def _register_adam_variants()

in timm/optim/_optim_factory.py [0:0]

• 94 lines of code
• 2 McCabe index (conditional complexity)

def _register_adam_variants(registry: OptimizerRegistry) -> None:
    """Register Adam-based optimizers"""
    adam_optimizers = [
        OptimInfo(
            name='adam',
            opt_class=torch.optim.Adam,
            description='torch.optim.Adam, Adaptive Moment Estimation',
            has_betas=True
        ),
        OptimInfo(
            name='adamw',
            opt_class=torch.optim.AdamW,
            description='torch.optim.AdamW, Adam with decoupled weight decay',
            has_betas=True
        ),
        OptimInfo(
            name='adamwlegacy',
            opt_class=AdamWLegacy,
            description='legacy impl of AdamW that pre-dates inclusion to torch.optim',
            has_betas=True
        ),
        OptimInfo(
            name='adamp',
            opt_class=AdamP,
            description='Adam with built-in projection to unit norm sphere',
            has_betas=True,
            defaults={'wd_ratio': 0.01, 'nesterov': True}
        ),
        OptimInfo(
            name='nadam',
            opt_class=torch.optim.NAdam,
            description='torch.optim.NAdam, Adam with Nesterov momentum',
            has_betas=True
        ),
        OptimInfo(
            name='nadamlegacy',
            opt_class=NAdamLegacy,
            description='legacy impl of NAdam that pre-dates inclusion in torch.optim',
            has_betas=True
        ),
        OptimInfo(
            name='nadamw',
            opt_class=NAdamW,
            description='Adam with Nesterov momentum and decoupled weight decay, mlcommons/algorithmic-efficiency impl',
            has_betas=True
        ),
        OptimInfo(
            name='radam',
            opt_class=torch.optim.RAdam,
            description='torch.optim.RAdam, Rectified Adam with variance adaptation',
            has_betas=True
        ),
        OptimInfo(
            name='radamlegacy',
            opt_class=RAdamLegacy,
            description='legacy impl of RAdam that predates inclusion in torch.optim',
            has_betas=True
        ),
        OptimInfo(
            name='radamw',
            opt_class=torch.optim.RAdam,
            description='torch.optim.RAdamW, Rectified Adam with variance adaptation and decoupled weight decay',
            has_betas=True,
            defaults={'decoupled_weight_decay': True}
        ),
        OptimInfo(
            name='adamax',
            opt_class=torch.optim.Adamax,
            description='torch.optim.Adamax, Adam with infinity norm for more stable updates',
            has_betas=True
        ),
        OptimInfo(
            name='adafactor',
            opt_class=Adafactor,
            description='Memory-efficient implementation of Adam with factored gradients',
        ),
        OptimInfo(
            name='adafactorbv',
            opt_class=AdafactorBigVision,
            description='Big Vision variant of Adafactor with factored gradients, half precision momentum',
        ),
        OptimInfo(
            name='adopt',
            opt_class=Adopt,
            description='Modified Adam that can converge with any β2 with the optimal rate',
        ),
        OptimInfo(
            name='adoptw',
            opt_class=Adopt,
            description='Modified AdamW (decoupled decay) that can converge with any β2 with the optimal rate',
            defaults={'decoupled': True}
        ),
    ]
    for opt in adam_optimizers:
        registry.register(opt)