Models Overview

Every model is a pure nn.Module implementing forward() and compute_loss(), built the same way regardless of which SSL paradigm it belongs to:

model = ModelClass(config: dict, in_channels: int)
# Supervised additionally takes num_classes: int
model = Supervised(config, in_channels, num_classes)

No model reaches into a trainer, logger, or datamodule — see Architecture for why that boundary is enforced.

Which one should I use?

Situation	Model	Why
Want a simple baseline, minimal code	DGI	One discriminator, no augmentation to tune
Want the most robust default on node classification	GraphCL	Classic contrastive, NT-Xent. Within 0.5 points of the best linear-probe accuracy in all six columns of the citation benchmark (GCN and GIN); cost quadratic in the number of nodes or graphs per batch
Want to avoid choosing/tuning structural augmentations	AFGRL	Mines positives from topology (kNN ∩ adjacency + k-means), no augmentation
Need cost linear in the batch, no negatives	BGRL	Teacher-student. Close to its paper on ogbn-arxiv and to the best methods with a GCN encoder; sensitive to the backbone (far behind with GIN) and to the predictor's normalization
Need embeddings that are decorrelated/non-collapsing by construction	VICReg / Barlow Twins	Loss terms explicitly penalize redundancy across dimensions
Exploring something closer to ViT-style self-distillation	GraphDINO	Prototypes + teacher temperature warmup
Have labels and want a supervised reference point	Supervised	Same API, direct CrossEntropyLoss

Config dataclasses

Every model's config dict is validated by a dedicated dataclass in src/graphssl/config/schema.py before the model is built — a misconfigured run fails at construction time, not mid-training.

Model	Config dataclass
Supervised	SupervisedConfig
DGI	DGIConfig
GraphCL	GraphCLConfig
BGRL	BGRLConfig
AFGRL	AFGRLConfig
VICReg	VICRegConfig
Barlow Twins	BarlowTwinsConfig
GraphDINO	GraphDINOConfig

See the Config API reference for the full field list of each.