OptimizerConfig (Python)

The OptimizerConfig Cython class wraps the C++ optimizer_params_t struct and is used to configure the optimiser and learning-rate scheduler when calling Analysis.AddModel.

from AnalysisG.core.lossfx import OptimizerConfig

op = OptimizerConfig()
op.Optimizer = "Adam"
op.Scheduler = "StepLR"
op.lr = 1e-3
op.step_size = 10
op.gamma = 0.5

Optimiser / Scheduler Selection

Property	Default	Description
Optimizer	''	Optimiser name: "Adam", "SGD", "RMSprop", "AdaGrad", "LBFGS".
Scheduler	''	LR-scheduler name: "StepLR", "CyclicLR", "ExponentialLR".

Hyper-parameter Properties

Property	Default	Description
lr	0.0	Learning rate.
lr_decay	0.0	L2 learning-rate decay (AdaGrad).
weight_decay	0.0	L2 regularisation weight decay.
eps	0.0	Numerical stability epsilon (Adam, RMSprop).
alpha	0.0	RMSprop smoothing constant.
momentum	0.0	SGD/RMSprop momentum factor.
dampening	0.0	SGD dampening for momentum.
betas	(0, 0)	Adam (β₁, β₂) tuple.
amsgrad	False	Use AMSGrad variant of Adam.
nesterov	False	Use Nesterov momentum (SGD).
centered	False	Use centred RMSprop.
initial_accumulator_value	0.0	AdaGrad initial accumulator value.
step_size	1	StepLR period in epochs.
gamma	0.1	StepLR/ExponentialLR multiplicative decay factor.
max_iter	0	LBFGS maximum iterations per step.
max_eval	0	LBFGS maximum function evaluations per step.
history_size	0	LBFGS history size.
tolerance_grad	0.0	LBFGS gradient-norm convergence tolerance.
tolerance_change	0.0	LBFGS step-size convergence tolerance.