Loss and Metrics¶
Loss functions and metrics are configured by specifying a field (e.g. total_energy, forces, etc.) and an error quantity to calculate for it (e.g. MeanSquaredError, MeanAbsoluteError, etc).
Loss functions and metrics are configured through MetricsManager objects in the training_module section of the config.
The loss function determines what the model optimizes during training, while metrics are used for monitoring training progress and conditioning training behavior (early stopping, learning rate scheduling, etc.).
Units¶
All loss components and metrics are in the physical units associated with the dataset. For example, if the dataset uses force units of eV/Å, a force mean-squared error (MSE) would have units of (eV/Å)².
Simplified Wrappers¶
Most users should use the simplified wrapper classes for common force field training scenarios. These wrappers automatically configure the appropriate metrics for you:
For Loss Functions:
EnergyOnlyLoss(for energy-only datasets)
For Validation/Test Metrics:
EnergyOnlyMetrics(for energy-only datasets)
When using simplified wrappers, the actual metric names logged during training may not be immediately obvious. Each wrapper creates specific metrics with predetermined names. To see exactly what metric names each wrapper produces, refer to their individual API documentation in the nequip.train metrics API reference.
Coefficients and Weighted Sum¶
Users can set coefficients (coeff) for each loss or metric term, which leads to the computation of a weighted_sum metric.
For loss functions,
weighted_sumis the actual loss value used for backpropagation.For validation/test metrics,
weighted_sumprovides a single monitoring metric that balances multiple quantities to be used for conditioning checkpointing, early stopping, learning rate scheduling, etc.
Coefficients are automatically normalized to sum to 1. For example:
coeffs:
total_energy: 3.0
forces: 1.0
becomes internally: total_energy: 0.75, forces: 0.25.
The weighted_sum is calculated as:
weighted_sum = (coeff_1 * metric_1) + (coeff_2 * metric_2) + ...
Coefficients only affect the weighted_sum calculation. The individual metrics (e.g., total_energy_rmse, forces_rmse) are logged with their actual computed values, unmodified by coefficients.
Metrics with coeff: null (or omitted from coeffs) are still computed and logged, but excluded from weighted_sum:
coeffs:
total_energy_rmse: 1.0 # included in weighted_sum
forces_rmse: 1.0 # included in weighted_sum
total_energy_mae: null # computed but not in weighted_sum
forces_mae: null # computed but not in weighted_sum
Here’s an example showing how to set up metrics and use weighted_sum for monitoring:
# Define the monitored metric once for consistency
monitored_metric: val0_epoch/weighted_sum
training_module:
_target_: nequip.train.EMALightningModule
# Loss function
loss:
_target_: nequip.train.EnergyForceLoss
coeffs:
total_energy: 1.0
forces: 1.0
# Validation metrics - weighted_sum will be used for monitoring
val_metrics:
_target_: nequip.train.EnergyForceMetrics
coeffs:
total_energy_rmse: 1.0
forces_rmse: 1.0
total_energy_mae: null # logged but not in weighted_sum
forces_mae: null # logged but not in weighted_sum
trainer:
_target_: lightning.Trainer
callbacks:
# Early stopping using the monitored metric
- _target_: lightning.pytorch.callbacks.EarlyStopping
monitor: ${monitored_metric}
patience: 20
min_delta: 1e-3
# Model checkpointing using the monitored metric
- _target_: lightning.pytorch.callbacks.ModelCheckpoint
monitor: ${monitored_metric}
filename: best
# Learning rate scheduler using the monitored metric
lr_scheduler:
scheduler:
_target_: torch.optim.lr_scheduler.ReduceLROnPlateau
factor: 0.6
patience: 5
monitor: ${monitored_metric}
Per-Species Force Loss Coefficients¶
For systems with very heterogeneous species, it can help to emphasize the force errors on some atomic types over others. EnergyForceLoss and EnergyForceStressLoss accept an optional per_type_forces_coeffs dict for this purpose. These are loss-aggregation coefficients — not to be confused with model parameters.
When supplied, the forces loss is computed per atom type and combined as a weighted mean sum(c_i * mse_i) / sum(c_i) instead of the default equal-mean over types. Equal coefficients reproduce the default behavior exactly.
loss:
_target_: nequip.train.EnergyForceLoss
per_atom_energy: true
coeffs:
total_energy: 1.0
forces: 1.0
# Emphasize H force errors relative to heavier species
per_type_forces_coeffs:
H: 5.0
O: 1.0
P: 1.0
Cs: 0.01
The dict must contain a strictly positive coefficient for every type in type_names (no missing keys, no zeros, no negatives). To deemphasize a species, give it a small positive coefficient rather than 0, since a force field with zero training signal on a species is rarely intended. Only the forces term is affected; per-structure terms (total_energy, stress) ignore this argument.
Per-type breakdowns logged during training (e.g. forces_mse_H, forces_mse_O) are the raw per-species MSEs. The coefficients are applied only to the aggregated forces_mse.
Logging per-species force MAE and RMSE¶
The loss only emits per-species MSE breakdowns (forces_mse_H, forces_mse_O, etc.). To also monitor per-species MAE and RMSE in physical units, the metrics wrapper accepts an extra_metrics list:
val_metrics:
_target_: nequip.train.EnergyForceMetrics
coeffs:
per_atom_energy_mae: 1.0
forces_mae: 1.0
# Per-species force breakdown (observation-only; omit `coeff` so they
# don't enter `weighted_sum`).
extra_metrics:
- name: per_type_forces_mae
field: forces
metric:
_target_: nequip.train.MeanAbsoluteError
per_type: true
- name: per_type_forces_rmse
field: forces
metric:
_target_: nequip.train.RootMeanSquaredError
per_type: true
# Reuse for train / test so all three log the same breakdown:
train_metrics: ${training_module.val_metrics}
test_metrics: ${training_module.val_metrics}
This logs:
per-species values:
per_type_forces_mae_H,per_type_forces_mae_O,per_type_forces_mae_P,per_type_forces_mae_Cs, and analogously for_rmse,the equal-mean aggregate over species:
per_type_forces_maeandper_type_forces_rmse.
To include a per-species metric in the monitored weighted_sum (e.g. to early-stop on per-species RMSE), give the entry a non-null coeff. The snippet above omits coeff, leaving these as observation-only.
Advanced Usage: Custom MetricsManager¶
For scenarios not covered by the simplified wrappers, you can use the full MetricsManager directly. Technical details and advanced examples are provided in the nequip.train.MetricsManager API documentation.
Common advanced use cases include:
Custom field modifiers beyond
PerAtomModifierPer-type metrics (separate metrics for each atom type, optionally combined as a weighted mean via
per_type_coeffs)Custom metric types (e.g.,
HuberLoss,StratifiedHuberForceLoss)Handling datasets with missing labels (using
ignore_nan: true)