ML4OPF Loss Functions

Sub-modules

ml4opf.loss_functions.ldf

Formulation-agnostic Lagrangian Dual Framework Loss Function.

ml4opf.loss_functions.objective

Use the formulation objective as the loss function

ml4opf.loss_functions.penalty

Penalize constraint violations

Classes

class LDFLoss (v: OPFViolation,
step_size: float,
kickin: int,
update_freq: int,
divide_by_counter: bool = True,
exclude_keys: str | list[str] | None = None)

Browse git

LDFLoss implements the Lagrangian Dual Framework.

exclude_keys is either None to use all violations, "all" to skip all violations, or a list of keys to skip specific violations.

Initialize LDFLoss module.

Ancestors

• torch.nn.modules.module.Module

Methods

def end_epoch(self)

Browse git

Call this method at the end of each epoch.

def forward(self,
base_loss: torch.Tensor,
exclude_keys: str | list[str] | None = None,
**calc_violation_inputs: torch.Tensor) ‑> torch.Tensor

Browse git

Compute the LDF Loss for a batch of samples.

def init_mults(self, shapes=None)

Browse git

Initialize λ and trackers to zeros.

def reset_trackers(self)

Browse git

Reset the violation trackers to zeros.

def start_epoch(self, epoch) ‑> str | None

Browse git

Call this method at the start of each epoch.

def update(self)

Browse git

Update the lagrangian dual multipliers (λ)

class ObjectiveLoss (v: OPFViolation,
reduction: str | None = 'mean')

Browse git

ObjectiveLoss is the original objective of the OPF.

It takes as input the same arguments as the corresponding formulation's ml4opf.loss_functions.objective method, and returns the objective value.

Initialize ObjectiveLoss module.

Args

v : OPFViolation

OPFViolation module.

reduction : Optional[str]

Reduction operation. Default: "mean".

Ancestors

• torch.nn.modules.module.Module

Class variables

var SUPPORTED_REDUCTIONS

Methods

def forward(self, *objective_args, **objective_kwargs) ‑> torch.Tensor

Browse git

Compute the objective value for a batch of samples.