import_q_scaler

By default, when training or resuming a NEP potential 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. The syntax is:

import_q_scaler <0 or 1>

with a default value of 0.

This is intended for resuming training from a restart state (nep.restart) that was transplanted from a different model, e.g. a foundation model restricted to a subset of species via a species-selection tool. In that case, the transplanted network weights were tuned under the q_scaler of the original model, and recomputing a different one from a smaller or less diverse training set can force the weights to spend part of the optimization readapting to the new descriptor scale, rather than only to the new training data. Unlike fine_tune, this keyword does not perform any species extraction: the nep.txt file it reads from is expected to already have the exact same architecture and species as the current run (nep validates this and raises an error on a mismatch), which is why the file is referred to implicitly rather than by a path argument, matching the way nep.restart is referred to. The mean/standard-deviation restart state (mu/sigma) is unaffected by this keyword: it is unconditionally read from nep.restart, exactly as in any other non-fine_tune run.