Losses

The Loss implementations measure the difference between model predictions and target spectra. Encodings are applied before loss calculation; see Encodings. They are configured as a non-empty list under trainer.loss. XANESNET builds a combined loss from the selected terms before each training update.

Available implementations

• mse (MSELoss) computes mean squared error.

• l1 (L1Loss) computes mean absolute error.

• bce (BCEWithLogitsLoss) computes binary cross-entropy with logits.

• emd (EMDLoss) compares the cumulative distributions of prediction and target spectra.

• wcc (WCCLoss) computes weighted cross-correlation.

• specplus (SpectralLossPlus) combines blurred, detail, and gradient terms.

• msssim (MultiScale_SSIM) computes a multi-scale structural-similarity loss for one-dimensional spectra.

Common configuration

Each entry may specify a positive loss_weight. Missing weights default to 1.0 and all weights are normalised to sum to one.

Distributed MSSSIM

With multi-process Lightning training, every msssim component must specify the data_range in the encoded target space. Choose the range for the actual spectra after encoding, not just the raw input. data_range: null tracks expanding bounds only during single-process training to avoid heavy communication overhead between distributed workers.

Example

trainer:
  loss:
    - loss_type: mse
      loss_weight: 0.7
    - loss_type: wcc
      loss_weight: 0.3
      gaussian_hwhm: 10

API reference

See also xanesnet.losses and xanesnet/schemas/losses/ for complete fields and defaults.