Run files: fit.yaml

Each aj fit writes a fit.yaml file in its --out directory. The file contains:

• all options of the run, with their final values;
• a provenance block: the ace-jax version, the coupling library, the command line and a timestamp.

aj fit --config reads a run file.

Reproduce a fit

aj fit --config fit/fit.yaml --out fit_again

This command gives the same model and metrics as the run that wrote fit/fit.yaml.

Vary a fit

Flags on the command line override the values in the file:

aj fit --config fit/fit.yaml --max-degree 12 --out fit_deg12     # a bigger basis
aj fit --config fit/fit.yaml --train more.xyz --out fit_more     # different data
aj fit --config fit/fit.yaml --model other.npz --r0 2.4 --out f3 # a saved basis instead of `basis:`

The fit logs each value that a flag overrides.

Write one by hand

A run file needs only the keys that you want to set. All other keys take the aj fit default.

• The keys are the flag names, with underscores in place of dashes.
• The basis settings go in a basis: block, with the aj basis flag names.

# fit_si.yaml
train: train.xyz
test: test.xyz
e0: lsq
m_per_species: 0
opt: lbfgs
basis:
  order: 3
  max_degree: 10

If your labels are not stored as energy, forces and virial, add energy_key:, force_key: and virial_key: lines.

aj fit --config fit_si.yaml --out fit

• Relative paths in the file (train, test, ood, data, model, ...) are relative to the directory of the file. Thus a run file works from any directory. out is relative to the current directory.
• model: basis.npz (a saved basis) and a basis: block are alternatives. Give only one.
• elements is optional. Without it, the fit builds the basis for the species in the data.
• If a key is spelled incorrectly or a value is not valid, the error gives the key. If the key is nearly correct, the error also gives the correct key.

From Python

The Python pipeline takes the same settings as a FitConfig; see the Python API. FitConfig(model=...) accepts the path of a basis file, a Basis or a BasisSpec.