Basic Shell Workflow

This short introduces a basic workflow where models are specified via JSON files and model fitting is achieved via a shell script.

At the start of a new project we

  • create a project folder
  • activate a Julia project in that folder
  • add ACEpotentials to the Julia project
  • generate a fitting script and example model specification file
mkdir myace
cd myace
julia --project=. -e 'using Pkg; Pkg.add("ACEpotentials"); using ACEpotentials; ACEpotentials.copy_runfit(@__DIR__())'

This should create two new files in the myace folder:

  • runfit.jl
  • example_params.json

Copy (or move) the example_params.json file to a new filename, e.g. 

cp example_params.json myace_params_1.json

then edit that file to specify the model hyperparameters, the fitting method (see also ACEfit.jl), and the path to the dataset (or, datasets if validation and or test sets are also provided). To produce a fit, use 

julia --project=. runfit.jl -p myace_params_1.json -o results_1

This will write all outputs to the ./results_1 folder, in particular results.json which contains the model specification, the fitted model parameters, and a dictionary of computed errors (rmse, mae). The list of required outputs and the output filename(s) can be changed in the model spec json.