CLI
Gracemaker
gracemaker [-h] [-l LOG] [-m] [-rl] [-r] [-rs RESTART_SUFFIX] [--no-intra-epoch-redo] [-p POTENTIAL] [-s] [-sf] [-e] [-nj] [--seed SEED] [-cm] [-t] [-cn CHECKPOINT_NAME] [--reset-epoch-and-step] [input]
Fitting utility for (graph) atomic cluster expansion potentials.
positional arguments:
input input YAML file, default: input.yaml
options:
-h, --help show this help message and exit
-l LOG, --log LOG log filename, default: log.txt
-m, --multigpu Single host/multi GPU distributed fit
-rl, --restart-latest
Restart from latest checkpoint (use separately from -r/-rs)
-r, --restart-best-test
Restart from latest best test checkpoint (use separately from -rs/-rl)
-rs RESTART_SUFFIX, --restart-suffix RESTART_SUFFIX
Suffix of checkpoint to restart from, i.e. .epoch_10 (use separately from -r/-rl)
--no-intra-epoch-redo
When resuming from a mid-epoch checkpoint
(intra_epoch_save=True), skip the interrupted epoch's
remaining batches entirely and bump the step counter
to the nominal end of that epoch. This keeps the LR
scheduler aligned with the original maxiter budget at
the cost of under-training by (nominal_end -
saved_step) gradient updates. Default is to
fast-forward the iterator and train only the unseen
tail.
-p POTENTIAL, --potential POTENTIAL
Potential configuration to load, model.yaml file
-s, --save-model Export model as TF saved model
-sf, --save--fs Export FS model as yaml to be loaded in CPP
-e, --eager Eager graph execution
-nj, --no-jit No JIT
--seed SEED Random seed (will overwrite value from input.yaml)
-cm, --check-model Check model consistency, without performing fit
-t, --template Generate a template 'input.yaml' file by dialog
-cn CHECKPOINT_NAME, --checkpoint-name CHECKPOINT_NAME
Explicit name of the checkpoint (omit .index suffix)
--reset-epoch-and-step
Reset epoch and step counters from prev. runs