Network Architectures

Sage uses a two-stage neural network: a frontend that extracts per-detector temporal features from the compressed time-domain input, and a backend that fuses the multi-detector feature maps and produces a compact feature vector.

Two complete networks are provided, differing only in the uncertainty output of the regression head:

Class	Regression head
MSCNN1D_2DResNetCBAM	Point estimates only (used with BCEWithPEregLoss)
MSCNN1D_2DResNetCBAM_Heteroscedastic	Mean + log-variance (used with BCEWithPEsigmaLoss)

• Frontend: Multi-Scale 1D CNN
• Backend: 2D ResNet with CBAM
• Complete Networks