How kentucky-kaneki/c-jepa-test works
T-JEPA is a tabular self-supervised representation learner, demonstrated as an Adult Income classification pipeline. The user-facing run loads and encodes data, pre-trains the model without labels, extracts context-encoder representations, trains a classifier, and reports test accuracy against a raw-feature baseline. The graph includes the documented feature embedding, masking, target encoder, EMA, and predictor stages; excerpts do not expose every constructor or call, so uncertain internal wiring is omitted.
Main entry points: run.py Pipeline runner · model.py T-JEPA model · trainer.py Pretraining loop · data.py Dataset loaders
Components 19
Experiment pipeline
- Pipeline runner
run.py - Raw-feature baseline
run.py - Accuracy comparison
run.py
Tabular data
- Adult loader
data.py - Preprocessor
data.py - Dataset loaders
data.py
T-JEPA model
- Feature embedding
model.py - Context-target masking
model.py - Context encoder
model.py - EMA target encoder
model.py - Prediction transformer
model.py - T-JEPA model
model.py
Training and evaluation
- Pretraining loop
trainer.py - Representation extraction
trainer.py - Downstream MLP
downstream.py - Classifier training
trainer.py - Accuracy evaluation
trainer.py
Other components
- Experiment user
- OpenML Adult
Connections 27
- Experiment user → Pipeline runner: starts
- Pipeline runner → Adult loader: loads data
- Adult loader → OpenML Adult: fetches Adult
- Pipeline runner → Preprocessor: builds loaders
- Preprocessor → Dataset loaders: encodes splits
- Pipeline runner → T-JEPA model: constructs
- Pipeline runner → Pretraining loop: invokes
- Pretraining loop → Dataset loaders: reads batches
- Pretraining loop → T-JEPA model: optimizes
- T-JEPA model → Context-target masking: samples masks
- T-JEPA model → Feature embedding: embeds features
- T-JEPA model → Context encoder: encodes context
- T-JEPA model → EMA target encoder: encodes targets
- T-JEPA model → Prediction transformer: predicts targets
- Pretraining loop → T-JEPA model: updates EMA
- Pipeline runner → Representation extraction: requests embeddings
- Representation extraction → Dataset loaders: reads batches
- Representation extraction → T-JEPA model: encodes features
- Pipeline runner → Downstream MLP: constructs
- Pipeline runner → Classifier training: trains classifier
- Classifier training → Downstream MLP: optimizes
- Pipeline runner → Accuracy evaluation: evaluates test
- Accuracy evaluation → Downstream MLP: scores
- Pipeline runner → Raw-feature baseline: trains baseline
- Pipeline runner → Accuracy comparison: reports accuracies
- Accuracy evaluation → Accuracy comparison: supplies score
- Raw-feature baseline → Accuracy comparison: supplies score