scimba_torch.flows.flow_trainer

Flow trainers for optimization and projection in flow spaces.

Classes

FlowTrainer(flow, full_data, **kwargs)	Abstract class for a nonlinear projector.
NaturalGradientFlowTrainer(flow, full_data, ...)	Natural gradient flow trainer for optimization.

class FlowTrainer(flow, full_data, **kwargs)

Bases: AbstractNonlinearProjector

Abstract class for a nonlinear projector.

This class defines a nonlinear projector with various projection options and an optimization method. It is used to solve projection problems in a given approximation space, using optimization methods.

Parameters:

• flow (DiscreteFlowSpace | ContinuousFlowSpace) – The flow space (discrete or continuous) where the projection takes place.

• full_data (tuple) – Tuple containing the full dataset for training.

• **kwargs (Any) – Additional parameters, such as the type of projection, losses, and optimizers.

space: DiscreteFlowSpace | ContinuousFlowSpace

The approximation space where the projection takes place.

y: Tensor

Target tensor for training.

full_data: tuple

Tuple containing the full dataset for training.

get_dof(flag_scope='all', flag_format='list')

Get degrees of freedom from the space.

Parameters:

• flag_scope (str) – Scope for the degrees of freedom. Defaults to “all”.

• flag_format (str) – Format for the degrees of freedom. Defaults to “list”.

Return type:

Any

Returns:

Degrees of freedom in the specified format.

sample_all_vars(**kwargs)

Sample variables from the full dataset.

Parameters:

**kwargs (Any) – Additional keyword arguments, including batch_size (int): Number of samples to draw. Defaults to 10.

Return type:

tuple

Returns:

A tuple containing sampled inputs and corresponding targets.

assembly_post_sampling(data, **kwargs)

Assemble the system after sampling.

Parameters:

• data (tuple) – Tuple containing sampled inputs and parameters.

• **kwargs (Any) – Additional keyword arguments, including flag_scope (str): scope for the last layer. Defaults to “all”.

Return type:

tuple

Returns:

A tuple containing the left-hand side and right-hand side of the system.

assembly(**kwargs)

Assembles the system of equations for the PDE.

(and weak boundary conditions if needed).

Parameters:

**kwargs (Any) –

Additional keyword arguments including:

• n_collocation (int): Number of collocation points for the PDE. Defaults to 1000.

• n_bc_collocation (int): Number of collocation points for the boundary conditions. Defaults to 1000.

Returns:

A tuple containing the assembled system of equations (Lo, f).

Return type:

tuple

class NaturalGradientFlowTrainer(flow, full_data, **kwargs)

Bases: FlowTrainer

Natural gradient flow trainer for optimization.

This class extends FlowTrainer to use natural gradient optimization with preconditioning and line search capabilities.

Parameters:

• flow (DiscreteFlowSpace | ContinuousFlowSpace) – The flow space (discrete or continuous) where the projection takes place.

• full_data (tuple) – Tuple containing the full dataset for training.

• **kwargs (Any) – Additional parameters including learning rate, line search options, etc.

default_lr: float

Default learning rate for the optimizer.

optimizer: OptimizerData

The optimizer used for parameter updates.

bool_linesearch: bool

Whether to use line search.

bool_preconditioner: bool

Whether to use preconditioning.

nb_epoch_preconditioner_computing: int

Number of epochs for preconditioner computation.

type_linesearch: str

Type of line search algorithm.

projection_data: dict

Data structure for projection settings.

preconditioner: EnergyNaturalGradientPreconditionerProjector

The preconditioner instance.

data_linesearch: dict

Parameters for line search configuration.