sage.architecture.custom_losses.loss_functions

Filename = custom_loss_functions.py Description = Repository of custom loss functions

Created on Fri Jan 28 19:08:44 2022

__author__ = nnarenraju __copyright__ = Copyright 2021, Sage __credits__ = nnarenraju __license__ = MIT Licence __version__ = 0.0.1 __maintainer__ = nnarenraju __email__ = nnarenraju@gmail.com __status__ = [‘inProgress’, ‘Archived’, ‘inUsage’, ‘Debugging’]

Github Repository: NULL

Documentation: NULL

Classes

BCEWithPEregLoss	Binary cross-entropy classification loss with MSE-based parameter
BCEWithPEsigmaLoss	Combined BCE + Heteroscedastic Regression Loss.

Module Contents

class BCEWithPEregLoss(regression_weight=1.0)

Bases: torch.nn.Module

Binary cross-entropy classification loss with MSE-based parameter estimation regularisation.

The total loss is:

L = BCE(ranking_stat, class_target)
  + regression_weight * MSE_signal(point_estimates, pe_targets)

where the MSE term is:

• computed only on signal samples (class_target == 1),

• weighted per-sample by the network’s current predicted signal probability p = sigmoid(ranking_stat) to focus regression updates on confident detections.

This is the simplest multi-task loss in Sage and does not model prediction uncertainty.

Parameters:

regression_weight (float) – Relative weight of the regression term vs. BCE.

Returns:

• torch.Tensor, shape (num_pe + 1,) – Stacked [total_loss, bce_loss, reg_loss, ...] (one entry per point-estimate parameter plus the total).

• Initialize internal Module state, shared by both nn.Module and ScriptModule.

regression_weight = 1.0

num_components

forward(outputs, targets)

Compute BCE + MSE regression loss.

Parameters:

• outputs (tuple) – (ranking_stat, point_estimates) ranking_stat: shape (B,) or (B, 1) — raw logits. point_estimates: shape (B, num_pe) — predicted parameters.

• targets (torch.Tensor, shape (B, num_pe + 1)) – Last column is the binary class label (0 = noise, 1 = signal). Preceding columns are the regression targets.

Returns:

[total_loss, bce_loss, reg_loss].

Return type:

torch.Tensor, shape (num_pe + 1,)

class BCEWithPEsigmaLoss(regression_weight=1.0, coupling_weight=1.0, beta=0.5, sigma_min=0.001, sigma_max=10.0, eps=1e-06)

Bases: torch.nn.Module

Combined BCE + Heteroscedastic Regression Loss.

• BCE for classification (ranking statistic).

• Regression term uses predicted mean and log-variance.

• Only computed for signal entries.

• Weighted per-sample by network’s predicted signal probability.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

• regression_weight (float)

• coupling_weight (float)

• beta (float)

• sigma_min (float)

• sigma_max (float)

• eps (float)

regression_weight = 1.0

coupling_weight = 1.0

beta

sigma_min

sigma_max

num_components

eps = 1e-06

forward(outputs, targets)

Compute heteroscedastic BCE + NLL regression + coupling loss.

Parameters:

• outputs (tuple) – (ranking_stat, point_estimates) ranking_stat: shape (B,) — raw classification logits. point_estimates: shape (B, 2 * num_pe) — concatenation of predicted means μ (first num_pe columns) and raw σ parameters (last num_pe columns), the latter mapped to a strictly-positive std via softplus (_sigma()).

• targets (torch.Tensor, shape (B, num_pe + 1)) – Last column is the binary class label; preceding columns are the physical regression targets.

Returns:

[total_loss, bce_loss, reg_loss, coupling_loss].

Return type:

torch.Tensor, shape (num_pe + 2,)