add the proposed cautious optimizer from https://arxiv.org/abs/2411.16085, but allow for attenuating unaligned updates with some factor instead of zeroing completely

Author: lucidrains

README.md

@@ -81,3 +81,12 @@ for _ in range(100):
     url     = {https://api.semanticscholar.org/CorpusID:273822148}
 }
 ```
+
+```bibtex
+@inproceedings{Liang2024CautiousOI,
+    title   = {Cautious Optimizers: Improving Training with One Line of Code},
+    author  = {Kaizhao Liang and Lizhang Chen and Bo Liu and Qiang Liu},
+    year    = {2024},
+    url     = {https://api.semanticscholar.org/CorpusID:274234738}
+}
+```

adam_atan2_pytorch/adam_atan2.py

@@ -21,6 +21,7 @@ def __init__(
         weight_decay = 0.,
         regen_reg_rate = 0.,
         decoupled_wd = False,
+        cautious_factor = 1., # set to 0. for zeroing out any updates not in same direction as gradient as in https://arxiv.org/abs/2411.16085
         a = 1.27,
         b = 1.
     ):
@@ -29,6 +30,7 @@ def __init__(
         assert weight_decay >= 0.
         assert regen_reg_rate >= 0.
         assert not (weight_decay > 0. and regen_reg_rate > 0.)
+        assert 0. <= cautious_factor <= 1.
 
         self._init_lr = lr
         self.decoupled_wd = decoupled_wd
@@ -40,6 +42,7 @@ def __init__(
             b = b,
             weight_decay = weight_decay,
             regen_reg_rate = regen_reg_rate,
+            cautious_factor = cautious_factor
         )
 
         super().__init__(params, defaults)
@@ -58,7 +61,7 @@ def step(
         for group in self.param_groups:
             for p in filter(lambda p: exists(p.grad), group['params']):
 
-                grad, lr, wd, regen_rate, beta1, beta2, a, b, state, init_lr = p.grad, group['lr'], group['weight_decay'], group['regen_reg_rate'], *group['betas'], group['a'], group['b'], self.state[p], self._init_lr
+                grad, lr, wd, regen_rate, cautious_factor, beta1, beta2, a, b, state, init_lr = p.grad, group['lr'], group['weight_decay'], group['regen_reg_rate'], group['cautious_factor'], *group['betas'], group['a'], group['b'], self.state[p], self._init_lr
 
                 # maybe decoupled weight decay
 
@@ -109,6 +112,13 @@ def step(
                 den = exp_avg_sq.mul(b * b / bias_correct2).sqrt_()
                 update = exp_avg.mul(1. / bias_correct1).atan2_(den)
 
+                # maybe cautious update - algorithm 2 in https://arxiv.org/abs/2411.16085
+
+                if cautious_factor < 1.:
+                    align_mask = (update * grad) > 0
+                    scale = torch.where(align_mask, torch.ones_like(grad), cautious_factor)
+                    update *= (scale / scale.mean().clamp(min = 1e-5))
+
                 # update parameters
 
                 p.add_(update, alpha = -lr * a)

adam_atan2_pytorch/adopt_atan2.py

@@ -28,6 +28,7 @@ def __init__(
         weight_decay = 0.,
         regen_reg_rate = 0.,
         decoupled_wd = True,
+        cautious_factor = 1., # set to 0. for zeroing out any updates not in same direction as gradient as in https://arxiv.org/abs/2411.16085
         a = 1.27,
         b = 1.
     ):
@@ -45,7 +46,8 @@ def __init__(
             a = a,
             b = b,
             weight_decay = weight_decay,
-            regen_reg_rate = regen_reg_rate
+            regen_reg_rate = regen_reg_rate,
+            cautious_factor = cautious_factor
         )
 
         super().__init__(params, defaults)
@@ -64,7 +66,7 @@ def step(
         for group in self.param_groups:
             for p in filter(lambda p: exists(p.grad), group['params']):
 
-                grad, lr, wd, regen_rate, beta1, beta2, a, b, state, init_lr = p.grad, group['lr'], group['weight_decay'], group['regen_reg_rate'], *group['betas'], group['a'], group['b'], self.state[p], self._init_lr
+                grad, lr, wd, regen_rate, cautious_factor, beta1, beta2, a, b, state, init_lr = p.grad, group['lr'], group['weight_decay'], group['regen_reg_rate'], group['cautious_factor'], *group['betas'], group['a'], group['b'], self.state[p], self._init_lr
 
                 # maybe decoupled weight decay
 
@@ -110,9 +112,18 @@ def step(
 
                 m.lerp_(update, 1. - beta1)
 
+                # maybe cautious update - algorithm 2 in https://arxiv.org/abs/2411.16085
+
+                scale = 1.
+
+                if cautious_factor < 1.:
+                    align_mask = (update * grad) > 0
+                    scale = torch.where(align_mask, torch.ones_like(grad), cautious_factor)
+                    scale /= scale.mean().clamp(min = 1e-5)
+
                 # then update parameters
 
-                p.add_(m, alpha = -lr * a)
+                p.add_(m * scale, alpha = -lr * a)
 
                 # update exp grad sq (v)
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "adam-atan2-pytorch"
-version = "0.1.16"
+version = "0.1.18"
 description = "Adam-atan2 for Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }