xnn.common.benchmark.metrics.prediction_pairs

xnn.common.benchmark.metrics.prediction_pairs(out, data, targets, atomic_energies=None, energy_per_atom=True)

The scored (prediction, reference) pairs of one batch, as flat detached tensors.

The per-batch step of collect_predictions() (see there for the energy handling and which labels are scored); the trainer accumulates its error metrics from the same pairs.

Parameters:

• out (dict of str to Tensor) – The model output for data.

• data (AtomicGraph) – The batch with its labels.

• targets (list of str) – Quantities to pair, as in collect_predictions().

• atomic_energies (torch.Tensor or None, optional) – Per-element reference energies to subtract, as in collect_predictions().

• energy_per_atom (bool, optional) – Divide energies by the atom count. Defaults to True.

Returns:

The targets present in the batch and in out, on the batch’s device.

Return type:

dict of str to (Tensor, Tensor)