Custom Loss Functions

Sage provides two multi-task loss functions that combine binary classification of detection with continuous parameter estimation. Both are subclasses of torch.nn.Module and return a stacked tensor of loss components so the training loop can track each term separately.

All losses expect the network to return a tuple (ranking_stat, point_estimates) and a target tensor of shape (B, num_pe + 1) where the last column is the binary class label (0 = noise, 1 = signal).

• BCEWithPEregLoss
• BCEWithPEsigmaLoss