pytorch-optimizer 3.3.4__tar.gz → 3.4.0__tar.gz

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/PKG-INFO

@@ -1,9 +1,9 @@
 Metadata-Version: 2.3
 Name: pytorch_optimizer
-Version: 3.3.4
+Version: 3.4.0
 Summary: optimizer & lr scheduler & objective function collections in PyTorch
 License: Apache-2.0
-Keywords: pytorch,deep-learning,optimizer,lr scheduler,A2Grad,ASGD,AccSGD,AdaBelief,AdaBound,AdaDelta,AdaFactor,AdaMax,AdamG,AdaMod,AdaNorm,AdaPNM,AdaSmooth,AdEMAMix,ADOPT,AdaHessian,Adai,Adalite,AdaLomo,AdamMini,AdamP,AdamS,Adan,AggMo,Aida,AliG,Amos,Apollo,APOLLO,AvaGrad,bSAM,CAME,DAdaptAdaGrad,DAdaptAdam,DAdaptAdan,DAdaptSGD,DAdaptLion,DeMo,DiffGrad,FAdam,Fromage,FTRL,GaLore,Grams,Gravity,GrokFast,GSAM,Kate,Lamb,LaProp,LARS,Lion,LOMO,Lookahead,MADGRAD,MARS,MSVAG,Muno,Nero,NovoGrad,OrthoGrad,PAdam,PCGrad,PID,PNM,Prodigy,QHAdam,QHM,RAdam,Ranger,Ranger21,RotoGrad,SAM,ScheduleFreeSGD,ScheduleFreeAdamW,ScheduleFreeRAdam,SGDP,Shampoo,ScalableShampoo,SGDW,SignSGD,SM3,SOAP,SopihaH,SPAM,SRMM,StableAdamW,SWATS,TAM,Tiger,TRAC,WSAM,Yogi,BCE,BCEFocal,Focal,FocalCosine,SoftF1,Dice,LDAM,Jaccard,Bi-Tempered,Tversky,FocalTversky,LovaszHinge,bitsandbytes,WSD,QGaLore
+Keywords: pytorch,deep-learning,optimizer,lr scheduler,A2Grad,ASGD,AccSGD,AdaBelief,AdaBound,AdaDelta,AdaFactor,AdaMax,AdamG,AdaMod,AdaNorm,AdaPNM,AdaSmooth,AdEMAMix,ADOPT,AdaHessian,Adai,Adalite,AdaLomo,AdamMini,AdamP,AdamS,Adan,AggMo,Aida,AliG,Amos,Apollo,APOLLO,AvaGrad,bSAM,CAME,DAdaptAdaGrad,DAdaptAdam,DAdaptAdan,DAdaptSGD,DAdaptLion,DeMo,DiffGrad,EXAdam,FAdam,FOCUS,Fromage,FTRL,GaLore,Grams,Gravity,GrokFast,GSAM,Kate,Lamb,LaProp,LARS,Lion,LOMO,Lookahead,MADGRAD,MARS,MSVAG,Muno,Nero,NovoGrad,OrthoGrad,PAdam,PCGrad,PID,PNM,Prodigy,PSGD,QHAdam,QHM,RAdam,Ranger,Ranger21,RotoGrad,SAM,ScheduleFreeSGD,ScheduleFreeAdamW,ScheduleFreeRAdam,SGDP,Shampoo,ScalableShampoo,SGDW,SignSGD,SM3,SOAP,SopihaH,SPAM,SRMM,StableAdamW,SWATS,TAM,Tiger,TRAC,WSAM,Yogi,BCE,BCEFocal,Focal,FocalCosine,SoftF1,Dice,LDAM,Jaccard,Bi-Tempered,Tversky,FocalTversky,LovaszHinge,bitsandbytes,WSD,QGaLore
 Author: kozistr
 Author-email: kozistr@gmail.com
 Maintainer: kozistr
@@ -29,8 +29,6 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development
 Classifier: Topic :: Software Development :: Libraries
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
-Provides-Extra: bitsandbytes
-Requires-Dist: bitsandbytes (>=0.44,<0.45) ; extra == "bitsandbytes"
 Requires-Dist: numpy (<=1.24.4) ; python_version < "3.9"
 Requires-Dist: numpy (>1.24.4) ; python_version >= "3.9"
 Requires-Dist: torch (>=1.10) ; python_version >= "3.8"
@@ -51,7 +49,7 @@ Description-Content-Type: text/markdown
 
