Merge pull request #2357 from huggingface/more_opt_stuff

Add caution to Adan. Add decouple decay option to LAMB.

Author: rwightman

timm/optim/_optim_factory.py

@@ -485,6 +485,20 @@ def _register_lamb_lars(registry: OptimizerRegistry) -> None:
             has_betas=True,
             defaults={'trust_clip': True}
         ),
+        OptimInfo(
+            name='lambw',
+            opt_class=Lamb,
+            description='LAMB with decoupled weight decay',
+            has_betas=True,
+            defaults={'decoupled_decay': True}
+        ),
+        OptimInfo(
+            name='lambcw',
+            opt_class=Lamb,
+            description='LAMB with trust ratio clipping for stability and decoupled decay',
+            has_betas=True,
+            defaults={'trust_clip': True, 'decoupled_decay': True}
+        ),
         OptimInfo(
             name='lars',
             opt_class=Lars,
@@ -544,6 +558,22 @@ def _register_cautious_optimizers(registry: OptimizerRegistry) -> None:
             description='Cautious Adopt',
             defaults={'caution': True}
         ),
+        OptimInfo(
+            name='cadan',
+            opt_class=Adan,
+            description='Cautious Adaptive Nesterov Momentum Algorithm',
+            defaults={'caution': True, 'no_prox': False},
+            has_betas=True,
+            num_betas=3
+        ),
+        OptimInfo(
+            name='cadanw',
+            opt_class=Adan,
+            description='Cautious Adaptive Nesterov Momentum with decoupled weight decay',
+            defaults={'caution': True, 'no_prox': True},
+            has_betas=True,
+            num_betas=3
+        ),
         OptimInfo(
             name='cadoptw',
             opt_class=Adopt,
@@ -557,6 +587,13 @@ def _register_cautious_optimizers(registry: OptimizerRegistry) -> None:
             has_betas=True,
             defaults={'caution': True}
         ),
+        OptimInfo(
+            name='clambw',
+            opt_class=Lamb,
+            description='Cautious LAMB with decoupled weight decay',
+            has_betas=True,
+            defaults={'caution': True, 'decoupled_decay': True}
+        ),
         OptimInfo(
             name='claprop',
             opt_class=LaProp,

timm/optim/adan.py

@@ -20,7 +20,7 @@
 # limitations under the License.
 
 import math
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 import torch
 from torch import Tensor
@@ -56,6 +56,7 @@ class Adan(Optimizer):
         eps: Term added to the denominator to improve numerical stability.
         weight_decay: Decoupled weight decay (L2 penalty)
         no_prox: How to perform the weight decay
+        caution: Enable caution from 'Cautious Optimizers'
         foreach: If True would use torch._foreach implementation. Faster but uses slightly more memory.
     """
 
@@ -66,7 +67,8 @@ def __init__(self,
             eps: float = 1e-8,
             weight_decay: float = 0.0,
             no_prox: bool = False,
-            foreach: bool = True,
+            caution: bool = False,
+            foreach: Optional[bool] = None,
     ):
         if not 0.0 <= lr:
             raise ValueError('Invalid learning rate: {}'.format(lr))
@@ -85,6 +87,7 @@ def __init__(self,
             eps=eps,
             weight_decay=weight_decay,
             no_prox=no_prox,
+            caution=caution,
             foreach=foreach,
         )
         super().__init__(params, defaults)
@@ -93,6 +96,7 @@ def __setstate__(self, state):
         super(Adan, self).__setstate__(state)
         for group in self.param_groups:
             group.setdefault('no_prox', False)
+            group.setdefault('caution', False)
 
     @torch.no_grad()
     def restart_opt(self):
@@ -118,6 +122,11 @@ def step(self, closure=None):
             with torch.enable_grad():
                 loss = closure()
 
+        try:
+            has_scalar_maximum = 'Scalar' in torch.ops.aten._foreach_maximum_.overloads()
+        except:
+            has_scalar_maximum = False
+
         for group in self.param_groups:
             params_with_grad = []
             grads = []
@@ -161,9 +170,19 @@ def step(self, closure=None):
             if not params_with_grad:
                 continue
 
-            kwargs = dict(
-                params=params_with_grad,
-                grads=grads,
+            if group['foreach'] is None:
+                use_foreach = not group['caution'] or has_scalar_maximum
+            else:
+                use_foreach = group['foreach']
+
+            if use_foreach:
+                func = _multi_tensor_adan
+            else:
+                func = _single_tensor_adan
+
+            func(
+                params_with_grad,
+                grads,
                 exp_avgs=exp_avgs,
                 exp_avg_sqs=exp_avg_sqs,
                 exp_avg_diffs=exp_avg_diffs,
@@ -178,13 +197,9 @@ def step(self, closure=None):
                 weight_decay=group['weight_decay'],
                 eps=group['eps'],
                 no_prox=group['no_prox'],
+                caution=group['caution'],
             )
 
