Loss Functions

Clorch implements the most common loss functions used in supervised learning. All loss functions are available in the clorch.nn.functional namespace.

Standard Losses

Mean Squared Error (MSE)

Standard loss for regression tasks.

(require '[clorch.nn.functional :as F])

(F/mse-loss input target)

L1 Loss (Mean Absolute Error)

More robust to outliers than MSE.

(F/l1-loss input target)

Smooth L1 Loss (Huber Loss)

Combines MSE and L1; stable and robust.

(F/smooth-l1-loss input target)

Classification Losses

Cross Entropy

Combines LogSoftmax and NLLLoss in one single class. Used for multi-class classification.

(F/cross-entropy logits targets)

NLL Loss (Negative Log Likelihood)

Used for multi-class classification (expects log-probabilities as input).

(F/nll-loss input target)

Binary Cross Entropy (BCE)

Used for binary classification.

;; Expects probabilities (after sigmoid)
(F/bce-loss input target)

;; More numerically stable: combines sigmoid + BCE
(F/bce-with-logits-loss input target)

Usage in Training Loop

(def loss (F/cross-entropy pred target))
(autograd/backward loss)

Loss tensors retain their autograd graphs. Extract required JVM values and let the per-batch scope close after the optimizer step; see Memory Management.