 ## The reasons why you use `pytorch-optimizer`.
 
-* Wide range of supported optimizers. Currently, **92 optimizers (+ `bitsandbytes`, `qgalore`, `torchao`)**, **16 lr schedulers**, and **13 loss functions** are supported!
+* Wide range of supported optimizers. Currently, **95 optimizers (+ `bitsandbytes`, `qgalore`, `torchao`)**, **16 lr schedulers**, and **13 loss functions** are supported!
 * Including many variants such as `ADOPT`, `Cautious`, `AdamD`, `StableAdamW`, and `Gradient Centrailiaztion`
 * Easy to use, clean, and tested codes
 * Active maintenance
@@ -241,6 +239,9 @@ get_supported_optimizers(['adam*', 'ranger*'])
 | Adam-ATAN2    | *Scaling Exponents Across Parameterizations and Optimizers*                                       |                                                                                                                | <https://arxiv.org/abs/2407.05872>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2024arXiv240705872E/exportcitation)                                                        |
 | SPAM          | *Spike-Aware Adam with Momentum Reset for Stable LLM Training*                                    | [github](https://github.com/TianjinYellow/SPAM-Optimizer)                                                      | <https://arxiv.org/abs/2501.06842>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2025arXiv250106842H/exportcitation)                                                        |
 | TAM           | *Torque-Aware Momentum*                                                                           |                                                                                                                | <https://arxiv.org/abs/2412.18790>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2024arXiv241218790M/exportcitation)                                                        |
+| FOCUS         | *First Order Concentrated Updating Scheme*                                                        | [github](https://github.com/liuyz0/FOCUS)                                                                      | <https://arxiv.org/abs/2501.12243>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2025arXiv250112243M/exportcitation)                                                        |
+| PSGD          | *Preconditioned Stochastic Gradient Descent*                                                      | [github](https://github.com/lixilinx/psgd_torch)                                                               | <https://arxiv.org/abs/1512.04202>                                                          | [cite](https://github.com/lixilinx/psgd_torch?tab=readme-ov-file#resources)                                                         |
+| EXAdam        | *The Power of Adaptive Cross-Moments*                                                             | [github](https://github.com/AhmedMostafa16/EXAdam)                                                             | <https://arxiv.org/abs/2412.20302>                                                          | [cite](https://github.com/AhmedMostafa16/EXAdam?tab=readme-ov-file#citation)                                                        |
 
