Getting Started

Installation

Install the package with pip into the environment where you will use it:

pip install active-learning-step

seamm-al select and seamm-al data need nothing more. Scoring, training, checking and the committee engine need torch and xnn 0.13 or later: run them in the MLFF environment SEAMM itself uses (see User Guide).

A first selection

Score a pool of candidate structures with a committee of five models, then choose the most uncertain within a budget of core-hours:

seamm-al score -m m0/best.pt m1/best.pt m2/best.pt m3/best.pt m4/best.pt \
    -o scores.csv candidates.extxyz
seamm-al select scores.csv --budget 5000 --cost cluster=4 --cost cell=2800

selection.csv lists the frames chosen, with each one’s rank and the running cost, and selected.extxyz holds them, copied unchanged from the pool, ready for labelling.

Before any committee exists, draw the first round at random instead, spread over the values of a column of the pool:

seamm-al score --no-models --carry source -o pool.csv candidates.extxyz
seamm-al select pool.csv --random 1 --stratify-column source \
    --cost cluster=4 --max-frames 50

In a SEAMM flowchart

Add an Active Learning step and, inside it, the sub-steps: for example Score Pool, taking the candidates from the job’s database or from files, then Select Frames with a budget. The sub-steps that need torch run in the MLFF environment of the xnn plug-in, set in the installation’s xnn.ini; nothing has to be installed there. Train Committee trains a committee and can save it as an xnn model, which the Model Chemistry step then offers for molecular dynamics that records where the members disagree. See In a flowchart: Active Learning step.

That should be enough to get started. For more detail, see the User Guide.