mmml extract-checkpoint-metrics¶
Plot training metrics from Orbax checkpoints.
Usage¶
mmml extract-checkpoint-metrics --help
Options¶
usage: mmml extract-checkpoint-metrics [-h] -o OUTPUT [--log-loss] [--quiet]
[--stride STRIDE]
[--max-epochs MAX_EPOCHS]
[--metrics-json METRICS_JSON] [--ef-only]
[--plot-style {dark,editorial_cm,editorial_dejavu_sans,editorial_dejavu_serif,editorial_stix,google,icml,mpl_classic,nature,science,tron,xmgrace}]
[--individual-dir INDIVIDUAL_DIR]
checkpoint_dir
Extract and plot training metrics from Orbax checkpoints
positional arguments:
checkpoint_dir Checkpoint directory containing epoch-* subdirectories
Execution:
--max-epochs MAX_EPOCHS
Cap the number of epoch checkpoints read after stride
(default: no cap).
Output & artifacts:
-o, --output OUTPUT Output plot file (PNG)
--log-loss Use log scale for loss axes (recommended)
--metrics-json METRICS_JSON
Optional path to write extracted metrics as JSON arrays.
--plot-style {dark,editorial_cm,editorial_dejavu_sans,editorial_dejavu_serif,editorial_stix,google,icml,mpl_classic,nature,science,tron,xmgrace}
Matplotlib style preset (default: google). Options:
nature, xmgrace, google, tron, mpl_classic.
Diagnostics & safety:
-h, --help show this help message and exit
--quiet Suppress output
Other options:
--stride STRIDE Read every Nth epoch checkpoint (default: 1 = all). Use
for large runs.
--ef-only Plot energy/forces panels only (omit dipole inset from
main layout).
--individual-dir INDIVIDUAL_DIR
If set, write one PNG per metric into this directory.
Examples: # Plot glycol training with log scale python -m
mmml.cli.extract_checkpoint_metrics \
examples/glycol/checkpoints/glycol_production/glycol_production-*/ \ --output
glycol_training.png \ --log-loss # Without log scale python -m
mmml.cli.extract_checkpoint_metrics \ checkpoints/run/run-uuid/ \ --output
training.png