 ## Supported LR Scheduler
 

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/README.md

@@ -10,7 +10,7 @@
 
 ## The reasons why you use `pytorch-optimizer`.
 
-* Wide range of supported optimizers. Currently, **92 optimizers (+ `bitsandbytes`, `qgalore`, `torchao`)**, **16 lr schedulers**, and **13 loss functions** are supported!
+* Wide range of supported optimizers. Currently, **95 optimizers (+ `bitsandbytes`, `qgalore`, `torchao`)**, **16 lr schedulers**, and **13 loss functions** are supported!
 * Including many variants such as `ADOPT`, `Cautious`, `AdamD`, `StableAdamW`, and `Gradient Centrailiaztion`
 * Easy to use, clean, and tested codes
 * Active maintenance
@@ -200,6 +200,9 @@ get_supported_optimizers(['adam*', 'ranger*'])
 | Adam-ATAN2    | *Scaling Exponents Across Parameterizations and Optimizers*                                       |                                                                                                                | <https://arxiv.org/abs/2407.05872>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2024arXiv240705872E/exportcitation)                                                        |
 | SPAM          | *Spike-Aware Adam with Momentum Reset for Stable LLM Training*                                    | [github](https://github.com/TianjinYellow/SPAM-Optimizer)                                                      | <https://arxiv.org/abs/2501.06842>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2025arXiv250106842H/exportcitation)                                                        |
 | TAM           | *Torque-Aware Momentum*                                                                           |                                                                                                                | <https://arxiv.org/abs/2412.18790>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2024arXiv241218790M/exportcitation)                                                        |
+| FOCUS         | *First Order Concentrated Updating Scheme*                                                        | [github](https://github.com/liuyz0/FOCUS)                                                                      | <https://arxiv.org/abs/2501.12243>                                                          | [cite](https://ui.adsabs.harvard.edu/abs/2025arXiv250112243M/exportcitation)                                                        |
+| PSGD          | *Preconditioned Stochastic Gradient Descent*                                                      | [github](https://github.com/lixilinx/psgd_torch)                                                               | <https://arxiv.org/abs/1512.04202>                                                          | [cite](https://github.com/lixilinx/psgd_torch?tab=readme-ov-file#resources)                                                         |
+| EXAdam        | *The Power of Adaptive Cross-Moments*                                                             | [github](https://github.com/AhmedMostafa16/EXAdam)                                                             | <https://arxiv.org/abs/2412.20302>                                                          | [cite](https://github.com/AhmedMostafa16/EXAdam?tab=readme-ov-file#citation)                                                        |
 
 ## Supported LR Scheduler
 

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "pytorch_optimizer"
-version = "3.3.4"
+version = "3.4.0"
 description = "optimizer & lr scheduler & objective function collections in PyTorch"
 license = "Apache-2.0"
 authors = ["kozistr <kozistr@gmail.com>"]
@@ -14,13 +14,13 @@ keywords = [
     "AdaDelta", "AdaFactor", "AdaMax", "AdamG", "AdaMod", "AdaNorm", "AdaPNM", "AdaSmooth", "AdEMAMix", "ADOPT",
     "AdaHessian", "Adai", "Adalite", "AdaLomo", "AdamMini", "AdamP", "AdamS", "Adan", "AggMo", "Aida", "AliG", "Amos",
     "Apollo", "APOLLO", "AvaGrad", "bSAM", "CAME", "DAdaptAdaGrad", "DAdaptAdam", "DAdaptAdan", "DAdaptSGD",
-    "DAdaptLion", "DeMo", "DiffGrad", "FAdam", "Fromage", "FTRL", "GaLore", "Grams", "Gravity", "GrokFast", "GSAM",
-    "Kate", "Lamb", "LaProp", "LARS", "Lion", "LOMO", "Lookahead", "MADGRAD", "MARS", "MSVAG", "Muno", "Nero",
-    "NovoGrad", "OrthoGrad", "PAdam", "PCGrad", "PID", "PNM", "Prodigy", "QHAdam", "QHM", "RAdam", "Ranger",
-    "Ranger21", "RotoGrad", "SAM", "ScheduleFreeSGD", "ScheduleFreeAdamW", "ScheduleFreeRAdam", "SGDP", "Shampoo",
-    "ScalableShampoo", "SGDW", "SignSGD", "SM3", "SOAP", "SopihaH", "SPAM", "SRMM", "StableAdamW", "SWATS", "TAM",
-    "Tiger", "TRAC", "WSAM", "Yogi", "BCE", "BCEFocal", "Focal", "FocalCosine", "SoftF1", "Dice", "LDAM", "Jaccard",
-    "Bi-Tempered", "Tversky", "FocalTversky", "LovaszHinge", "bitsandbytes", "WSD", "QGaLore",
+    "DAdaptLion", "DeMo", "DiffGrad", "EXAdam", "FAdam", "FOCUS", "Fromage", "FTRL", "GaLore", "Grams", "Gravity",
+    "GrokFast", "GSAM", "Kate", "Lamb", "LaProp", "LARS", "Lion", "LOMO", "Lookahead", "MADGRAD", "MARS", "MSVAG",
+    "Muno", "Nero", "NovoGrad", "OrthoGrad", "PAdam", "PCGrad", "PID", "PNM", "Prodigy", "PSGD", "QHAdam", "QHM",
+    "RAdam", "Ranger", "Ranger21", "RotoGrad", "SAM", "ScheduleFreeSGD", "ScheduleFreeAdamW", "ScheduleFreeRAdam",
+    "SGDP", "Shampoo", "ScalableShampoo", "SGDW", "SignSGD", "SM3", "SOAP", "SopihaH", "SPAM", "SRMM", "StableAdamW",
+    "SWATS", "TAM", "Tiger", "TRAC", "WSAM", "Yogi", "BCE", "BCEFocal", "Focal", "FocalCosine", "SoftF1", "Dice",
+    "LDAM", "Jaccard", "Bi-Tempered", "Tversky", "FocalTversky", "LovaszHinge", "bitsandbytes", "WSD", "QGaLore",
 ]
 classifiers = [
     "License :: OSI Approved :: Apache Software License",
@@ -52,29 +52,27 @@ numpy = [
     { version = "<=1.24.4", python = "<3.9" },
 ]
 torch = { version = ">=1.10", python = ">=3.8", source = "torch" }
-bitsandbytes = { version = "^0.44", optional = true }
 
