Loss Functions

A loss function (or criterion) is a fundamental component of training a neural network. It calculates a single value that measures how well the model's prediction matches the target label. The goal of training is to minimize this value.

Standard LibTorch Losses

LibTorch provides a comprehensive set of standard loss functions that are sufficient for most common tasks. These can be used directly within any xTorch project.

Common Examples:

• torch::nn::MSELoss: Mean Squared Error, for regression.
• torch::nn::L1Loss: Mean Absolute Error, for regression.
• torch::nn::CrossEntropyLoss: For multi-class classification.
• torch::nn::BCELoss: Binary Cross Entropy, for binary classification.
• torch::nll_loss: Negative Log Likelihood Loss, often used with LogSoftmax.

For a complete list and usage instructions, please refer to the official PyTorch C++ API documentation.

xTorch Extended Losses

For more advanced or specialized tasks—such as object detection, metric learning, or dealing with class imbalance—standard loss functions may not be optimal. To address this, xTorch provides a rich library of modern and specialized loss functions from the research literature.

These implementations are ready to use, allowing you to easily experiment with advanced training objectives.

Usage

All xTorch loss functions are implemented as torch::nn::Modules and are located in the xt::losses namespace. They can be instantiated and used just like any standard torch::nn module.

They integrate seamlessly with the xt::Trainer.

#include <xtorch/xtorch.hh>
 
int main() {
    // Assume we have model outputs and targets
    auto output = torch::randn({16, 10}); // Batch of 16, 10 classes
    auto target = torch::randint(0, 10, {16}); // 16 target labels
 
    // --- Standalone Usage ---
    // Instantiate a loss function with its options
    xt::losses::FocalLoss focal_loss(xt::losses::FocalLossOptions().gamma(2.0));
    // Calculate the loss
    torch::Tensor loss = focal_loss(output, target);
    std::cout << "Calculated Focal Loss: " << loss.item<float>() << std::endl;
 
 
    // --- Integration with xt::Trainer ---
    xt::Trainer trainer;
    // The trainer can accept any callable, including our instantiated module
    trainer.set_loss_fn(focal_loss);
 
    // Now, when trainer.fit() is called, it will use the FocalLoss internally.
    // trainer.fit(model, data_loader, nullptr, device);
}

Available Loss Functions

Below is the comprehensive list of specialized loss functions available in the xt::losses module.

!!! info "Constructor Options" Many of these loss functions have tunable hyperparameters (like gamma in FocalLoss or margin in TripletLoss). These are configured via an Options struct passed to the constructor. Please refer to the specific header file in <xtorch/losses/> for details on the available settings.