The Command Line#
Installing xnn provides the xnn command with five subcommands. Reading
structure files needs the ase extra. Every --ckpt accepts a trainer
checkpoint, a portable model directory, a hub name, a URL or a Zenodo DOI.
xnn train#
xnn train --config configs/train.yaml
xnn train --config configs/train.yaml --set optim.epochs=50 model.cutoff=6.0
torchrun --nproc-per-node 2 -m xnn train --config configs/train.yaml # data parallel
Reads data.train_path / val_path / test_path (and
replay_path for multi-head fine-tuning) with ASE, trains, and writes
best.pt and last.pt to output_dir. --set KEY=VALUE overrides
any config key. See Train a Model.
xnn benchmark#
xnn benchmark --config configs/benchmark.yaml
xnn benchmark --config configs/benchmark.yaml --set "metrics={'energy': ['mae']}"
Scores pre-trained checkpoints on one dataset and writes the results table in every configured format. See Benchmark Several Models.
xnn export#
xnn export --ckpt runs/exp/best.pt --out deployed.pt
xnn export --ckpt mace-off23-small --out mace_off.pt
Writes a self-contained TorchScript file with the whole-system entry point
forward and the pair-style forward_lammps. Options: --config for
a checkpoint without an embedded config, --head for a multi-head
checkpoint, --total-charge and --spin-multiplicity for the charge
state, --no-dispersion to skip a recorded subtracted dispersion. See
Export to LAMMPS and TorchScript.
xnn mdi#
xnn mdi --ckpt runs/exp/best.pt -mdi "-role ENGINE -name xnn -method TCP -port 8021 -hostname localhost"
Serves a checkpoint as a MolSSI Driver Interface engine (mdi extra) for
LAMMPS fix mdi/qm or any MDI driver. Options: --device, --dtype
(default: the checkpoint’s), --dispersion d3|d4|"{...}" to add a term
to a plain checkpoint, --no-dispersion, --head, --total-charge,
--eeq-reuse (carry the D4 charge solve between MD steps) and --fast
auto|on|off. See Deployment.
xnn models#
xnn models list --cached # registered and cached models
xnn models info mace-mh-0 # the model card
xnn models pull mace-off23-small # download and convert into the cache
xnn models pack runs/exp/best.pt my-model --license MIT --archive