optim

class trojanzoo.optim.Optimizer(iteration=20, stop_threshold=None, loss_fn=None, **kwargs)

An abstract input optimizer class that inherits trojanzoo.utils.module.Process.

Parameters:

• iteration (int) – Optimization iteration. Defaults to 20.

• stop_threshold (float | None) – Threshold used in early stop check. Defaults to None (no early stop).

• loss_fn (Callable) – Loss function (it’s usually reduction='none').

• **kwargs – Keyword Arguments passed to trojanzoo.utils.module.Process.

early_stop_check(*args, current_idx=None, adv_input=None, loss_values=None, loss_fn=None, stop_threshold=None, loss_kwargs={}, **kwargs)

Early stop check using stop_threshold.

Parameters:

• current_idx (torch.Tensor) – The indices of adv_input need to check (Other indices have early stopped).

• adv_input (torch.Tensor) – The entire batched adversairl input tensor with shape (N, *).

• loss_values (torch.Tensor) – Batched loss tensor with shape (N). If None, use loss_fn and adv_input to calculate. Defaults to None.

• loss_fn (collections.abc.Callable | None) – Loss function (it’s usually reduction='none'). Defaults to self.loss_fn.

• stop_threshold (float | None) – Threshold used in early stop check. None means using self.stop_threshold. Defaults to self.stop_threshold.

• loss_kwargs (dict[str, torch.Tensor]) – Keyword arguments passed to loss_fn, which will also be selected according to current_idx.

• *args – Any positional argument (unused).

• **kwargs – Any keyword argument (unused).

Returns:

torch.Tensor – Batched torch.BoolTensor with shape (N).

optimize(_input, *args, iteration=None, loss_fn=None, stop_threshold=None, output=None, **kwargs)

Main optimize method.

Parameters:

• _input (torch.Tensor) – The batched input tensor to optimize.

• iteration (int | None) – Optimization iteration. Defaults to self.iteration.

• loss_fn (Callable) – Loss function (it’s usually reduction='none'). Defaults to self.loss_fn.

• stop_threshold (float | None) – Threshold used in early stop check. None means using self.stop_threshold. Defaults to self.stop_threshold.

• output (int | Iterable[str]) – Output level integer or output items. If int, call get_output_int(). Defaults to self.output.

Returns:

(torch.Tensor, torch.Tensor) – batched adversarial input tensor and batched optimization iterations (-1 if not reaching self.threshold).

output_info(*args, mode='start', _iter=0, iteration=0, output=None, indent=None, **kwargs)

Output information.

Parameters:

• mode (str) – The output mode (e.g., 'start', 'end', 'middle', 'memory'). Should be legal strings in get_output_int(). Defaults to 'start'.

• _iter (int) – Current iteration. Defaults to 0.

• iteration (int) – Total iteration. Defaults to 0.

• output (Iterable[str]) – Output items. Defaults to self.output.

• indent (int) – The space indent for the entire string. Defaults to self.indent.

• *args – Any positional argument (unused).

• **kwargs – Any keyword argument (unused).

preprocess_input(*args, adv_input=None, org_input=None, **kwargs)

Optimize input tensor for 1 iteration.

Parameters:

• adv_input (torch.Tensor) – The entire batched adversairl input tensor with shape (N, *).

• org_input (torch.Tensor) – The entire batched original input tensor with shape (N, *).

abstractmethod update_input(current_idx, adv_input, org_input, *args, **kwargs)

Optimize input tensor for 1 iteration.

Parameters:

• current_idx (torch.Tensor) – The indices of adv_input need to optimize (Other indices have early stopped).

• adv_input (torch.Tensor) – The entire batched adversairl input tensor with shape (N, *).

• org_input (torch.Tensor) – The entire batched original input tensor with shape (N, *).