Skip to content

developer_notes.md

Latest commit

 

History

History
35 lines (25 loc) · 1.23 KB

developer_notes.md

File metadata and controls

35 lines (25 loc) · 1.23 KB
Raw

Developer Notes

Checkpoint Compatibility

The neural model code has been cleaned up, but some module attribute names intentionally remain close to the original research code. The pretrained checkpoint files store PyTorch state_dict keys such as:

mapping
linear1.weight
qt.weight
project0.weight
A.weight
B.weight
C.weight
dec.project0.weight

Renaming those module attributes would break direct loading of checkpoints/true_new_64/epoch8_*.tar. For this reason:

  • Internal temporary variables in models.py were renamed for readability.
  • Public aliases were added, for example KoopmanSpace and KoopSpace.
  • Legacy-compatible aliases remain, for example Koop_space and decoder.
  • The old misspelled config key liner_dec is still emitted by the config bridge, while the runtime attribute is exposed as linear_decoder.

This keeps the code cleaner without requiring checkpoint conversion.

Refactor Boundary

The current open-source-ready path covers the control simulation:

Model checkpoint -> Koopman matrix construction -> MPC optimization -> simulation outputs

The original training and prediction-evaluation scripts depended on large HighD .mat files and were not moved into the default runnable package.