Losses

Keras to DL4J loss function mapping for model import

Keras Loss Functions

DL4J supports all standard Keras loss functions except logcosh. Loss function mapping is implemented in KerasLossUtils.

---

Loss Mapping Table

Keras Loss	DL4J Loss Class	Supported
mean_squared_error / mse	LossMSE	Yes
mean_absolute_error / mae	LossMAE	Yes
mean_absolute_percentage_error / mape	LossMAPE	Yes
mean_squared_logarithmic_error / msle	LossMSLE	Yes
squared_hinge	LossSquaredHinge	Yes
hinge	LossHinge	Yes
categorical_hinge	LossCategoricalHinge	Yes
logcosh	—	No
categorical_crossentropy	LossMCXENT	Yes
sparse_categorical_crossentropy	LossSparseMCXENT	Yes
binary_crossentropy	LossBinaryXENT	Yes
kullback_leibler_divergence / kld	LossKLD	Yes
poisson	LossPoisson	Yes
cosine_proximity	LossCosineProximity	Yes

---

Loss Descriptions

mean_squared_error

Mean of squared differences between predictions and targets. Standard regression loss.

MSE = mean((y_true - y_pred)^2)

Use for regression tasks with normally distributed errors.

mean_absolute_error

Mean of absolute differences. More robust to outliers than MSE.

MAE = mean(|y_true - y_pred|)

mean_absolute_percentage_error

Scale-independent percentage error.

MAPE = mean(|y_true - y_pred| / |y_true|) * 100

mean_squared_logarithmic_error

Useful when targets span several orders of magnitude.

MSLE = mean((log(1 + y_pred) - log(1 + y_true))^2)

squared_hinge

Squared hinge loss for binary classification with labels in {-1, +1}.

squared_hinge = mean(max(0, 1 - y_true * y_pred)^2)

hinge

Standard SVM-style hinge loss.

hinge = mean(max(0, 1 - y_true * y_pred))

categorical_hinge

Multi-class hinge loss variant.

categorical_crossentropy

Standard multi-class cross-entropy. Requires one-hot encoded targets and softmax output.

categorical_crossentropy = -sum(y_true * log(y_pred))

sparse_categorical_crossentropy

Same as categorical_crossentropy but accepts integer class indices rather than one-hot vectors.

binary_crossentropy

Binary cross-entropy for binary or multi-label classification with sigmoid output.

binary_crossentropy = -y_true * log(y_pred) - (1 - y_true) * log(1 - y_pred)

kullback_leibler_divergence

Measures divergence between two distributions.

KLD = sum(y_true * log(y_true / y_pred))

Use for variational autoencoders and distributional matching.

poisson

Poisson loss for count-based predictions.

poisson = mean(y_pred - y_true * log(y_pred))

cosine_proximity

Negative cosine similarity. Used for metric learning and similarity tasks.

cosine_proximity = -sum(y_true * y_pred) / (||y_true|| * ||y_pred||)

---

logcosh (Not Supported)

logcosh (log of the hyperbolic cosine of the prediction error) is not implemented in DL4J. Use mean_absolute_error as an alternative for outlier-robust regression tasks.

---

Notes

Loss functions are read from the Keras training configuration embedded in the HDF5 file. They are only present when the model was compiled (model.compile(...)) and saved with model.save(). When loading for inference only, pass enforceTrainingConfig=false to bypass training configuration parsing.