Loss Functions

Every loss is a standalone nn.Module in losses/, independently testable and usable outside any model — registered in the LOSSES registry via @LOSSES.register("name").

Registry name	Class	Used by
nt_xent	NTXentLoss	GraphCL
vicreg	VICRegLoss	VICReg
barlow_twins	BarlowTwinsLoss	Barlow Twins
cosine_regression	CosineRegressionLoss	BGRL, AFGRL
—	DINOLoss	GraphDINO

Combining losses

CombinedLoss mixes and weights any set of registered losses without touching model code:

from graphssl.losses import CombinedLoss

fn = CombinedLoss.from_config([
    {"name": "nt_xent", "weight": 0.7, "tau": 0.5},
    {"name": "vicreg",  "weight": 0.3, "invariance": 25.0, "variance": 25.0, "covariance": 1.0},
])
loss = fn(z1, z2)

Or programmatically:

from graphssl.losses import CombinedLoss, NTXentLoss, VICRegLoss

fn = CombinedLoss([(NTXentLoss(tau=0.5), 0.7), (VICRegLoss(), 0.3)])

Reference

See the Losses API for full signatures and the formula for each loss (also documented on each model's own page).