Fix inconsistencies in fgsm_tutorial

[QasimKhan5x]

Fixes #1032

Description

1. The architecture of the model is changed. The tutorial says that it uses the same MNIST model from pytorch/examples/mnist. However, the dimensions of the convolutional layers were different. Therefore, the exact model is copy-pasted as is.
2. The model in the MNIST example was trained using a transform in the data loader that normalized the data using mean=0.1307 and std=0.3081. However, no normalization was being applied in this example. Thus, the same normalization is applied to the test_loader in this example.
3. The state_dict of the MNIST was being loaded on the CPU even though use_cuda=True and device is being dynamically defined. Therefore, map_location=device instead of cpu
4. A denormalization function is added to unnormalize the data before applying FGSM on it, since FGSM takes unnormalized input
5. The test function is modified. After computing the gradient of the normalized data, the data is unnormalized. FGSM is then applied on unnormalized data. Finally, the perturbed data is again normalized before applying the model on it.

I also checked the code without step 5, i.e., not performing normalization and denormalization inside the test function. The observation was that the accuracy stayed the same between 98-99%, although it was supposed to fall. Applying step 5 rectified this issue.

Checklist

• The issue that is being fixed is referred in the description (see above "Fixes #ISSUE_NUMBER")
• Only one issue is addressed in this pull request
• Labels from the issue that this PR is fixing are added to this pull request
• No unnecessary issues are included into this pull request.

[netlify]

✅ Deploy Preview for pytorch-tutorials-preview ready!

Name	Link
🔨 Latest commit	eb1338d
🔍 Latest deploy log	https://app.netlify.com/sites/pytorch-tutorials-preview/deploys/648782d9aa44ac0008bd8563
😎 Deploy Preview	https://deploy-preview-2419--pytorch-tutorials-preview.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify site settings.

[QasimKhan5x]

the CI is failing because the tutorial relies on "data/lenet_mnist_model.pth", which contains the weights of the old model. I don't have access to this file, otherwise I would change it.

[malfet]

@QasimKhan5x I can update the file, but weights referenced from the tutorial are indeed for small model:

>>> import torch
>>> for k, v in torch.load("lenet_mnist_model.pth", weights_only=True, map_location='cpu').items():
...   print(k, v.shape)
conv1.weight torch.Size([10, 1, 5, 5])
conv1.bias torch.Size([10])
conv2.weight torch.Size([20, 10, 5, 5])
conv2.bias torch.Size([20])
fc1.weight torch.Size([50, 320])
fc1.bias torch.Size([50])
fc2.weight torch.Size([10, 50])
fc2.bias torch.Size([10])

I.e. your PR should include new location to download the model in both tutorial itself as well as here:

tutorials/Makefile

Line 85 in 47b9ea4

wget -nv -N https://s3.amazonaws.com/pytorch-tutorial-assets/lenet_mnist_model.pth -P $(DATADIR)

[QasimKhan5x]

@malfet I trained a model using the pytorch/examples/mnist code. Should I upload it to my own publicly accessible S3 bucket and provide its link?

[malfet]

@QasimKhan5x any public location will do, though google drive is preferable (as tutorials are most frequently run on colab, and one can connect collab to g-drive, can't they?)

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/tutorials/2419

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit eb1338d:
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[QasimKhan5x]

@malfet completed the changes