Skip to content

Modifying DPLossFastGradientClipping to add support for generative tasks with ghost clipping (#716) #722

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 1 commit into from
Closed

Modifying DPLossFastGradientClipping to add support for generative tasks with ghost clipping (#716) #722

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 20 additions & 6 deletions opacus/utils/fast_gradient_clipping_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,7 +68,9 @@ def backward(self):
reduced_loss.backward(retain_graph=True)
self.optimizer.zero_grad()
coeff = self.module.get_clipping_coef()
second_loss_per_sample = coeff * self.loss_per_sample
second_loss_per_sample = (
coeff.to(self.loss_per_sample.device) * self.loss_per_sample
)
second_loss = torch.sum(second_loss_per_sample)
self.module.disable_hooks()
second_loss.backward()
Expand Down Expand Up @@ -104,15 +106,27 @@ def __init__(
self.loss_reduction = loss_reduction
self.criterion.reduction = "none"

def __call__(self, input, target) -> DPTensorFastGradientClipping:
def __call__(self, input, target, shape=None) -> DPTensorFastGradientClipping:
"""
Redefining the forward function to compute per-sample loss and wrap it in DPTensorFastGradientClipping
"""

loss_per_sample = self.criterion(
input,
target,
)
loss_per_sample = self.criterion(input, target)

if shape is not None and loss_per_sample.shape[0] == shape[0] * shape[1]:
# Note that the privacy unit for generative NLP tasks is per sequence.
# The shape variable is the shape of the logits before flattening i.e., [batch_size, sequence_lenght, vocab_size].
# This variable is necessary for ghost clipping to work with generative NLP tasks.
loss_per_sample = loss_per_sample.view(shape[0], shape[1]) # BxT
if self.loss_reduction == "mean":
loss_per_sample = loss_per_sample.mean(dim=1) # B
elif self.loss_reduction == "sum":
loss_per_sample = loss_per_sample.sum(dim=1) # B
else:
raise ValueError(
f"loss_reduction = {self.loss_reduction}. Only 'sum' and 'mean' losses are supported"
)

return DPTensorFastGradientClipping(
self.module, self.optimizer, loss_per_sample, self.loss_reduction
)
Loading