Skip to content
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

Jump to bottom

Fix missing move to model device for EkfacInfluence implementation #570

Merged
merged 4 commits into from
May 3, 2024
Merged

Fix missing move to model device for EkfacInfluence implementation #570

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
cafa32a
Fix missing move to model device for EkfacInfluence implementation
schroedk May 2, 2024
a151422
Update CHANGELOG.md
schroedk May 2, 2024
36ea3ba
Add device move in influence_from_factors method in base class TorchI…
schroedk May 3, 2024
dfedbe6
Merge remote-tracking branch 'origin/develop' into fix/569-missing-de…
schroedk May 3, 2024
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
7 changes: 7 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,12 @@
# Changelog

## Unreleased

### Fixed

- Fixed missing move of tensors to model device in `EkfacInfluence`
implementation [PR #570](https://github.com/aai-institute/pyDVL/pull/570)

## 0.9.1 - Bug fixes, logging improvement

### Fixed
Expand Down
17 changes: 10 additions & 7 deletions src/pydvl/influence/torch/influence_function_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -303,13 +303,13 @@ def influences_from_factors(
"""
if mode == InfluenceMode.Up:
return (
z_test_factors
z_test_factors.to(self.model_device)
@ self._loss_grad(x.to(self.model_device), y.to(self.model_device)).T
)
elif mode == InfluenceMode.Perturbation:
return torch.einsum(
"ia,j...a->ij...",
z_test_factors,
z_test_factors.to(self.model_device),
self._flat_loss_mixed_grad(
x.to(self.model_device), y.to(self.model_device)
),
Expand Down Expand Up @@ -1195,7 +1195,7 @@ def _get_kfac_blocks(
data, disable=not self.progress, desc="K-FAC blocks - batch progress"
):
data_len += x.shape[0]
pred_y = self.model(x)
pred_y = self.model(x.to(self.model_device))
loss = empirical_cross_entropy_loss_fn(pred_y)
loss.backward()

Expand Down Expand Up @@ -1319,7 +1319,7 @@ def _update_diag(
data, disable=not self.progress, desc="Update Diagonal - batch progress"
):
data_len += x.shape[0]
pred_y = self.model(x)
pred_y = self.model(x.to(self.model_device))
loss = empirical_cross_entropy_loss_fn(pred_y)
loss.backward()

Expand Down Expand Up @@ -1526,7 +1526,10 @@ def influences_from_factors_by_layer(
influences = {}
for layer_id, layer_z_test in z_test_factors.items():
end_idx = start_idx + layer_z_test.shape[1]
influences[layer_id] = layer_z_test @ total_grad[:, start_idx:end_idx].T
influences[layer_id] = (
layer_z_test.to(self.model_device)
@ total_grad[:, start_idx:end_idx].T
)
start_idx = end_idx
return influences
elif mode == InfluenceMode.Perturbation:
Expand All @@ -1539,7 +1542,7 @@ def influences_from_factors_by_layer(
end_idx = start_idx + layer_z_test.shape[1]
influences[layer_id] = torch.einsum(
"ia,j...a->ij...",
layer_z_test,
layer_z_test.to(self.model_device),
total_mixed_grad[:, start_idx:end_idx],
)
start_idx = end_idx
Expand Down Expand Up @@ -1626,7 +1629,7 @@ def explore_hessian_regularization(
being dictionaries containing the influences for each layer of the model,
with the layer name as key.
"""
grad = self._loss_grad(x, y)
grad = self._loss_grad(x.to(self.model_device), y.to(self.model_device))
influences_by_reg_value = {}
for reg_value in regularization_values:
reg_factors = self._solve_hvp_by_layer(
Expand Down
Loading