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

[Fix] schedulefree #351

Closed
wants to merge 4 commits into from
Closed

[Fix] schedulefree #351

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3d54dfb
fix: schedulefree
Feb 17, 2025
63497b8
fix: restore weight_decouple argument in ScheduleFreeSGD
Feb 19, 2025
5ba1182
fix: palm betas scheduling in ScheduleFreeAdamW/ScheduleFreeRAdam
Feb 19, 2025
10dd63e
fix: skip updates when lr is 0 in ScheduleFreeRAdam
Feb 19, 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
76 changes: 35 additions & 41 deletions pytorch_optimizer/optimizer/schedulefree.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,21 +132,15 @@ def step(self, closure: CLOSURE = None) -> LOSS:
if len(state) == 0:
state['z'] = p.clone()

self.apply_weight_decay(
p=p,
grad=grad,
lr=lr,
weight_decay=group['weight_decay'],
weight_decouple=group['weight_decouple'],
fixed_decay=group['fixed_decay'],
)

z = state['z']

grad.mul_(lr)
grad.add_(p, alpha=group['weight_decay'] * (1.0 if group['fixed_decay'] else lr))

p.lerp_(z, weight=checkpoint)
p.add_(grad, alpha=lr * (momentum * (1.0 - checkpoint) - 1))
p.add_(grad, alpha=momentum * (1.0 - checkpoint) - 1)

z.sub_(grad, alpha=lr)
z.sub_(grad)

return loss

Expand Down Expand Up @@ -259,9 +253,9 @@ def step(self, closure: CLOSURE = None) -> LOSS:

beta1, beta2 = group['betas']

bias_correction2_sq: float = math.sqrt(1.0 - beta2 ** group['step'])
bias_correction2: float = 1.0 - beta2 ** group['step']

lr: float = group['lr'] * schedule * bias_correction2_sq
lr: float = group['lr'] * schedule
lr_max = group['lr_max'] = max(lr, group['lr_max'])

weight = (group['step'] ** group['r']) * (lr_max ** group['weight_lr_power'])
Expand All @@ -271,7 +265,9 @@ def step(self, closure: CLOSURE = None) -> LOSS:

if group['use_palm']:
beta2: float = 1.0 - group['step'] ** -0.8
debias: float = (1.0 - beta2) / (1.0 - beta2 ** group['step'])
debias: float = 1.0 - (1.0 - beta2) / (1.0 - beta2 ** group['step'])
# unnecessary bias correction when PaLM beta2 scheduling
bias_correction2 = 1.0
else:
debias: float = beta2

Expand All @@ -289,31 +285,27 @@ def step(self, closure: CLOSURE = None) -> LOSS:
state['z'] = p.clone()
state['exp_avg_sq'] = torch.zeros_like(p)

self.apply_weight_decay(
p=p,
grad=grad,
lr=lr,
weight_decay=group['weight_decay'],
weight_decouple=group['weight_decouple'],
fixed_decay=group['fixed_decay'],
)
if not group['weight_decouple']:
grad.add_(p, alpha=group['weight_decay'])

z, exp_avg_sq = state['z'], state['exp_avg_sq']
exp_avg_sq.mul_(debias).addcmul_(grad, grad, value=1.0 - debias)

de_nom = self.apply_ams_bound(
ams_bound=group['ams_bound'],
exp_avg_sq=exp_avg_sq,
exp_avg_sq=exp_avg_sq.div(bias_correction2),
max_exp_avg_sq=state.get('max_exp_avg_sq', None),
eps=group['eps'],
)

grad.div_(de_nom)
grad.mul_(lr)
if group['weight_decouple']:
grad.add_(p, alpha=group['weight_decay'] * (1.0 if group['fixed_decay'] else lr))

p.lerp_(z, weight=checkpoint)
p.add_(grad, alpha=lr * (beta1 * (1.0 - checkpoint) - 1))
p.add_(grad, alpha=beta1 * (1.0 - checkpoint) - 1)

