Losses API¶

Losses consume a `ModelOutput` and a labeled `Batch`, returning a scalar (or elementwise, when supported) tensor for optimization. Select a loss with its `LOSSES` registry key in YAML; direct construction is useful for tests and custom loops. See Losses for choosing a task-appropriate objective.

Contract¶

chimera_ml.losses.base.BaseLoss ¶

Bases: ABC

Loss contract used by Trainer.

Implementations receive a ModelOutput and labeled Batch and return a differentiable tensor consumed by the optimizer.

Regression¶

`MSELoss` and `MAELoss` delegate to PyTorch reductions and expect predictions and targets with matching shapes.

chimera_ml.losses.regression.MSELoss ¶

Bases: BaseLoss

Mean squared error for matching prediction and target shapes.

Parameters:

Name	Type	Description	Default
`reduction`	`str`	PyTorch reduction mode: mean, sum, or none.	`'mean'`

chimera_ml.losses.regression.mse_loss ¶

mse_loss(**params)

chimera_ml.losses.regression.MAELoss ¶

Bases: BaseLoss

Mean absolute error for matching prediction and target shapes.

Parameters:

Name	Type	Description	Default
`reduction`	`str`	PyTorch reduction mode: mean, sum, or none.	`'mean'`

chimera_ml.losses.regression.mae_loss ¶

mae_loss(**params)

Classification¶

Cross entropy and focal loss consume class-index targets. Focal loss additionally supports focusing, alpha weighting, reduction, and label smoothing.

chimera_ml.losses.classification.CrossEntropyLoss ¶

Bases: BaseLoss

Cross-entropy loss for class logits and integer class targets.

Parameters:

Name	Type	Description	Default
`label_smoothing`	`float`	Amount of uniform label smoothing passed to PyTorch.	`0.0`

chimera_ml.losses.classification.cross_entropy_loss ¶

cross_entropy_loss(**params)

chimera_ml.losses.focal.FocalLoss ¶

Bases: BaseLoss

Multi-class focal loss for logits and integer class targets.

Parameters:

Name	Type	Description	Default
`gamma`	`float`	Focusing exponent; larger values down-weight easy examples.	`2.0`
`alpha`	`float \| list[float] \| None`	Optional scalar or per-class weighting factor.	`None`
`reduction`	`str`	mean, sum, or none.	`'mean'`
`label_smoothing`	`float`	Optional smoothing applied to the selected class probability.	`0.0`

chimera_ml.losses.focal.focal_loss ¶

focal_loss(**params)

Multilabel and agreement objectives¶

`BCEWithLogitsLoss` expects multi-hot targets. `CCCLoss` optimizes one minus Lin's concordance correlation coefficient across flattened output dimensions.

chimera_ml.losses.multilabel.BCEWithLogitsLoss ¶

Bases: BaseLoss

Binary cross-entropy loss for multi-label logits and targets.

Parameters:

Name	Type	Description	Default
`pos_weight`	`Tensor \| None`	Optional positive-class weighting tensor.	`None`
`reduction`	`str`	PyTorch reduction mode, typically mean, sum, or none.	`'mean'`

chimera_ml.losses.multilabel.bce_with_logits_loss ¶

bce_with_logits_loss(**params)

chimera_ml.losses.ccc.CCCLoss ¶

Bases: BaseLoss

Concordance Correlation Coefficient loss for regression.

The loss is 1 - mean(CCC) across output dimensions.

output.preds and batch.targets must have the same shape: (B,), (B, D), or (B, ...). Inputs are flattened to (B, Dflat).

chimera_ml.losses.ccc.ccc_loss ¶

ccc_loss(**params)