mmml physnet-evaluate¶
Evaluate PhysNet checkpoint.
Usage¶
mmml physnet-evaluate --help
Options¶
usage: mmml physnet-evaluate [-h] --checkpoint CHECKPOINT --data DATA
[-o OUTPUT_DIR] [--natoms NATOMS]
[--batch-size BATCH_SIZE] [--seed SEED]
[--num-samples NUM_SAMPLES]
[--subtract-atom-energies] [--subtract-mean]
[--plots] [--no-save-npz]
[--use-ema | --no-use-ema]
Evaluate PhysNetJAX checkpoint on NPZ (energies, forces, dipoles).
Input & configuration:
--checkpoint CHECKPOINT
PhysNet checkpoint root (directory containing epoch-*
orbax runs), same as mmml physnet-md --checkpoint
--data DATA NPZ with R, Z, N, E, F (and optionally D / Dxyz / dipole
if model predicts dipoles)
Execution:
--batch-size BATCH_SIZE
Batch size for inference (default: 16). Remainder
samples are skipped.
--seed SEED PRNG seed for batch shuffling (default: 0).
Output & artifacts:
-o, --output-dir OUTPUT_DIR
Directory for metrics.json and optional plots (default:
./physnet_evaluate_out)
--plots Write parity plots (requires matplotlib).
--no-save-npz Do not write predictions.npz (default: save).
Diagnostics & safety:
-h, --help show this help message and exit
Other options:
--natoms NATOMS Padded atom count (must match training). Default:
inferred from NPZ Z/R width.
--num-samples NUM_SAMPLES
If set, evaluate at most this many structures (after
shuffle split).
--subtract-atom-energies
Subtract atomic reference energies from E (same option
as training data prep).
--subtract-mean Subtract mean energy from E (training-style).
--use-ema, --no-use-ema
Evaluate the checkpoint's EMA params (default: on). Use
--no-use-ema for the live training weights.
Evaluate a trained PhysNet (PhysNetJAX) checkpoint on an NPZ dataset. Runs real
model inference (orbax checkpoint + EF forward), reports energy / force / dipole
errors in kcal/mol (and eV where noted), optional parity plots. Usage: mmml
physnet-evaluate --checkpoint out/ckpts/run --data splits/test.npz -o eval_out/
mmml physnet-evaluate --checkpoint out/ckpts/run --data splits/test.npz \
--natoms 64 --batch-size 32 --plots --num-samples 500