pretrain in defense about trojanzoo

[ain-soph]

I’ll close this issue if there is no further question,

[ain-soph]

It's unnecessary because the model saved by previous attack is loaded. You can just remove --pretrained in your command.

I'll consider removing it to make things clean.

You may close this issue if you have no further concern.

trojanzoo/trojanvision/defenses/abstract.py

Lines 53 to 56 in caad6ed

	if not self.original:
	self.attack.load(**kwargs)
	if isinstance(self.attack, TrojanNet):
	self.model = self.attack.combined_model

[programehr]

Thank you. I'll close it.

[programehr]

It's unnecessary because the model saved by previous attack is loaded. You can just remove --pretrained in your command.

I'll consider removing it to make things clean.

You may close this issue if you have no further concern.

trojanzoo/trojanvision/defenses/abstract.py

Lines 53 to 56 in caad6ed

	if not self.original:
	self.attack.load(**kwargs)
	if isinstance(self.attack, TrojanNet):
	self.model = self.attack.combined_model

Hello again,

Excuse me, but I think it does have an effect. It loads the (possibly trojaned) model only if you specify the pre-train option. Otherwise, it will create a model from scratch.

Am I right?

[ain-soph]

No, --pretrained loads the clean pretrained model at model initialization, while defense.detect loads the previously saved attack result model.

if --original is claimed, then defense.detect will not load the attack model.