-            if group['foreach']:
-                _multi_tensor_adan(**kwargs)
-            else:
-                _single_tensor_adan(**kwargs)
-
         return loss
 
 
@@ -206,6 +221,7 @@ def _single_tensor_adan(
         weight_decay: float,
         eps: float,
         no_prox: bool,
+        caution: bool,
 ):
     for i, param in enumerate(params):
         grad = grads[i]
@@ -227,6 +243,12 @@ def _single_tensor_adan(
         step_size_diff = lr * beta2 / bias_correction2
         step_size = lr / bias_correction1
 
+        if caution:
+            # Apply caution as per 'Cautious Optimizers' - https://arxiv.org/abs/2411.16085
+            mask = (exp_avg * grad > 0).to(grad.dtype)
+            mask.div_(mask.mean().clamp_(min=1e-3))
+            exp_avg = exp_avg * mask
+
         if no_prox:
             param.mul_(1 - lr * weight_decay)
             param.addcdiv_(exp_avg, denom, value=-step_size)
@@ -257,6 +279,7 @@ def _multi_tensor_adan(
         weight_decay: float,
         eps: float,
         no_prox: bool,
+        caution: bool,
 ):
     if len(params) == 0:
         return
@@ -282,6 +305,15 @@ def _multi_tensor_adan(
     step_size_diff = lr * beta2 / bias_correction2
     step_size = lr / bias_correction1
 
+    if caution:
+        # Apply caution as per 'Cautious Optimizers' - https://arxiv.org/abs/2411.16085
+        masks = torch._foreach_mul(exp_avgs, grads)
+        masks = [(m > 0).to(g.dtype) for m, g in zip(masks, grads)]
+        mask_scale = [m.mean() for m in masks]
+        torch._foreach_maximum_(mask_scale, 1e-3)
+        torch._foreach_div_(masks, mask_scale)
+        exp_avgs = torch._foreach_mul(exp_avgs, masks)
+
     if no_prox:
         torch._foreach_mul_(params, 1 - lr * weight_decay)
         torch._foreach_addcdiv_(params, exp_avgs, denom, value=-step_size)

timm/optim/lamb.py

@@ -94,6 +94,7 @@ def __init__(
             trust_clip: bool = False,
             always_adapt: bool = False,
             caution: bool = False,
+            decoupled_decay: bool = False,
     ):
         defaults = dict(
             lr=lr,
@@ -106,13 +107,15 @@ def __init__(
             trust_clip=trust_clip,
             always_adapt=always_adapt,
             caution=caution,
+            decoupled_decay=decoupled_decay,
         )
         super().__init__(params, defaults)
 
     def __setstate__(self, state):
         super().__setstate__(state)
         for group in self.param_groups:
             group.setdefault('caution', False)
+            group.setdefault('decoupled_decay', False)
 
     def _get_clip_grad_norm(self):
         max_grad_norm = self.defaults['max_grad_norm']
@@ -199,7 +202,10 @@ def step(self, closure=None):
 
                 weight_decay = group['weight_decay']
                 if weight_decay != 0:
-                    update.add_(p, alpha=weight_decay)
+                    if group.get('decoupled_decay', False):
+                        p.add_(p, alpha=-group['lr'] * weight_decay)
+                    else:
+                        update.add_(p, alpha=weight_decay)
 
                 if weight_decay != 0 or group['always_adapt']:
                     # Layer-wise LR adaptation. By default, skip adaptation on parameters that are

timm/optim/mars.py

@@ -54,10 +54,13 @@ def _mars_single_tensor_step(
             if c_t_norm > 1.:
                 c_t = c_t / c_t_norm
         exp_avg.mul_(beta1).add_(c_t, alpha=one_minus_beta1)
+
         if caution:
+            # Apply caution as per 'Cautious Optimizers' - https://arxiv.org/abs/2411.16085
             mask = (exp_avg * grad > 0).to(grad.dtype)
             mask.div_(mask.mean().clamp_(min=1e-3))
             exp_avg = exp_avg * mask
+
         if mars_type == "adamw":
             exp_avg_sq.mul_(beta2).addcmul_(c_t, c_t, value=1. - beta2)
             bias_correction1 = 1.0 - beta1 ** step