import_q_scaler
For ordinary structural NEP, when training or resuming without fine_tune, nep computes q_scaler, the per-descriptor-component normalization used by the neural network, from the training structures of the current run at generation 0.
This keyword instead tells nep to read q_scaler from the nep.txt file present in the run directory, skipping that computation.
This naturally requires such a file to be present.
Syntax
import_q_scaler <0 or 1>
Parameters
The flag is 0 (default) or 1. A value of 1 imports the descriptor normalization from nep.txt.
Output
None.
Examples
import_q_scaler 1
Notes
Use this option to retain a model’s descriptor normalization when continuing training on different data.
Provide nep.txt with exactly the same architecture and species as the current run.
This option does not extract species or change the search state read from nep.restart.
For extraction from a foundation model, see fine_tune.
NEP-spin training and restart
Spin restart automatically restores the scaler from the matching nep.txt,
even with import_q_scaler 0. The model and restart must have compatible
architecture and element order; missing or invalid scalers are errors.
For a fresh Spin search, import_q_scaler 1 imports only the scaler.
Parameters are newly initialized and the energy reference is fitted to the
current training data. On restart, the saved energy reference and search state
are retained. Cross-species fine_tune remains unsupported for Spin.
See Descriptor scaling and Restart behavior for initialization
and restart details.