-[tool.poetry.dev-dependencies]
-isort = { version = "^5", python = ">=3.8" }
-black = { version = "^24", python = ">=3.8" }
+[tool.poetry.group.dev.dependencies]
+isort = [
+    { version = "<6", python = "<3.9" },
+    { version = "^6", python = ">=3.9" },
+]
+black = [
+    { version = "<25", python = "<3.9" },
+    { version = "^25", python = ">=3.9" },
+]
 ruff = "*"
 pytest = "*"
 pytest-cov = "*"
 
-[tool.poetry.extras]
-bitsandbytes = ["bitsandbytes"]
-
 [[tool.poetry.source]]
 name = "torch"
 url = "https://download.pytorch.org/whl/cpu"
 priority = "explicit"
 
 [tool.ruff]
-src = [
-    "pytorch_optimizer",
-    "tests",
-    "examples",
-]
+src = ["pytorch_optimizer", "tests", "examples"]
 target-version = "py312"
 line-length = 119
 exclude = [
@@ -97,8 +95,10 @@ select = [
     "TID", "ARG", "ERA", "RUF", "YTT", "PL", "Q"
 ]
 ignore = [
-    "A005", "B905", "D100", "D102", "D104", "D105", "D107", "D203", "D213", "D413", "PIE790", "PLR0912", "PLR0913",
-    "PLR0915", "PLR2004", "RUF013", "Q003", "ARG002",
+    "A005", "B905",
+    "D100", "D102", "D104", "D105", "D107", "D203", "D213", "D413",
+    "PLR0912", "PLR0913", "PLR0915", "PLR2004",
+    "Q003", "ARG002",
 ]
 fixable = ["ALL"]
 unfixable = ["F401"]
@@ -107,10 +107,10 @@ flake8-quotes.docstring-quotes = "double"
 flake8-quotes.inline-quotes = "single"
 
 [tool.ruff.lint.extend-per-file-ignores]
-"hubconf.py" = ["D", "INP001"]
 "examples/visualize_optimizers.py" = ["D103", "D400", "D415"]
-"**/__init__.py" = ["F401"]
 "{tests,examples}/*.py" = ["D", "S101"]
+"**/__init__.py" = ["F401"]
+"hubconf.py" = ["D", "INP001"]
 
 [tool.ruff.lint.isort]
 combine-as-imports = false
@@ -129,9 +129,9 @@ testpaths = "tests"
 
 [tool.coverage.run]
 omit = [
-    "./pytorch_optimizer/optimizer/rotograd.py",
     "./pytorch_optimizer/optimizer/adam_mini.py",
     "./pytorch_optimizer/optimizer/demo.py",
+    "./pytorch_optimizer/optimizer/rotograd.py",
 ]
 
 [build-system]

