chowfi/JEPA-world-model
Train a JEPA world model on a set of pre-collected trajectories from an environment involving an agent in two rooms.
What's novel
Train a JEPA world model on a set of pre-collected trajectories from an environment involving an agent in two rooms.
Code Analysis
11 files read · 4 roundsTrains a JEPA-style self-supervised world model (CNN encoder + recurrent CNN predictor) on 2D agent-in-maze trajectories to predict future latent states conditioned on actions, then evaluates via a frozen-encoder probing head that regresses agent (x,y) location.
Strengths
The recurrent predictor architecture that rolls spatial latents forward one action at a time is a reasonable design choice, and the project demonstrates a working end-to-end pipeline from training to probing evaluation. The DDP adaptation, while incomplete, shows awareness of multi-GPU training patterns.
Weaknesses
The training loop contains a critical bug where the VICReg regularization terms are computed but never used in the actual loss, the target encoder is updated via EMA but never actually used to produce prediction targets (contradicting the README), and the hardcoded 4×4 spatial dimensions will crash on the actual 65×65 data. There are zero tests, duplicated code (Prober in two files), TODO comments left in production code, and the DDP file is an incomplete copy-paste with dummy data paths.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
5
Contributors
3
Files
53
Active weeks
2
Repository
Language
Python
Stars
3
Forks
1
License
—