Struct Loss 

pub struct Loss;

Loss function implementations for neural network training.

Provides forward (loss computation) and backward (gradient computation) passes for common loss functions used in GNN training.

Numerical Stability

All loss functions use epsilon clamping and gradient clipping to prevent numerical instability with extreme prediction values (near 0 or 1).

Implementations

impl Loss

pub fn compute( loss_type: LossType, predictions: &Array2<f32>, targets: &Array2<f32>, ) -> Result<f32>

Compute the loss value between predictions and targets.

Arguments

• loss_type - The type of loss function to use
• predictions - Model predictions as a 2D array
• targets - Ground truth targets as a 2D array (same shape as predictions)

Returns

• Ok(f32) - The computed scalar loss value
• Err(GnnError) - If shapes don’t match or computation fails

Example

use ndarray::Array2;
use ruvector_gnn::training::{Loss, LossType};

let predictions = Array2::from_shape_vec((2, 2), vec![0.9, 0.1, 0.2, 0.8]).unwrap();
let targets = Array2::from_shape_vec((2, 2), vec![1.0, 0.0, 0.0, 1.0]).unwrap();
let loss = Loss::compute(LossType::Mse, &predictions, &targets).unwrap();
assert!(loss >= 0.0);

pub fn gradient( loss_type: LossType, predictions: &Array2<f32>, targets: &Array2<f32>, ) -> Result<Array2<f32>>

Compute the gradient of the loss with respect to predictions.

Arguments

• loss_type - The type of loss function to use
• predictions - Model predictions as a 2D array
• targets - Ground truth targets as a 2D array (same shape as predictions)

Returns

• Ok(Array2<f32>) - Gradient array with same shape as predictions
• Err(GnnError) - If shapes don’t match or computation fails

Example

use ndarray::Array2;
use ruvector_gnn::training::{Loss, LossType};

let predictions = Array2::from_shape_vec((2, 2), vec![0.9, 0.1, 0.2, 0.8]).unwrap();
let targets = Array2::from_shape_vec((2, 2), vec![1.0, 0.0, 0.0, 1.0]).unwrap();
let grad = Loss::gradient(LossType::Mse, &predictions, &targets).unwrap();
assert_eq!(grad.shape(), predictions.shape());