Steganografiya quyidagi sohalarda qo'llaniladi, lekin ular bilan cheklanmaydi

loss.backward(retain_graph=True)
self.optimizer.step()

def watermark_loss(self):
"""Compute loss on watermark.

Returns:
wmloss (float): loss on watermark data

"""
wmloss = None
for _, data in enumerate(self.triggerloader):
inputs, labels = data
outputs = self.model(inputs.to(self.device))
if not wmloss:
wmloss = self.args.criterion(outputs, labels.to(self.device))
else:
wmloss += self.args.criterion(outputs, labels.to(self.device))
return wmloss

def get_ownership(self):
"""Compute ownership information

Returns:_ownership_(dict):_information_on_trigger_data__"""_ownership_=_{}_WM_X,_WM_y_=_[],_[]'>Returns:
ownership (dict): information on trigger data

"""
ownership = {}
WM_X, WM_y = [], []
for _, data in enumerate(self.triggerloader):
inputs, labels = data
WM_X += list(inputs.cpu().detach().numpy())
WM_y += list(labels.cpu().detach().numpy())

ownership['inputs'] = WM_X
ownership['labels'] = WM_y
ownership['bounds'] = (min(WM_y), max(WM_y))

return ownership

def train(self):
"""Train the model on watermarked data.

Args:
trainloader (object): training data
epochs (int): number of epochs

Returns:
ownership (dict): Ownership information (None
if model not watermarked)

"""
self.model.to(self.device)
logger.info('Training')
pbar = tqdm.tqdm(range(self.args.epochs),
disable=not self.args.verbose)
for _ in pbar:
for idx, data in enumerate(self.trainloader):
X, y = data
self.train_step(X, y, idx)