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/__init__.py

@@ -46,6 +46,7 @@ from pytorch_optimizer.optimizer import (
     ASGD,
     BSAM,
     CAME,
+    FOCUS,
     FTRL,
     GSAM,
     LARS,
@@ -105,6 +106,7 @@ from pytorch_optimizer.optimizer import (
     DeMo,
     DiffGrad,
     DynamicLossScaler,
+    EXAdam,
     FAdam,
     Fromage,
     GaLore,
@@ -112,6 +114,7 @@ from pytorch_optimizer.optimizer import (
     Gravity,
     GrokFastAdamW,
     Kate,
+    Kron,
     Lamb,
     LaProp,
     Lion,

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/base/optimizer.py

@@ -6,7 +6,16 @@ import torch
 from torch.optim import Optimizer
 
 from pytorch_optimizer.base.exception import NegativeLRError, NegativeStepError
-from pytorch_optimizer.base.types import BETAS, CLOSURE, DEFAULTS, HUTCHINSON_G, LOSS, PARAMETERS, STATE
+from pytorch_optimizer.base.types import (
+    BETAS,
+    CLOSURE,
+    DEFAULTS,
+    HUTCHINSON_G,
+    LOSS,
+    OPTIMIZER_INSTANCE_OR_CLASS,
+    PARAMETERS,
+    STATE,
+)
 
 
 class BaseOptimizer(ABC, Optimizer):
@@ -15,6 +24,18 @@ class BaseOptimizer(ABC, Optimizer):
     def __init__(self, params: PARAMETERS, defaults: DEFAULTS) -> None:
         super().__init__(params, defaults)
 
+    @staticmethod
+    def load_optimizer(optimizer: OPTIMIZER_INSTANCE_OR_CLASS, **kwargs) -> Optimizer:
+        r"""Build torch.optim.Optimizer class."""
+        if isinstance(optimizer, Optimizer):
+            return optimizer
+
+        if 'params' in kwargs:
+            params = kwargs.pop('params')
+            return optimizer(params, **kwargs)
+
+        raise ValueError('need to pass `params` when you pass the `torch.optim.Optimizer` instance.')
+
     @staticmethod
     @torch.no_grad()
     def set_hessian(param_groups: PARAMETERS, state: STATE, hessian: List[torch.Tensor]) -> None:
@@ -317,7 +338,7 @@ class BaseOptimizer(ABC, Optimizer):
     @staticmethod
     def validate_options(x: str, name: str, options: List[str]) -> None:
         if x not in options:
-            opts: str = ' or '.join([f'\'{option}\'' for option in options]).strip()
+            opts: str = ' or '.join([f"'{option}'" for option in options]).strip()
             raise ValueError(f'[-] {name} {x} must be one of ({opts})')
 
     @staticmethod

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/loss/jaccard.py

@@ -49,7 +49,7 @@ class JaccardLoss(_Loss):
     def __init__(
         self,
         mode: CLASS_MODE,
-        classes: List[int] = None,
+        classes: Optional[List[int]] = None,
         log_loss: bool = False,
         from_logits: bool = True,
         label_smooth: float = 0.0,
@@ -59,7 +59,7 @@ class JaccardLoss(_Loss):
 
         if classes is not None:
             if mode == 'binary':
-                raise ValueError('[-] Masking classes is not supported with mode=binary')
+                raise ValueError('masking classes is not supported with mode=binary')
 
             classes = torch.LongTensor(classes)
 

