How lucas-maes/le-wm works
LeWorldModel is a compact JEPA world-model implementation: training learns latent next-state predictions from pixel sequences, while evaluation uses a loaded model to score candidate action rollouts for planning in control environments. The repository’s main internal layer is the JEPA model, built from encoder, action-embedding, prediction, and projection components, with a Gaussian latent regularizer in the training objective. Hydra configuration selects datasets, model/training settings, environments, and planning solvers. Training and evaluation depend on stable-pretraining and stable-worldmodel for orchestration and environment services; those libraries are external. Evidence for unsampled config wiring comes from the README and config paths, not inspected config contents.
Main entry points: eval.py Evaluation runner · train.py Training runner · jepa.py JEPA model · solver Planning solver
Components 20
Data and training
- HDF5 datasets
- Data preprocessing
utils.py - Training runner
train.py - Training framework
- Checkpoint saving
utils.py
JEPA model
- JEPA model
jepa.py - Image encoder
jepa.py - Action embedder
module.py - AR predictor
module.py - Projection heads
module.py - SIGReg regularizer
module.py
Planning and evaluation
- Evaluation runner
eval.py - Control environment
- World-model policy
- Planning solver
config/eval/solver - Checkpoint loader
- Evaluation results
eval.py
Configuration
- Training settings
config/train - Evaluation settings
config/eval
Other components
- Engineer
Connections 22
- Engineer → Training runner: starts training
README.md - Engineer → Evaluation runner: starts evaluation
README.md - Training settings → Training runner: configures
README.md - Evaluation settings → Evaluation runner: configures
README.md - Training runner → HDF5 datasets: loads data
train.py - Training runner → Data preprocessing: builds transforms
train.py - Training runner → Training framework: delegates training
train.py - Training runner → SIGReg regularizer: uses regularizer
train.py - Training runner → JEPA model: trains model
- Training framework → Checkpoint saving: runs callback
- JEPA model → Image encoder: encodes pixels
jepa.py - JEPA model → Action embedder: embeds actions
jepa.py - JEPA model → AR predictor: predicts embeddings
jepa.py - JEPA model → Projection heads: projects embeddings
jepa.py - Evaluation runner → HDF5 datasets: loads data
eval.py - Evaluation runner → Control environment: creates and evaluates
eval.py - Evaluation runner → Checkpoint loader: loads model
eval.py - Evaluation runner → Planning solver: instantiates solver
eval.py - Evaluation runner → World-model policy: creates policy
eval.py - Evaluation runner → Evaluation results: writes results
eval.py - World-model policy → Planning solver: uses solver
- Checkpoint loader → JEPA model: loads checkpoint