The result of using the Wasserstein attack is the "nan" value.

[iunyu]

I was generating adversarial inputs to various attacks using ART.
when I was using Wasserstein attack, An unexpected warning occurs and the result is output as a nan value.

This warning is as follows...

...site-packages/art/attacks/evasion/wasserstein.py:489: RuntimeWarning: invalid value encountered in log 
alpha = np.log(self._local_transport(var_k, exp_beta, self.kernel_size)) - np.log(x_init)
.../site-packages/art/attacks/evasion/wasserstein.py:497: RuntimeWarning: invalid value encountered in greater
beta[beta > 1e-10] = np.real(lambertw(beta[beta > 1e-10]))

it is generating attacks on a resnet50 model and the dataset is cifar10
and I have used the parameter of the attack described in the #776 bug report.

attack = Wasserstein(estimator=classifier,
        targeted=False,
        regularization=3000.0,
        p=2,
        kernel_size=5,
        eps_step=0.001,
        norm="inf",
        ball="wasserstein",
        eps=0.001,
        eps_iter=10,
        eps_factor=1.17,
        max_iter=400,
        conjugate_sinkhorn_max_iter=400,
        projected_sinkhorn_max_iter=400,
        batch_size=1,
        verbose=True)

- art version: 1.8.0

Why am I getting this error?
thank you.

[beat-buesser]

Hi @iunyu Thank you very much for using ART! Could you please provide more details about your experiment, maybe in a Google Colab notebook?

[iunyu]

First, thank you for your prompt reply!

these are my full code.

import torchvision.transforms as transforms
import sys
sys.path.append('./pytorch-cifar')
sys.path.append('./pytorch-cifar/models') # clone https://github.com/kuangliu/pytorch-cifar.git
from models import *
import torch.optim as optim
from torchsummary import summary
from art.attacks.evasion import Wasserstein
from art.estimators.classification import PyTorchClassifier
import torch.nn.functional as F
import torchvision
import numpy as np

GPU_NUM = 1
device = torch.device(f'cuda:{GPU_NUM}' if torch.cuda.is_available() else 'cpu')
torch.cuda.set_device(device) # change allocation of current GPU

print('==> Preparing data..')
transform_test = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, shuffle=False, num_workers=2, batch_size=100)

def main():
    print('==> Building model..')
    net = ResNet50()
    summary(net.cuda(), (3, 32, 32))

    net = net.to(device)
    checkpoint = torch.load('./checkpoint/ckpt_resnet50.pth') # it is ths pretrained model RESNET50
    net.load_state_dict(checkpoint['net'])
    net.eval()

    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4)
    classifier = PyTorchClassifier(
        model=net,
        loss=criterion,
        optimizer=optimizer,
        input_shape=(3,32,32),
        nb_classes=10,
    )

    result = torch.zeros([1, 3, 32, 32])
    attack = Wasserstein(estimator=classifier,
                         targeted=False,
                         regularization=3000.0,
                         p=2,
                         kernel_size=5,
                         eps_step=0.001,
                         norm="inf",
                         ball="wasserstein",
                         eps=0.001,
                         eps_iter=10,
                         eps_factor=1.17,
                         max_iter=400,
                         conjugate_sinkhorn_max_iter=400,
                         projected_sinkhorn_max_iter=400,
                         batch_size=1,
                         verbose=True)

    for batch_idx, (data, target) in enumerate(testloader):
        data = data.numpy()
        x_test_adv = attack.generate(x=data)
        result = torch.cat([result, x_test_adv], dim=0)

    np.save('result/10_WS_input',result)

if __name__ == '__main__':
    main()

[beat-buesser]

Great, thank you. From where do you get the checkpoint?

[iunyu]

I'm new to this research, sorry for the inconvenience :(

checkpoint = torch.load('./checkpoint/ckpt_resnet50.pth') # it is ths pretrained model RESNET50
The file used in this research is as follows.

https://github.com/iunyu/dnn_model/blob/91323a5af5dae7801828452664307e14e371fc02/ckpt_resnet50.pth

Thank you!

[billbradley]

Was there ever a resolution to this issue? I am also getting NaNs (from a finetuned Resnet34 model).