| # Debug Tools |
|
|
| ## Puzzle Eval-Stack Oracle Replay |
|
|
| The previous root-level `debug_puzzle.py` now lives in this directory. Run it |
| from the repository root: |
|
|
| ```bash |
| MUJOCO_GL=egl python debug/debug_puzzle.py --config-name=puzzle |
| ``` |
|
|
| ## PushT Multi-Step Latent Rollout Drift |
|
|
| `debug_pusht_rollout.py` freezes a trained HyperbolicJEPA checkpoint and replays |
| ground-truth PushT actions from the offline dataset. It compares each predicted |
| latent against the encoded real future frame for blocks `1...5`. |
|
|
| The script writes: |
|
|
| - `rollout_drift.png`: Euclidean, tangent-space, and Lorentz error curves. |
| - `rollout_drift_summary.json`: run metadata and aggregated metrics. |
| - `rollout_drift_summary.csv`: aggregated metrics for plotting or tables. |
| - `rollout_drift_per_sequence.csv`: per-window metrics for deeper analysis. |
|
|
| Each curve contains two modes: |
|
|
| - `autoregressive`: append each predicted latent and continue rollout. |
| - `teacher_forced`: restart each one-step prediction from real encoded frames. |
|
|
| Interpretation: |
|
|
| - Low teacher-forced error with rapidly rising autoregressive error indicates |
| compounding rollout drift. |
| - High error from block 1 indicates a one-step dynamics or action-protocol |
| issue. |
| - Similar low curves with poor planning SR point toward candidate sampling or |
| goal-cost alignment instead of dynamics quality. |
|
|
| Example: |
|
|
| ```bash |
| MUJOCO_GL=egl python debug/debug_pusht_rollout.py \ |
| --policy /data_nvme/user/zliu681/le-wm-main/lewm_cache/pusht/hyperbolic_pusht/lewm_hyperbolic_pusht_epoch_10 \ |
| --dataset pusht/pusht_expert_train \ |
| --device auto \ |
| --num-sequences 512 \ |
| --context-blocks 3 \ |
| --max-blocks 5 \ |
| --frameskip 5 |
| ``` |
|
|
