mmml compare-npz¶
Reference vs model NPZ plots.
Usage¶
mmml compare-npz --help
Options¶
usage: mmml compare-npz [-h] [--reference REFERENCE] [--predictions PREDICTIONS]
[--checkpoint CHECKPOINT] [--data DATA] [-o OUTPUT_DIR]
[--max-frames MAX_FRAMES] [--stride STRIDE]
[--cutoff CUTOFF] [--use-dcmnet-dipole]
[--energy-unit ENERGY_UNIT] [--force-unit FORCE_UNIT]
[--no-plots] [--save-predictions]
Compare reference and model NPZ data (metrics + plots).
Input & configuration:
--checkpoint CHECKPOINT
Model checkpoint JSON/pkl/dir; run inference on --data
--data DATA Labeled NPZ for --checkpoint mode (R,Z,E,F,...)
Scientific model:
--cutoff CUTOFF Model cutoff override for checkpoint inference
--energy-unit ENERGY_UNIT
Energy unit label for plots (default: infer from
reference NPZ)
--force-unit FORCE_UNIT
Force unit label for plots (default: eV/Å)
Output & artifacts:
-o, --output-dir OUTPUT_DIR
Output directory for metrics and plots
--no-plots Skip matplotlib plots
--save-predictions With --checkpoint, save inference NPZ to output dir
Diagnostics & safety:
-h, --help show this help message and exit
Other options:
--reference REFERENCE
Reference NPZ (PySCF / QM labels)
--predictions PREDICTIONS
Model prediction NPZ (E, F, D, ...)
--max-frames MAX_FRAMES
Max structures to compare (default: all)
--stride STRIDE Frame stride for --checkpoint mode (default: 1)
--use-dcmnet-dipole Use DCMNet dipole from joint checkpoint
Compare reference (PySCF/QM) and model NPZ trajectories with metrics and plots.
Modes ----- 1. Two NPZ files (reference labels vs model predictions): mmml
compare-npz --reference ref.npz --predictions pred.npz -o out/ 2. Checkpoint
inference against labeled NPZ (same file holds R,Z,E,F,...): mmml compare-npz
--checkpoint params.json --data test.npz -o out/ --max-frames 200 Issue #12:
per-atom / per-element force analysis and richer validation plots.