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/optimizer/__init__.py

@@ -40,8 +40,10 @@ from pytorch_optimizer.optimizer.came import CAME
 from pytorch_optimizer.optimizer.dadapt import DAdaptAdaGrad, DAdaptAdam, DAdaptAdan, DAdaptLion, DAdaptSGD
 from pytorch_optimizer.optimizer.demo import DeMo
 from pytorch_optimizer.optimizer.diffgrad import DiffGrad
+from pytorch_optimizer.optimizer.exadam import EXAdam
 from pytorch_optimizer.optimizer.experimental.ranger25 import Ranger25
 from pytorch_optimizer.optimizer.fadam import FAdam
+from pytorch_optimizer.optimizer.focus import FOCUS
 from pytorch_optimizer.optimizer.fp16 import DynamicLossScaler, SafeFP16Optimizer
 from pytorch_optimizer.optimizer.fromage import Fromage
 from pytorch_optimizer.optimizer.ftrl import FTRL
@@ -69,6 +71,7 @@ from pytorch_optimizer.optimizer.pcgrad import PCGrad
 from pytorch_optimizer.optimizer.pid import PID
 from pytorch_optimizer.optimizer.pnm import PNM
 from pytorch_optimizer.optimizer.prodigy import Prodigy
+from pytorch_optimizer.optimizer.psgd import Kron
 from pytorch_optimizer.optimizer.qhadam import QHAdam
 from pytorch_optimizer.optimizer.qhm import QHM
 from pytorch_optimizer.optimizer.radam import RAdam
@@ -289,8 +292,11 @@ OPTIMIZER_LIST: List[OPTIMIZER] = [
     LaProp,
     MARS,
     SGDSaI,
+    FOCUS,
     Grams,
     SPAM,
+    Kron,
+    EXAdam,
     Ranger25,
 ]
 OPTIMIZERS: Dict[str, OPTIMIZER] = {str(optimizer.__name__).lower(): optimizer for optimizer in OPTIMIZER_LIST}

pytorch_optimizer-3.4.0/pytorch_optimizer/optimizer/exadam.py

@@ -0,0 +1,123 @@
+import numpy as np
+import torch
+
+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 EXAdam(BaseOptimizer):
+    r"""The Power of Adaptive Cross-Moments.
+
+    :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.
+    :param fixed_decay: bool. fix weight decay.
+    :param eps: float. term added to the denominator to improve numerical stability.
+    """
+
+    def __init__(
+        self,
+        params: PARAMETERS,
+        lr: float = 1e-3,
+        betas: BETAS = (0.9, 0.999),
+        weight_decay: float = 0.0,
+        weight_decouple: bool = True,
+        fixed_decay: bool = False,
+        eps: float = 1e-8,
+        **kwargs,
+    ):
+        self.validate_learning_rate(lr)
+        self.validate_betas(betas)
+        self.validate_non_negative(weight_decay, 'weight_decay')
+        self.validate_non_negative(eps, 'eps')
+
+        self.sq2: float = np.sqrt(2)
+
+        defaults: DEFAULTS = {
+            'lr': lr,
+            'betas': betas,
+            'weight_decay': weight_decay,
+            'weight_decouple': weight_decouple,
+            'fixed_decay': fixed_decay,
+            'eps': eps,
+        }
+
+        super().__init__(params, defaults)
+
+    def __str__(self) -> str:
+        return 'EXAdam'
+
+    @torch.no_grad()
+    def reset(self):
+        for group in self.param_groups:
+            group['step'] = 0
+            for p in group['params']:
+                state = self.state[p]
+
+                state['exp_avg'] = torch.zeros_like(p)
+                state['exp_avg_sq'] = 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:
+            if 'step' in group:
+                group['step'] += 1
+            else:
+                group['step'] = 1
+
+            beta1, beta2 = group['betas']
+
+            bias_correction1: float = self.debias(beta1, group['step'])
+            bias_correction2: float = self.debias(beta2, group['step'])
+
+            step_size: float = group['lr'] * np.log(np.sqrt(group['step'] + 1) * self.sq2)
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+
+                grad = p.grad
+                if grad.is_sparse:
+                    raise NoSparseGradientError(str(self))
+
+                state = self.state[p]
+                if len(state) == 0:
+                    state['exp_avg'] = torch.zeros_like(p)
+                    state['exp_avg_sq'] = torch.zeros_like(p)
+
+                self.apply_weight_decay(
+                    p=p,
+                    grad=grad,
+                    lr=group['lr'],
+                    weight_decay=group['weight_decay'],
+                    weight_decouple=group['weight_decouple'],
+                    fixed_decay=group['fixed_decay'],
+                )
+
+                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
+                exp_avg.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1.0 - beta2)
+
+                d1 = 1.0 + exp_avg_sq.div(exp_avg_sq.add(group['eps'])) * (1.0 - bias_correction2)
+
+                exp_avg_p2 = exp_avg.pow(2)
+                d2 = 1.0 + exp_avg_p2.div(exp_avg_p2.add(group['eps'])) * (1.0 - bias_correction1)
+
+                m_tilde = exp_avg.div(bias_correction1) * d1
+                v_tilde = exp_avg_sq.div(bias_correction2) * d2
+
+                g_tilde = grad.div(bias_correction1) * d1
+
+                update = (m_tilde + g_tilde) / v_tilde.sqrt().add_(group['eps'])
+
+                p.add_(update, alpha=-step_size)
+
+        return loss

