There is a discrepancy when using automatic differentiation to solve the attacker's optimization problem between:
1) the official implementation of ResNet18 from Pytorch (as imported by from torchvision.models import resnet18)
and
2) the customer implementation of ResNet18 as in https://github.com/marsplus/SG-Unlearn/blob/Zonglin/src/resnet.py
There is a discrepancy when using automatic differentiation to solve the attacker's optimization problem between: 1) the official implementation of ResNet18 from Pytorch (as imported by
from torchvision.models import resnet18
) and 2) the customer implementation of ResNet18 as in https://github.com/marsplus/SG-Unlearn/blob/Zonglin/src/resnet.pySpecifically, this line fails: https://github.com/marsplus/SG-Unlearn/blob/14453a8d3e2b51f98f7bf8619a4de6815c465202/src/models.py#L715