fix(loss): exclude mixed_type padding atoms from the training loss#5738

Merged

wanghan-iapcm merged 17 commits into

deepmodeling:masterfrom

wanghan-iapcm:fix-loss-mixed-type-padding

Jul 12, 2026

Commits on Jul 4, 2026

fix(loss): recover the per-atom mask in the pt loss so mixed_type ghosts can be excluded
Han Wang
committed
fix(loss): per-frame normalize dos/tensor losses to exclude mixed_type padding
Han Wang
committed
fix(loss): per-frame normalize the energy loss to exclude mixed_type padding
Han Wang
committed
test(loss): cover atom_ener Huber invariant and mask the rmse_f display metric
Han Wang
committed
fix(loss): normalize extensive property loss by per-frame real atom count
Han Wang
committed
fix(loss): per-frame normalize the spin energy loss to exclude mixed_type padding
Han Wang
committed
fix(loss): per-frame normalize the spin MAE loss paths for mixed_type padding
Han Wang
committed

Commits on Jul 5, 2026

Commits on Jul 9, 2026

fix(loss): drop dead None stores flagged by CodeQL
Han Wang
committed
test(loss): xfail-track force_mag MAE frame-normalization debt
Han Wang
committed

Commits on Jul 10, 2026

docs(loss): reference tracking issues for mixed_type padding follow-ups
Han Wang
committed

Commits on Jul 11, 2026

ci: re-trigger readthedocs build (prior build reaped by RTD resource limit)
Han Wang
committed

Commits on Jul 12, 2026

test(pt): run PropertyLoss padding tests on env.DEVICE
Han Wang
committed
fix(loss): make ener_spin force_mag MAE batch-size independent
Han Wang
committed
Merge upstream/master into fix-loss-mixed-type-padding
Han Wang
committed