def _register_other_optimizers()

in timm/optim/_optim_factory.py [0:0]
119 lines of code
2 McCabe index (conditional complexity)

def _register_other_optimizers(registry: OptimizerRegistry) -> None:
    """Register miscellaneous optimizers"""
    other_optimizers = [
        OptimInfo(
            name='adabelief',
            opt_class=AdaBelief,
            description='Adapts learning rate based on gradient prediction error',
            has_betas=True,
            defaults={'rectify': False}
        ),
        OptimInfo(
            name='radabelief',
            opt_class=AdaBelief,
            description='Rectified AdaBelief with variance adaptation',
            has_betas=True,
            defaults={'rectify': True}
        ),
        OptimInfo(
            name='adadelta',
            opt_class=torch.optim.Adadelta,
            description='torch.optim.Adadelta, Adapts learning rates based on running windows of gradients'
        ),
        OptimInfo(
            name='adagrad',
            opt_class=torch.optim.Adagrad,
            description='torch.optim.Adagrad, Adapts learning rates using cumulative squared gradients',
            defaults={'eps': 1e-8}
        ),
        OptimInfo(
            name='adan',
            opt_class=Adan,
            description='Adaptive Nesterov Momentum Algorithm',
            defaults={'no_prox': False},
            has_betas=True,
            num_betas=3
        ),
        OptimInfo(
            name='adanw',
            opt_class=Adan,
            description='Adaptive Nesterov Momentum with decoupled weight decay',
            defaults={'no_prox': True},
            has_betas=True,
            num_betas=3
        ),
        OptimInfo(
            name='adahessian',
            opt_class=Adahessian,
            description='An Adaptive Second Order Optimizer',
            has_betas=True,
            second_order=True,
        ),
        OptimInfo(
            name='kron',
            opt_class=Kron,
            description='PSGD optimizer with Kronecker-factored preconditioner',
            has_momentum=True,
        ),
        OptimInfo(
            name='kronw',
            opt_class=Kron,
            description='PSGD optimizer with Kronecker-factored preconditioner and decoupled weight decay',
            has_momentum=True,
            defaults={'decoupled_decay': True}
        ),
        OptimInfo(
            name='laprop',
            opt_class=LaProp,
            description='Separating Momentum and Adaptivity in Adam',
            has_betas=True,
        ),
        OptimInfo(
            name='lion',
            opt_class=Lion,
            description='Evolved Sign Momentum optimizer for improved convergence',
            has_eps=False,
            has_betas=True
        ),
        OptimInfo(
            name='madgrad',
            opt_class=MADGRAD,
            description='Momentum-based Adaptive gradient method',
            has_momentum=True
        ),
        OptimInfo(
            name='madgradw',
            opt_class=MADGRAD,
            description='MADGRAD with decoupled weight decay',
            has_momentum=True,
            defaults={'decoupled_decay': True}
        ),
        OptimInfo(
            name='mars',
            opt_class=Mars,
            description='Unleashing the Power of Variance Reduction for Training Large Models',
            has_betas=True,
        ),
        OptimInfo(
            name='novograd',
            opt_class=NvNovoGrad,
            description='Normalized Adam with L2 norm gradient normalization',
            has_betas=True
        ),
        OptimInfo(
            name='rmsprop',
            opt_class=torch.optim.RMSprop,
            description='torch.optim.RMSprop, Root Mean Square Propagation',
            has_momentum=True,
            defaults={'alpha': 0.9}
        ),
        OptimInfo(
            name='rmsproptf',
            opt_class=RMSpropTF,
            description='TensorFlow-style RMSprop implementation, Root Mean Square Propagation',
            has_momentum=True,
            defaults={'alpha': 0.9}
        ),
    ]
    for opt in other_optimizers:
        registry.register(opt)
    registry.register_foreach_default('lion')