DCGAN.md

Deep Convolutional Generative Adversarial Network (DC-GAN)

A Generative Adversarial Network (GAN) comprises two neural networks: a Generator and a Discriminator, engaged in a minimax game. The generator fabricates synthetic images out of a noisy input, while the discriminator evaluates the authenticity of these samples, distinguishing between real data and the generated ones. Through iterative training, the generator learns to produce increasingly realistic outputs that deceive the discriminator, while the discriminator enhances its ability to differentiate genuine from fake data.

Parameters

Argument	Description	Default	Choices
--train	Train model	False
--sample	Sample from model	False
--outlier_detection	Outlier detection	False
--batch_size	Batch size	128
--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
--n_epochs	Number of epochs	100
--lrg	Learning rate generator	0.0002
--lrd	Learning rate discriminator	0.0002
--beta1	Beta1	0.5
--beta2	Beta2	0.999
--latent_dim	Latent dimension	100
--img_size	Image size	32
--channels	Channels	1
--sample_and_save_freq	Sample interval	5
--checkpoint	Checkpoint path	None
--discriminator_checkpoint	Discriminator checkpoint path	None
--n_samples	Number of samples	9
--d	d	128
--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 DCGAN.py --help

Training

Adversarial losses are used during training. The generator is encouraged to generate images that fool the discriminator into classifying them as real, while the discriminator is trained as a binary classifier to distinguish between real and generated images. The model can be trained using the following command:

python DCGAN.py --train --dataset svhn

Sampling

To sample from a GAN, you input a noisy latent vector of a predefined size into the generator network. This latent vector serves as a random seed that the generator uses to generate synthetic data samples.

python DCGAN.py --sample --dataset svhn --checkpoint ./../../models/DCGAN/DCGAN_svhn.pt

Outlier Detection

To perform outlier detection, only the Discriminator will be used:

python DCGAN.py --outlier_detection --dataset svhn --out_dataset cifar10 --discriminator_checkpoint ./../../models/DCGAN/DCDisc_svhn.pt