Loss Layers

Loss layers are Module wrappers around the functional loss ops, making it easy to slot a loss function into a model definition or training loop as a named module. CELoss wraps cross-entropy loss for classification and MSELoss wraps mean squared error for regression. Neither layer has trainable parameters.

import simplegrad as sg
import simplegrad.nn as nn

criterion = nn.CELoss()
logits = sg.normal((8, 10), requires_grad=True)
targets = sg.Tensor([0, 1, 2, 3, 4, 5, 6, 7])
loss = criterion(logits, targets)
loss.backward()

All loss layers inherit from Module and share its methods.

---

CELoss

CELoss

Bases: Module

Cross-entropy loss layer with built-in softmax.

Parameters:

• dim (int, default: -1 ) –

  Class dimension. Defaults to -1 (last dim).

• reduction (str | None, default: 'mean' ) –

  "mean", "sum", or None. Defaults to "mean".

Attributes

Attribute	Type	Description
.dim	int	Axis over which softmax is computed. Defaults to -1.
.reduction	str \| None	"mean", "sum", or None for no reduction.

Methods

Method	Description
.forward()	Compute cross-entropy loss from logits and integer targets.

---

MSELoss

MSELoss

Bases: Module

Mean squared error loss layer.

Parameters:

• reduction (str | None, default: 'mean' ) –

  "mean", "sum", or None. Defaults to "mean".

Attributes

Attribute	Type	Description
.reduction	str \| None	"mean", "sum", or None for no reduction.

Methods

Method	Description
.forward()	Compute mean squared error between predictions and targets.