Init dates are reduced when forecast steps exceed `end_date`

[MatKbauer]

What happened?

When running a three-months inference with the following command

uv run --offline inference --from-run-id izu35lvf --options test_config.samples_per_mini_epoch=368 test_config.output.num_samples=368 training_config.forecast.num_steps=120 test_config.start_date=2023-10-01T00:00 test_config.end_date=2023-12-31T00:00

the console logs the following information

0: TimeWindowHandler: start=2023-10-01T00:00:00.000000000, end=2023-12-31T00:00:00.000000000, len=06:00:00, step=06:00:00
0: forecast_steps : 120
0: There are only 242 available_samples, samples_per_mini_epoch reduced to 242 to avoid repeating data. Set repeat_data_in_mini_epoch to True if this is undesired.

That is, the init dates are adjusted such that the last rollout step still lies within the end_date.

However, with start_date and end_date, we explicitly want to specify the init times, independend of the forecast steps. The desired behavior, thus, is to have 368 init dates, when specifying the above start_date and end_date, and unrolling a forecast for forecast.num_steps into the future from each of these init dates.

What are the steps to reproduce the bug?

No response

Hedgedoc link to logs and more information. This ticket is public, do not attach files directly.

No response

[clessig]

The current approach does not properly check the available samples in the input dataset and too aggressively truncates based on the lead time. This only should happen if any/all datasets are not available--it's not entirely clear if any or all should apply here. All can be dangerous when we don't have clear visibility which samples have which input data; all can be too conservative.