Loss Overview

The built-in trainer selects a task loss from problem_type. The same loss modules are public for ordinary PyTorch loops, and loss_for(problem_type) returns the standard module for a supported type.

Loss Table

Problem type	Loss	Output shape	Label shape
`binary`	`BinaryClassificationLoss`	`[batch]` or `[batch, 1]`	`[batch]`
`multiclass`	`MulticlassClassificationLoss`	`[batch, classes]`	`[batch]` integer IDs
`multilabel`	`MultilabelClassificationLoss`	`[batch, labels]`	`[batch, labels]`
`regression`	`RegressionLoss`	`[batch]`	`[batch]`
`forecasting`	`RegressionLoss`	`[batch]`	`[batch]`

An unsupported problem type raises ValueError from loss_for.

Classification

Binary and multilabel tasks use binary cross entropy with logits. Multiclass tasks use cross entropy over mutually exclusive class logits. Pass raw logits to loss modules; the sigmoid or softmax lives inside the loss, and activating twice silently flattens gradients.

Binary labels arrive as anything castable to float. Multiclass labels are integer class IDs, and multilabel labels are multi-hot float rows matching the logit shape.

Regression and Forecasting

RegressionLoss uses Huber loss with a configurable delta, which bounds the influence of outlier targets compared to plain mean squared error. Normalize targets during data preparation when their scale is large or highly skewed, and store the inverse transform with the trained artifact so predictions can be mapped back to real units.

from relational_transformers import RegressionLoss

loss = RegressionLoss(delta=2.0)
value = loss(predictions, labels)

Custom Loss Functions

A custom loss is any torch.nn.Module accepting predictions and labels. Two model outputs are the natural attachment points:

target_scores when adapting the published scalar decoder, for example with a class weight or a ranking objective;
target_features when the task needs its own head, giving the loss a [batch, 512] feature to build on.

logits = model.forward(batch, output="target_scores").scores
loss = my_loss(logits, labels)
loss.backward()

The custom PyTorch loop example shows where this slots into a full training step.