problem about saving the intermediate results and config problem

[andyclsr]

hi, your work is excellent but i encounter some problems.
first , i read similar problems #153 and
your https://ain-soph.github.io/trojanzoo/trojanvision/configs.html ,but constructions about configs are hard to understand and the example you provide is limited like this : https://github.com/ain-soph/trojanzoo/blob/main/trojanzoo/configs/dataset.yml.

my problem is i don't know how to store intermediate results and i think it is my config problem
i do something like this : and for example attack.yml is only " attack_dir: /data/attack/"
and my file structure:

my main.py :

i still wonder about the use of config file and it seem like nothing was stored.
i hope you could give a concrete example about the use of it. thx anyway

[andyclsr]

can concretely tell me how to store the attacked model after this step:

[ain-soph]

attack_dir sets the folder_path of Attack instance, which is attack.folder_path. This property is used in many places including saving method (see following)

trojanzoo/trojanvision/attacks/abstract.py

Lines 110 to 112 in 1f417cc

	def attack(self, epochs: int, **kwargs):
	kwargs['validate_fn'] = kwargs.get('validate_fn', self.validate_fn)
	kwargs['save_fn'] = kwargs.get('save_fn', self.save)

trojanzoo/trojanvision/attacks/abstract.py

Lines 185 to 192 in 1f417cc

	def save(self, filename: str = None, **kwargs):
	r"""Save attack results to files."""
	filename = filename or self.get_filename(**kwargs)
	file_path = os.path.join(self.folder_path, filename)
	np.save(file_path + '.npy', self.mark.mark.detach().cpu().numpy())
	F.to_pil_image(self.mark.mark).save(file_path + '.png')
	self.model.save(file_path + '.pth')
	print('attack results saved at: ', file_path)

The save_fn is passed to model._train method and used in

trojanzoo/trojanzoo/utils/train.py

Lines 201 to 203 in 1f417cc

	if save:
	save_fn(file_path=file_path, folder_path=folder_path,
	suffix=suffix, verbose=verbose)

---

This structure is to save model during training intervals. If you simply want to save model in your main script after training, you can always use model._model, which is a standard torch.nn.Module. You can use torch.save(model._model.state_dict(), save_path)

[ain-soph]

To control the interval to save models, use validate_interval parameter (command-line argument as well)

trojanzoo/trojanzoo/utils/train.py

Line 180 in 1f417cc

if validate_interval != 0 and (_epoch % validate_interval == 0 or _epoch == epochs):

trojanzoo/trojanzoo/trainer.py

Line 123 in 1f417cc

group.add_argument('--validate_interval', type=int,

[andyclsr]

To control the interval to save models, use validate_interval parameter (command-line argument as well)

trojanzoo/trojanzoo/utils/train.py

Line 180 in 1f417cc

if validate_interval != 0 and (_epoch % validate_interval == 0 or _epoch == epochs):

trojanzoo/trojanzoo/trainer.py

Line 123 in 1f417cc

group.add_argument('--validate_interval', type=int,

what if i want to save multiple models? not just updating one model.

[ain-soph]

@andyclsr You have 2 options:

1. pass your own save_fn to model._train in attack.attack
2. overload save_fn in your own Attack class

trojanzoo/trojanvision/attacks/abstract.py

Lines 110 to 112 in 1f417cc

	def attack(self, epochs: int, **kwargs):
	kwargs['validate_fn'] = kwargs.get('validate_fn', self.validate_fn)
	kwargs['save_fn'] = kwargs.get('save_fn', self.save)

[andyclsr]

ok I know it. really thank you for your generous explanations! wish you a good day!