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/optimizer/experimental/ranger25.py

@@ -11,7 +11,7 @@ from pytorch_optimizer.base.types import BETAS, CLOSURE, DEFAULTS, LOSS, PARAMET
 class Ranger25(BaseOptimizer):
     r"""Mixin' every fancy optimizer hacks.
 
-    ADOPT + AdEMAMix + Cautious + StableAdamW + Adam-Atan2
+    ADOPT + AdEMAMix + Cautious + StableAdamW + Adam-Atan2 + OrthoGrad
 
     :param params: PARAMETERS. iterable of parameters to optimize or dicts defining parameter groups.
     :param lr: float. learning rate.
@@ -23,6 +23,7 @@ class Ranger25(BaseOptimizer):
     :param t_alpha_beta3: Optional[float]. total number of iterations is preferred when needed.
     :param cautious: bool. whether to use the Cautious variant.
     :param stable_adamw: bool. whether to use stable AdamW variant.
+    :param orthograd: bool. whether to use orthograd variant.
     :param eps: Optional[float]. term added to the denominator to improve numerical stability. when eps is None and
         stable_adamw is False, adam-atan2 feature will be used.
     """
@@ -39,6 +40,7 @@ class Ranger25(BaseOptimizer):
         t_alpha_beta3: Optional[float] = None,
         cautious: bool = True,
         stable_adamw: bool = True,
+        orthograd: bool = True,
         eps: Optional[float] = 1e-8,
         **kwargs,
     ):
@@ -51,6 +53,7 @@ class Ranger25(BaseOptimizer):
 
         self.cautious = cautious
         self.stable_adamw: bool = stable_adamw if isinstance(eps, float) else False
+        self.orthograd = orthograd
 
         defaults: DEFAULTS = {
             'lr': lr,
@@ -97,6 +100,21 @@ class Ranger25(BaseOptimizer):
             beta3,
         )
 
+    @torch.no_grad()
+    def orthogonalize_gradients(self, params, eps: float = 1e-16) -> None:
+        for p in params:
+            if p.grad is None or p.grad.is_sparse:
+                continue
+
+            w = p.view(-1)
+            g = p.grad.view(-1)
+
+            proj = torch.dot(w, g).div_(torch.dot(w, w).add_(eps))
+            g_ortho = g.to(dtype=torch.float32, copy=True).sub_(w, alpha=proj)
+            g_ortho_scaled = g_ortho.mul_(g.norm(2).div_(g_ortho.norm(2).add_(eps)))
+
+            p.grad.copy_(g_ortho_scaled.view_as(p.grad))
+
     @torch.no_grad()
     def step(self, closure: CLOSURE = None) -> LOSS:
         loss: LOSS = None
@@ -104,6 +122,10 @@ class Ranger25(BaseOptimizer):
             with torch.enable_grad():
                 loss = closure()
 
+        if self.orthograd:
+            for group in self.param_groups:
+                self.orthogonalize_gradients(group['params'])
+
         for group in self.param_groups:
             if 'step' in group:
                 group['step'] += 1

