model	False
--sample	Sample from model	False
--outlier_detection	Outlier detection	False
--dataset	Dataset name	mnist	mnist, cifar10, fashionmnist, chestmnist, octmnist, tissuemnist, pneumoniamnist, svhn, tinyimagenet, cifar100, places365, dtd, imagenet
--no_wandb	Disable Wandb	False
--out_dataset	Outlier dataset name	fashionmnist	mnist, cifar10, fashionmnist, chestmnist, octmnist, tissuemnist, pneumoniamnist, svhn, tinyimagenet, cifar100, places365, dtd, imagenet
--batch_size	Batch size	8
--n_epochs	Number of epochs	100
--lr	Learning rate	1e-3
--warm_up	Warm up	200
--grad_clip	Gradient clip	1.0
--drop_prob	Dropout probability	0.2
--num_blocks	Number of blocks	10
--num_components	Number of components in the mixture	32
--num_dequant_blocks	Number of blocks in dequantization	2
--num_channels	Number of channels in Flow++	96
--use_attn	Use attention	False
--sample_and_save_freq	Sample interval	5
--checkpoint	Checkpoint path to VQVAE	None
--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 FlowPP.py --help

Training

You can train this model with the following command:

python FlowPP.py --train --dataset mnist

Sampling

To sample, please provide the checkpoint:

python FlowPP.py --sample --dataset fashionmnist --checkpoint ./../../models/FlowPP/FlowPP_mnist.pt

Outlier Detection

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

python FlowPP.py --outlier_detection --dataset mnist --out_dataset fashionmnist --checkpoint ./../../models/FlowPP/FlowPP_mnist.pt