Skip to content

Disk #706

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 10 commits into from
Closed

Disk #706

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
d2e6300
DiSK Optimizer
564612540 Dec 23, 2024
2c538ee
Test optimizers
564612540 Dec 23, 2024
545eca2
delete test file
564612540 Dec 23, 2024
35d793f
fix bugs for optimizer with states
564612540 Dec 24, 2024
4981c00
simplify if-else in __init__.py
564612540 Jan 10, 2025
38b8f37
Merge branch 'main' into disk
564612540 Jan 10, 2025
a092cc0
update code format
564612540 Jan 10, 2025
5251d29
Merge branch 'disk' of https://github.com/564612540/opacus into disk
564612540 Jan 10, 2025
968f933
fix import issue
564612540 Jan 15, 2025
164aa53
Merge branch 'main' into disk
iden-kalemaj Jan 15, 2025
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
Prev Previous commit
Next Next commit
fix bugs for optimizer with states
  • Loading branch information
@564612540
564612540 committed Dec 24, 2024
commit 35d793fa11d2a7f2d64e8a525a5167a5b12ab7ca
11 changes: 11 additions & 0 deletions research/disk_optimizer/optimizers/KFadaclipoptimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,16 +69,27 @@ def step(self, closure=required) -> Optional[float]:
loss = self._compute_one_closure(closure)

if self.pre_step():
tmp_states = []
first_step = False
for p in self.params:
grad = p.grad
state = self.state[p]
if "kf_d_t" not in state:
state = dict()
first_step = True
state["kf_d_t"] = torch.zeros_like(p.data).to(p.data)
state["kf_m_t"] = grad.clone().to(p.data)
state["kf_m_t"].lerp_(grad, weight=self.kappa)
p.grad = state["kf_m_t"].clone().to(p.data)
state["kf_d_t"] = -p.data.clone().to(p.data)
if first_step:
tmp_states.append(state)
self.original_optimizer.step()
for p in self.params:
if first_step:
tmp_state = tmp_states.pop(0)
self.state[p]['kf_d_t'] = tmp_state['kf_d_t']
self.state[p]['kf_m_t'] = tmp_state['kf_m_t']
del tmp_state
self.state[p]["kf_d_t"].add_(p.data, alpha=1.0)
return loss
11 changes: 11 additions & 0 deletions research/disk_optimizer/optimizers/KFddpoptimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,17 +73,28 @@ def step(self, closure=required) -> Optional[float]:
loss = self._compute_one_closure(closure)

if self.pre_step():
tmp_states = []
first_step = False
for p in self.params:
grad = p.grad
state = self.state[p]
if "kf_d_t" not in state:
state = dict()
first_step = True
state["kf_d_t"] = torch.zeros_like(p.data).to(p.data)
state["kf_m_t"] = grad.clone().to(p.data)
state["kf_m_t"].lerp_(grad, weight=self.kappa)
p.grad = state["kf_m_t"].clone().to(p.data)
state["kf_d_t"] = -p.data.clone().to(p.data)
if first_step:
tmp_states.append(state)
self.reduce_gradients()
self.original_optimizer.step()
for p in self.params:
if first_step:
tmp_state = tmp_states.pop(0)
self.state[p]['kf_d_t'] = tmp_state['kf_d_t']
self.state[p]['kf_m_t'] = tmp_state['kf_m_t']
del tmp_state
self.state[p]["kf_d_t"].add_(p.data, alpha=1.0)
return loss
11 changes: 11 additions & 0 deletions research/disk_optimizer/optimizers/KFddpoptimizer_fast_gradient_clipping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,17 +70,28 @@ def step(self, closure=required) -> Optional[float]:
loss = self._compute_one_closure(closure)

if self.pre_step():
tmp_states = []
first_step = False
for p in self.params:
grad = p.grad
state = self.state[p]
if "kf_d_t" not in state:
state = dict()
first_step = True
state["kf_d_t"] = torch.zeros_like(p.data).to(p.data)
state["kf_m_t"] = grad.clone().to(p.data)
state["kf_m_t"].lerp_(grad, weight=self.kappa)
p.grad = state["kf_m_t"].clone().to(p.data)
state["kf_d_t"] = -p.data.clone().to(p.data)
if first_step:
tmp_states.append(state)
self.reduce_gradients()
self.original_optimizer.step()
for p in self.params:
if first_step:
tmp_state = tmp_states.pop(0)
self.state[p]['kf_d_t'] = tmp_state['kf_d_t']
self.state[p]['kf_m_t'] = tmp_state['kf_m_t']
del tmp_state
self.state[p]["kf_d_t"].add_(p.data, alpha=1.0)
return loss
11 changes: 11 additions & 0 deletions research/disk_optimizer/optimizers/KFoptimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,16 +121,27 @@ def step(self, closure=required) -> Optional[float]:
loss = self._compute_one_closure(closure)

if self.pre_step():
tmp_states = []
first_step = False
for p in self.params:
grad = p.grad
state = self.state[p]
if "kf_d_t" not in state:
state = dict()
first_step = True
state["kf_d_t"] = torch.zeros_like(p.data).to(p.data)
state["kf_m_t"] = grad.clone().to(p.data)
state["kf_m_t"].lerp_(grad, weight=self.kappa)
p.grad = state["kf_m_t"].clone().to(p.data)
state["kf_d_t"] = -p.data.clone().to(p.data)
if first_step:
tmp_states.append(state)
self.original_optimizer.step()
for p in self.params:
if first_step:
tmp_state = tmp_states.pop(0)
self.state[p]['kf_d_t'] = tmp_state['kf_d_t']
self.state[p]['kf_m_t'] = tmp_state['kf_m_t']
del tmp_state
self.state[p]["kf_d_t"].add_(p.data, alpha=1.0)
return loss
Loading