Loss Functions Module

lossfx wraps PyTorch’s loss functions, optimisers, and learning-rate schedulers into a single C++ class, and provides weight-initialisation utilities for torch::nn::Sequential networks. It is used internally by model_template and exposed to Python via OptimizerConfig in AnalysisG.core.lossfx.

Class: `lossfx`

Header: <templates/lossfx.h>

Inheritance: tools, notification

Supported Loss Functions

The following PyTorch loss functions are selectable by name (case-insensitive):

Name string	PyTorch class
`"bce"`	`torch::nn::BCELoss`
`"bce_with_logits"`	`torch::nn::BCEWithLogitsLoss`
`"cosine_embedding"`	`torch::nn::CosineEmbeddingLoss`
`"cross_entropy"`	`torch::nn::CrossEntropyLoss`
`"ctc"`	`torch::nn::CTCLoss`
`"hinge_embedding"`	`torch::nn::HingeEmbeddingLoss`
`"huber"`	`torch::nn::HuberLoss`
`"kl_div"`	`torch::nn::KLDivLoss`
`"l1"`	`torch::nn::L1Loss`
`"margin_ranking"`	`torch::nn::MarginRankingLoss`
`"mse"`	`torch::nn::MSELoss`
`"multi_label_margin"`	`torch::nn::MultiLabelMarginLoss`
`"multi_label_soft_margin"`	`torch::nn::MultiLabelSoftMarginLoss`
`"multi_margin"`	`torch::nn::MultiMarginLoss`
`"nll"`	`torch::nn::NLLLoss`
`"poisson_nll"`	`torch::nn::PoissonNLLLoss`
`"smooth_l1"`	`torch::nn::SmoothL1Loss`
`"soft_margin"`	`torch::nn::SoftMarginLoss`
`"triplet_margin"`	`torch::nn::TripletMarginLoss`
`"triplet_margin_with_distance"`	`torch::nn::TripletMarginWithDistanceLoss`

Supported Optimisers

Name string (`opt_enum`)	PyTorch class
`"Adam"`	`torch::optim::Adam`
`"Adagrad"`	`torch::optim::Adagrad`
`"AdamW"`	`torch::optim::AdamW`
`"LBFGS"`	`torch::optim::LBFGS`
`"RMSprop"`	`torch::optim::RMSprop`
`"SGD"`	`torch::optim::SGD`

Supported Learning-Rate Schedulers

Name string (`scheduler_enum`)	PyTorch class
`"StepLR"`	`torch::optim::StepLR`
`"ReduceLROnPlateau"`	`torch::optim::ReduceLROnPlateauScheduler`

Public Fields

Field	Type	Description
`variable`	`std::string`	Name of the output feature this loss function is attached to.
`lss_cfg`	`loss_opt`	Loss-option struct (reduction type, smoothing, margin, …).

Public Methods

Signature	Description
`lossfx()`	Default constructor.
`lossfx(std::string var, std::string enx)`	Constructs with feature name var and loss-function name enx.
`torch::Tensor loss(torch::Tensor* pred, torch::Tensor* truth)`	Computes the loss using the configured loss function.
`torch::Tensor loss(torch::Tensor* pred, torch::Tensor* truth, loss_enum lss)`	Computes the loss using an explicit `loss_enum` value.
`bool build_loss_function()`	Builds the loss function from `lss_cfg`. Returns `false` on failure.
`bool build_loss_function(loss_enum lss)`	Builds the loss function for the given enum value.
`torch::optim::Optimizer* build_optimizer(optimizer_params_t* op, std::vector<torch::Tensor>* params)`	Constructs and returns the PyTorch optimiser.
`void build_scheduler(optimizer_params_t* op, torch::optim::Optimizer* opx)`	Wraps the given optimiser with a learning-rate scheduler.
`void weight_init(torch::nn::Sequential* data, mlp_init method)`	Applies the weight-initialisation strategy method to data.
`void loss_opt_string(std::string name)`	Parses a composite option string like `"cross_entropy\|mean"` and populates `lss_cfg`.
`void to(torch::TensorOptions*)`	Moves the loss function to the specified device.
`void step()`	Calls the learning-rate scheduler step (if a scheduler is configured).