pytorch_optimizer-3.4.0/pytorch_optimizer/optimizer/focus.py

@@ -0,0 +1,95 @@
+import torch
+
+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 FOCUS(BaseOptimizer):
+    r"""First Order Concentrated Updating Scheme.
+
+    :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 gamma: float. control the strength of the attraction.
+    :param weight_decay: float. weight decay (L2 penalty).
+    """
+
+    def __init__(
+        self,
+        params: PARAMETERS,
+        lr: float = 1e-2,
+        betas: BETAS = (0.9, 0.999),
+        gamma: float = 0.1,
+        weight_decay: float = 0.0,
+        **kwargs,
+    ):
+        self.validate_learning_rate(lr)
+        self.validate_betas(betas)
+        self.validate_range(gamma, 'gamma', 0.0, 1.0, '[)')
+        self.validate_non_negative(weight_decay, 'weight_decay')
+
+        defaults: DEFAULTS = {'lr': lr, 'betas': betas, 'gamma': gamma, 'weight_decay': weight_decay}
+
+        super().__init__(params, defaults)
+
+    def __str__(self) -> str:
+        return 'FOCUS'
+
+    @torch.no_grad()
+    def reset(self):
+        for group in self.param_groups:
+            group['step'] = 0
+            for p in group['params']:
+                state = self.state[p]
+
+                state['exp_avg'] = torch.zeros_like(p)
+                state['pbar'] = 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:
+            if 'step' in group:
+                group['step'] += 1
+            else:
+                group['step'] = 1
+
+            beta1, beta2 = group['betas']
+
+            bias_correction2: float = self.debias(beta2, group['step'])
+
+            weight_decay: float = group['weight_decay']
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+
+                grad = p.grad
+                if grad.is_sparse:
+                    raise NoSparseGradientError(str(self))
+
+                state = self.state[p]
+                if len(state) == 0:
+                    state['exp_avg'] = torch.zeros_like(p)
+                    state['pbar'] = torch.zeros_like(p)
+
+                exp_avg, pbar = state['exp_avg'], state['pbar']
+
+                exp_avg.mul_(beta1).add_(grad, alpha=1.0 - beta1)
+                pbar.mul_(beta2).add_(p, alpha=1.0 - beta2)
+
+                pbar_hat = pbar / bias_correction2
+
+                if weight_decay > 0.0:
+                    p.add_(pbar_hat, alpha=-group['lr'] * weight_decay)
+
+                update = (p - pbar_hat).sign_().mul_(group['gamma']).add_(torch.sign(exp_avg))
+
+                p.add_(update, alpha=-group['lr'])
+
+        return loss

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/optimizer/ftrl.py

@@ -25,7 +25,7 @@ class FTRL(BaseOptimizer):
         beta: float = 0.0,
         lambda_1: float = 0.0,
         lambda_2: float = 0.0,
-        **kwargs
+        **kwargs,
     ):
         self.validate_learning_rate(lr)
         self.validate_non_negative(beta, 'beta')

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/optimizer/lookahead.py

@@ -29,13 +29,7 @@ class Lookahead(BaseOptimizer):
         self.validate_range(alpha, 'alpha', 0.0, 1.0)
         self.validate_options(pullback_momentum, 'pullback_momentum', ['none', 'reset', 'pullback'])
 
-        if isinstance(optimizer, Optimizer):
-            self.optimizer = optimizer
-        elif 'params' in kwargs:
-            params = kwargs.pop('params')
-            self.optimizer = optimizer(params, **kwargs)
-        else:
-            raise ValueError('Need to pass `params` when you pass the torch.optim.Optimizer instance.')
+        self.optimizer: Optimizer = self.load_optimizer(optimizer, **kwargs)
 
         self._optimizer_step_pre_hooks: Dict[int, Callable] = {}
         self._optimizer_step_post_hooks: Dict[int, Callable] = {}

{pytorch_optimizer-3.3.4 → pytorch_optimizer-3.4.0}/pytorch_optimizer/optimizer/muon.py

@@ -95,10 +95,10 @@ class Muon(BaseOptimizer):
 
         return new_params
 
-    def set_muon_state(self, params: PARAMETERS, adamw_params: PARAMETERS, threshold: int = 8192) -> None:
+    def set_muon_state(self, params: PARAMETERS, adamw_params: PARAMETERS) -> None:
         r"""Set use_muon flag."""
         for p in params:
-            self.state[p]['use_muon'] = p.ndim >= 2 and p.size(0) < threshold
+            self.state[p]['use_muon'] = p.ndim >= 2
 
         for p in adamw_params:
             self.state[p]['use_muon'] = False