Generating examples using CelebA

[ozturkoktay]

Hi, I'm trying to run code for CelebA dataset. Can you share an example like you shared for CIFAR10 (QuickStart.ipynb) ?

[HanxunH]

Hi,

I believe you can change the CIFAR10 dataset in the notebook to the CelebA on your local machine. Don't forget to change noise spatial dimension size.

[saigontrade88]

Hello,
I am interested in your paper. I am also trying to modify the QuickStart notebook to generate and evaluate the unlearnable dataset on the CelebA-50 dataset. I want to ask about the spatial dimension size for the input images and perturbation setting for the min-min noise algorithm. I currently use (the train dataset's size, 32, 32, 3) and ResNet-18 as the source and target model. The clean test accuracy when trained on the unlearnable dataset is ~90% so I suspect that I used the wrong values for these two parameters. I originally set the dimension to (128, 128, 3) and I got an error with the forward pass. Many thanks.

`
num_of_classes = 50
train_transform = [
transforms.CenterCrop((32, 32)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
]
clean_train_dataset = CelebAMini(root='datasets',
split="train",
target_type="identity",
download=False,
num_of_classes=num_of_classes,
transform=train_transform)
clean_test_dataset = CelebAMini(root='datasets',
split="all",
target_type="identity",
download=False,
num_of_classes=num_of_classes,
transform=test_transform)

noise_generator = toolbox.PerturbationTool(epsilon=0.03137254901960784,
num_steps=20,
step_size=0.0031372549019607846)

base_model = ResNet18(num_of_classes)
`
Also, based on the notebook, can you suggest on how to save the pertubed images? CelebAMini class only keeps a list of filenames. Many thanks.

[HanxunH]

Hi, thanks for your interest.

 I want to ask about the spatial dimension size for the input images and perturbation setting for the min-min noise algorithm.

I'm not quite sure what is the original resolution of CelebAMini, the noise shape should be the same as the original image.

can you suggest on how to save the pertubed images?

Here is an suggetion that should works. In the CelebAMini class, for the __getitem__(self, index) function, return the name of the file. In the for loop over the data loader, you can get filenames, images, labels, then the filenames are coorsponding to the images sampled from the dataloader. transforms.ToPILImage() can help to convert the tensor to image format and then save them.

[saigontrade88]

Thanks, you for your correct suggestions. The QuickStart Jupyter Notebook also works for ResNet-18 and CelebA.