Adversarial example generation by FGSM: different normalization of training vs test images?

[hookxs]

In the Adversarial example generation tutorial the classifier from https://github.com/pytorch/examples/tree/master/mnist is used. However, this classifier is trained with input normalization transforms.Normalize((0.1307,), (0.3081,)) while in the FGSM tutorial no normalization is used and the perturbed images are clamped to [0,1] - is this not a contradiction?

[Melcfrn]

Hi,
Furthermore, dimensions in convolution layers are differents.
The file should have been changed. But the link to the model saved is good.

Clément

[QasimKhan5x]

As per my understanding, the following changes are needed in this tutorial:

1. The Net in the Adversarial example generation tutorial does not match the one in MNIST tutorial. The model definition needs to be changed.
2. The Adversarial example generation does not apply transforms.Normalize((0.1307,), (0.3081,)) in its test_loader, whereas the dataloaders in the MNIST example use it. Therefore, I should add this transform.

@svekars if I'm understanding this correctly, may I make these changes and proceed with a pull request?

[QasimKhan5x]

/assigntome