z.sub_(grad, alpha=lr)
z.sub_(grad)

return loss

Expand Down Expand Up @@ -428,19 +420,22 @@ def step(self, closure: CLOSURE = None) -> LOSS:
n_sma_threshold=4,
degenerated_to_sgd=group['degenerated_to_sgd'],
)
if n_sma > 4:
# cancel bias correction2
lr = lr / bias_correction2_sq

lr_max = group['lr_max'] = max(lr, group['lr_max'])
lr_max = group['lr_max'] = max(lr, group['lr_max'], 0.0)

weight = (group['step'] ** group['r']) * (lr_max ** group['weight_lr_power'])
weight_sum = group['weight_sum'] = group['weight_sum'] + weight

checkpoint: float = weight / weight_sum if weight_sum != 0.0 else 0.0

adaptive_y_lr: float = lr * (beta1 * (1.0 - checkpoint) - 1.0)

if group['use_palm']:
beta2: float = 1.0 - group['step'] ** -0.8
debias: float = (1.0 - beta2) / (1.0 - beta2 ** group['step'])
debias: float = 1.0 - (1.0 - beta2) / (1.0 - beta2 ** group['step'])
# unnecessary bias correction when PaLM beta2 scheduling
bias_correction2_sq = 1.0
else:
debias: float = beta2

Expand All @@ -458,25 +453,24 @@ def step(self, closure: CLOSURE = None) -> LOSS:
state['z'] = p.clone()
state['exp_avg_sq'] = torch.zeros_like(p)

if not group['weight_decouple']:
grad.add_(p, alpha=group['weight_decay'])

z, exp_avg_sq = state['z'], state['exp_avg_sq']
exp_avg_sq.mul_(debias).addcmul_(grad, grad, value=1.0 - debias)

if n_sma > 4.0:
de_nom = exp_avg_sq.sqrt().div_(bias_correction2_sq).add_(group['eps'])
grad.div_(de_nom)

self.apply_weight_decay(
p=p,
grad=grad,
lr=lr,
weight_decay=group['weight_decay'],
weight_decouple=group['weight_decouple'],
fixed_decay=group['fixed_decay'],
)
if lr > 0.0:
grad.mul_(lr)
if group['weight_decouple']:
grad.add_(p, alpha=group['weight_decay'] * (1.0 if group['fixed_decay'] else lr))

p.lerp_(z, weight=checkpoint)
p.add_(grad, alpha=adaptive_y_lr)
p.lerp_(z, weight=checkpoint)
p.add_(grad, alpha=beta1 * (1.0 - checkpoint) - 1.0)

z.sub_(grad, alpha=lr)
z.sub_(grad)

return loss
4 changes: 3 additions & 1 deletion tests/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -502,7 +502,9 @@
(Adalite, {'lr': 1e0, 'weight_decay': 1e-3}, 5),
(ScheduleFreeSGD, {'lr': 1e0, 'weight_decay': 1e-3}, 5),
(ScheduleFreeAdamW, {'lr': 1e0, 'weight_decay': 1e-3}, 5),
(ScheduleFreeAdamW, {'lr': 1e-2, 'weight_decay': 1e-3, 'use_palm': True}, 5),
(ScheduleFreeAdamW, {'lr': 1e0, 'weight_decay': 1e-3, 'use_palm': True}, 5),
(ScheduleFreeRAdam, {'lr': 5e0, 'weight_decay': 1e-3}, 10),
(ScheduleFreeRAdam, {'lr': 5e0, 'weight_decay': 1e-3, 'use_palm': True}, 10),
(ScheduleFreeRAdam, {'lr': 1e0, 'weight_decay': 1e-3, 'degenerated_to_sgd': True}, 5),
(ScheduleFreeRAdam, {'lr': 1e0, 'weight_decay': 1e-3, 'use_palm': True, 'degenerated_to_sgd': True}, 5),
(FAdam, {'lr': 1e0, 'weight_decay': 1e-3}, 5),
Expand Down