FlexMeasures / flexmeasures / 39306731112

87%

Build Type

push

github

Committed by web-flow

Commit Message

Hand the trained model to the prediction in memory, keeping a file only when asked (#2701)

* forecasting: hand the trained model to the prediction in memory

Context:
- #2111 and #2683: every forecast cycle pickled its trained model to a
  folder, only for the prediction to read it back in the same job a moment
  later, and nothing in production deleted it. A model can be hundreds of MB
  (589 MB for 12 files in one development checkout), the default folder was
  relative to the worker's working directory, and two runs for the same
  sensor shared file names. #2479's parallel training uses threads, so no
  model crosses a process boundary through the file.

Change:
- TrainPipeline.run returns the trained model, and PredictPipeline accepts
  it, reading a model file only when none is handed over.
- model-save-dir defaults to None: a model is kept as a file only when it is
  given, as an explicit request.
- Deleting a kept model after prediction still works, and is skipped when
  no file was kept.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: F.N. Claessen <claessen@seita.nl>

* forecasting/tests: pin a forecast keeping its model only when asked

Context:
- #2111.

Change:
- A forecast run from an empty working directory makes its forecasts and
  writes no model file anywhere; given a model-save-dir, it keeps one model
  per cycle there, which can be read back. On main, the first case fails:
  two model files appear under the working directory.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: F.N. Claessen <claessen@seita.nl>

* changelog: forecasts keep no model file unless asked

Context:
- #2111, #2683.

Change:
- Infrastructure entry for #2701, telling hosts the old artifacts folder can
  be deleted, and a CLI changelog entry for --model-save-dir.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: F.N. Claessen <claessen@seita.nl>

* forecasting: type the model handed from traini... (continued)

Coverage Stats

13 of 15 new or added lines in 4 files covered. (86.67%)

6 existing lines in 2 files now uncovered.

22105 of 25472 relevant lines covered (86.78%)

0.87 hits per line