pulsatrix::MSELoss Class Reference

MSE = mean((prediction - target)^2). More...

#include <mse_loss.hpp>

Public Member Functions
	MSELoss (DeviceBackend *backend)
	Constructs an MSE loss.
float	forward (const Tensor &prediction, const Tensor &target)
	Computes the loss value and caches prediction/target for backward().
Tensor	backward () const
	Computes the gradient w.r.t. the prediction: (2/n) * (prediction - target).

Detailed Description

MSE = mean((prediction - target)^2).

Note

Not a Module subclass. Losses are the seed point relevance/gradient propagation starts from, not something a propagate_relevance rule is defined for – LRP explains a model's prediction, not the loss function used to train it. Deliberate scope decision (see mission_conv2d_losses.md's Objective 1), not an oversight.

Constructor & Destructor Documentation

MSELoss()

pulsatrix::MSELoss::MSELoss ( DeviceBackend *  backend )

explicit

Constructs an MSE loss.

Parameters

backend

Backend to compute through. Not owned; must outlive this loss.

Member Function Documentation

backward()

Tensor pulsatrix::MSELoss::backward

(

)

const

Computes the gradient w.r.t. the prediction: (2/n) * (prediction - target).

Returns

Gradient tensor, same shape as the prediction passed to forward().

Note

Must be called after forward() – uses the cached prediction/target.

Device-generic; the gradient is on the prediction's device.

forward()

float pulsatrix::MSELoss::forward	(	const Tensor &	prediction,
		const Tensor &	target
	)

Computes the loss value and caches prediction/target for backward().

Parameters

prediction	Model output.
target	Ground truth. Must match prediction's shape.

Returns

The scalar MSE value.

Exceptions

std::invalid_argument

if prediction and target are on different devices.

Note

Device-generic: runs on Cpu, Cuda or Hip tensors (GPU-native-kernels Mission 1).

---

The documentation for this class was generated from the following file:

• include/pulsatrix/mse_loss.hpp