Merge pull request #113 from kozistr/feature/lion-optimizer

[Feature] Lion optimizer

Author: kozistr

README.rst

@@ -122,6 +122,8 @@ Supported Optimizers
 +--------------+-------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------------------+-----------------------------------------------------------------------------------------------+
 | NovoGrad     | *Stochastic Gradient Methods with Layer-wise Adaptive Moments for Training of Deep Networks*    | `github <https://github.com/lonePatient/NovoGrad-pytorch>`__                      | `https://arxiv.org/abs/1905.11286 <https://arxiv.org/abs/1905.11286>`__                       |
 +--------------+-------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------------------+-----------------------------------------------------------------------------------------------+
+| Lion         | *Symbolic Discovery of Optimization Algorithms*                                                 | `github <https://github.com/google/automl/tree/master/lion>`__                    | `https://arxiv.org/abs/2302.06675 <https://arxiv.org/abs/2302.06675>`__                       |
++--------------+-------------------------------------------------------------------------------------------------+-----------------------------------------------------------------------------------+-----------------------------------------------------------------------------------------------+
 
 Useful Resources
 ----------------
@@ -323,6 +325,8 @@ Citations
 
 `NovoGrad <https://ui.adsabs.harvard.edu/abs/2019arXiv190511286G/exportcitation>`__
 
+`Lion <https://github.com/google/automl/tree/master/lion#citation>`__
+
 Citation
 --------
 

docs/optimizer_api.rst

@@ -264,3 +264,11 @@ NovoGrad
 
 .. autoclass:: pytorch_optimizer.NovoGrad
     :members:
+
+.. _Lion:
+
+Lion
+----
+
+.. autoclass:: pytorch_optimizer.Lion
+    :members:

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "pytorch_optimizer"
-version = "2.4.2"
+version = "2.5.0"
 description = "optimizer & lr scheduler implementations in PyTorch with clean-code, strict types. Also, including useful optimization ideas."
 license = "Apache-2.0"
 authors = ["kozistr <[email protected]>"]

pytorch_optimizer/__init__.py

@@ -32,6 +32,7 @@
 from pytorch_optimizer.optimizer.gsam import GSAM
 from pytorch_optimizer.optimizer.lamb import Lamb
 from pytorch_optimizer.optimizer.lars import LARS
+from pytorch_optimizer.optimizer.lion import Lion
 from pytorch_optimizer.optimizer.lookahead import Lookahead
 from pytorch_optimizer.optimizer.madgrad import MADGRAD
 from pytorch_optimizer.optimizer.nero import Nero
@@ -98,6 +99,7 @@
     AdaFactor,
     Apollo,
     NovoGrad,
+    Lion,
 ]
 OPTIMIZERS: Dict[str, OPTIMIZER] = {str(optimizer.__name__).lower(): optimizer for optimizer in OPTIMIZER_LIST}
 

