Skip to content

metadata_passthrough

metadata_passthrough

Passthrough loss that forwards target keys into metadata.

MetadataPassthrough 

MetadataPassthrough(keys_mapping)

Bases: BaseLoss

Passthrough to add target keys to metadata without computing loss.

Useful for adding auxiliary data to metadata for visualization/analysis without affecting training.

Initialize metadata passthrough.

Parameters:

Name Type Description Default
keys_mapping dict[str, str]

Mapping from target keys to metadata keys. Example: {"phase_label": "phase_label"} extracts targets["phase_label"] and stores it in metadata["phase_label"].

 required
Source code in src/versatil/metrics/losses/metadata_passthrough.py 
16
17
18
19
20
21
22
23
24
25
def __init__(self, keys_mapping: dict[str, str]):
    """Initialize metadata passthrough.

    Args:
        keys_mapping: Mapping from target keys to metadata keys.
            Example: {"phase_label": "phase_label"} extracts targets["phase_label"]
            and stores it in metadata["phase_label"].
    """
    super().__init__()
    self.keys_mapping = resolve_dict_keys(dict(keys_mapping))

get_required_keys 

get_required_keys()

Get required target keys.

Source code in src/versatil/metrics/losses/metadata_passthrough.py 
27
28
29
def get_required_keys(self) -> set[str]:
    """Get required target keys."""
    return set(self.keys_mapping.keys())

forward 

forward(predictions, targets, is_pad=None)

Extract keys from targets and add to metadata.

Source code in src/versatil/metrics/losses/metadata_passthrough.py 
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
def forward(
    self,
    predictions: dict[str, torch.Tensor],
    targets: dict[str, torch.Tensor],
    is_pad: torch.Tensor | None = None,
) -> LossOutput:
    """Extract keys from targets and add to metadata."""
    device = next(iter(predictions.values())).device
    metadata = {}
    for target_key, metadata_key in self.keys_mapping.items():
        if target_key in targets:
            metadata[metadata_key] = targets[target_key].detach()
    return LossOutput(
        total_loss=torch.tensor(0.0, device=device),
        component_losses={},
        metadata=metadata,
    )