synth_metric

Monitor_metric_name can be accessed by combining the stage (train | val) with the property name and the metric type (mae or rmse), for example: val/energy_mae, val/energy_per_atom_rmse, val/forces_mae, val/stress_rmse, val/loss, train/loss, etc.

Specifically, we have a composite metric, val/synth_metric, for which a mixing ratio can be specified. This is particularly useful when training on large datasets with outliers.

Note that the val/synth_metric is only supported for the validation, train/synth_metric is not supported. All weights other than the one corresponding to monitor_metric_name must also be for val. These weights are internally normalized.

Example

synth_metric:
    monitor_metric_name: val/synth_metric # use for variable interpolation
    # MAE
    val/energy_mae: 1.0
    val/energy_per_atom_mae: 5.0
    val/forces_mae: 5.0
    val/stress_mae: 2.5
    # RMSE
    val/energy_per_atom_rmse: 1.0
    val/forces_rmse: 1.0
    val/stress_rmse: 1.0