pytorch_optimizer/optimizer/lion.py

@@ -0,0 +1,94 @@
+import torch
+from torch.optim.optimizer import Optimizer
+
+from pytorch_optimizer.base.exception import NoSparseGradientError
+from pytorch_optimizer.base.optimizer import BaseOptimizer
+from pytorch_optimizer.base.types import BETAS, CLOSURE, DEFAULTS, LOSS, PARAMETERS
+
+
+class Lion(Optimizer, BaseOptimizer):
+    r"""Symbolic Discovery of Optimization Algorithms.
+
+    :param params: PARAMETERS. iterable of parameters to optimize or dicts defining parameter groups.
+    :param lr: float. learning rate.
+    :param betas: BETAS. coefficients used for computing running averages of gradient and the squared hessian trace.
+    :param weight_decay: float. weight decay (L2 penalty).
+    :param weight_decouple: bool. the optimizer uses decoupled weight decay as in AdamW.
+    """
+
+    def __init__(
+        self,
+        params: PARAMETERS,
+        lr: float = 1e-4,
+        betas: BETAS = (0.9, 0.99),
+        weight_decay: float = 0.0,
+        weight_decouple: bool = True,
+    ):
+        self.lr = lr
+        self.betas = betas
+        self.weight_decay = weight_decay
+        self.weight_decouple = weight_decouple
+
+        self.validate_parameters()
+
+        defaults: DEFAULTS = {
+            'lr': lr,
+            'betas': betas,
+            'weight_decay': weight_decay,
+        }
+        super().__init__(params, defaults)
+
+    def validate_parameters(self):
+        self.validate_learning_rate(self.lr)
+        self.validate_betas(self.betas)
+        self.validate_weight_decay(self.weight_decay)
+
+    @property
+    def __str__(self) -> str:
+        return 'Lion'
+
+    @torch.no_grad()
+    def reset(self):
+        for group in self.param_groups:
+            for p in group['params']:
+                state = self.state[p]
+
+                state['exp_avg'] = torch.zeros_like(p)
+
+    @torch.no_grad()
+    def step(self, closure: CLOSURE = None) -> LOSS:
+        loss: LOSS = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            beta1, beta2 = group['betas']
+            weight_decay = group['weight_decay']
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+
+                grad = p.grad
+                if grad.is_sparse:
+                    raise NoSparseGradientError(self.__str__)
+
+                state = self.state[p]
+
+                if len(state) == 0:
+                    state['exp_avg'] = torch.zeros_like(p)
+
+                update = exp_avg = state['exp_avg']
+
+                if weight_decay > 0.0:
+                    if self.weight_decouple:
+                        p.mul_(1.0 - group['lr'] * weight_decay)
+                    else:
+                        grad.add_(p, alpha=weight_decay)
+
+                update.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+                exp_avg.mul_(beta2).add_(grad, alpha=1.0 - beta2)
+
+                p.add_(update.sign(), alpha=-group['lr'])
+
+        return loss

tests/constants.py

@@ -21,6 +21,7 @@
     DiffGrad,
     DiffRGrad,
     Lamb,
+    Lion,
     Nero,
     NovoGrad,
     RAdam,
@@ -70,6 +71,7 @@
     'adams',
     'adafactor',
     'novograd',
+    'lion',
 ]
 
 VALID_LR_SCHEDULER_NAMES: List[str] = [
@@ -161,6 +163,8 @@
     (Apollo, {'lr': 5e-1, 'weight_decay': 1e-3, 'rebound': 'belief'}, 10),
     (Apollo, {'lr': 5e-1, 'weight_decay': 1e-3, 'weight_decay_type': 'stable', 'warmup_steps': 0}, 50),
     (NovoGrad, {'lr': 5e-1, 'weight_decay': 1e-3, 'grad_averaging': True}, 50),
+    (Lion, {'lr': 5e-1, 'weight_decay': 1e-3}, 10),
+    (Lion, {'lr': 5e-1, 'weight_decay': 1e-3, 'weight_decouple': False}, 10),
 ]
 ADAMD_SUPPORTED_OPTIMIZERS: List[Tuple[Any, Dict[str, Union[float, bool, int]], int]] = [
     (build_lookahead, {'lr': 5e-1, 'weight_decay': 1e-3, 'adamd_debias_term': True}, 10),

tests/test_load_optimizers.py

@@ -16,4 +16,4 @@ def test_load_optimizers_invalid(invalid_optimizer_names):
 
 
 def test_get_supported_optimizers():
-    assert len(get_supported_optimizers()) == 27
+    assert len(get_supported_optimizers()) == 28

tests/test_optimizer_parameters.py

@@ -22,7 +22,7 @@ def test_learning_rate(optimizer_name):
 
 @pytest.mark.parametrize('optimizer_name', VALID_OPTIMIZER_NAMES)
 def test_epsilon(optimizer_name):
-    if optimizer_name in ('nero', 'shampoo', 'scalableshampoo', 'dadaptsgd', 'adafactor'):
+    if optimizer_name in ('nero', 'shampoo', 'scalableshampoo', 'dadaptsgd', 'adafactor', 'lion'):
         pytest.skip(f'skip {optimizer_name} optimizer')
 
     optimizer = load_optimizer(optimizer_name)