Skip to content

Task: Support Optional Target During Inference for Beyond-Training-Date Forecasting #2256

@sbAsma

Description

@sbAsma

Describe the task. It can be a feature, documentation, etc.

Currently, the model requires a target tensor during inference, which limits its use to periods where ground truth data is available. This issue proposes making the target optional to enable inference on future timescales where no reference data exists (e.g., operational forecasting or out-of-distribution scenario simulation).

This requires:

  • Decoupling the target from the inference pipeline so the forward pass can run without it
  • Gracefully skipping loss computation and validation metrics when no target is provided
  • Ensuring compatibility with existing training/evaluation workflows (i.e., no regression when target is present)

Hedgedoc URL, if you are keeping notes, plots, logs in hedgedoc.

No response

Area

  • datasets, data readers, data preparation and transfer
  • model
  • science
  • infrastructure and engineering
  • evaluation, export and visualization
  • documentation
  • performance

Metadata

Metadata

Assignees

Labels

infraIssues related to infrastructure

Type

No type

Projects

Status

No status

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions