Fixing minor bugs and typos so that the main runs

mit-han-lab / dlg

[NeurIPS 2019] Deep Leakage From Gradients

https://dlg.mit.edu/

MIT License

396 stars 105 forks source link

Closed jeandut closed 4 years ago

jeandut commented 4 years ago

There are a few bugs in the code on master:

the output layer of the LeNet defined in models/vision is of size 1000 instead of 100
one import is missing in utils: torch.nn.functional as F
I took the liberty to add some seeding

jeandut commented 4 years ago

Also added a plt.show for the ones with funny matplotlib backends.

jeandut commented 4 years ago

To have it completely reproducible on GPU one needs to add:

torch.backends.cudnn.deterministic = True  

torch.backends.cudnn.benchmark = False

I did not want to add them to the PR for performance issues

jeandut commented 4 years ago

Glad I could help