Module dfdx::losses

Standard loss functions such as mse_loss(), cross_entropy_with_logits_loss(), and more.

Functions

binary_cross_entropy_with_logits_loss

Binary Cross Entropy With Logits in numerically stable way.

cross_entropy_with_logits_loss

Cross entropy loss. This computes: -(logits.log_softmax() * target_probs).sum(-1).mean()

huber_loss

Huber Loss uses absolute error when the error is higher than beta, and squared error when the error is lower than beta.

kl_div_with_logits_loss

KL Divergence loss. This computes (target_probs * (target_probs.log() - logits.log_softmax())).sum(-1).mean()

mae_loss

Mean absolute error. This computes (&targ - pred).abs().mean()

mse_loss

Mean Squared Error. This computes (&targ - pred).square().mean().

rmse_loss

Root Mean square error. This computes (&targ - pred).square().mean().sqrt()

smooth_l1_loss

Smooth l1 loss (closely related to Huber Loss) uses absolute error when the error is higher than beta, and squared error when the error is lower than beta.