[Draft] Make processing service support model retraining

[mohamedelabbas1996]

Summary

Antenna is gaining the ability to retrain a classifier head from species that people have verified. Only some models are worth retraining — a classifier head over a frozen backbone is cheap, a detector or a model trained end to end is not — so Antenna needs to be told which is which before it can offer retraining for a pipeline.

This adds that one flag. A model declares trainable, the worker reports it in the algorithm config it registers with Antenna, and Antenna uses it to decide whether retraining applies. Nothing else changes: no model is trainable unless it says so, and nothing here trains anything.

List of Changes

#	What it does	How
1	A model can declare that its head can be retrained	trainable attribute on InferenceBaseClass, defaulting to False
2	Antenna is told which algorithms can be retrained	trainable field on AlgorithmConfigResponse
3	Both places that build an algorithm response pass the flag through	make_algorithm_response() and make_algorithm_config_response()

Detailed Description

The flag is per algorithm rather than per service. A processing service usually hosts several pipelines and only one of them has a retrainable head, so a single service-level flag would not tell Antenna which head to touch.

The default is False everywhere, and the builders read it with getattr(model, "trainable", False), so a model written before this change is treated as not trainable rather than being assumed retrainable.

This matches the field Antenna reads from /info, which it mirrors onto its own Algorithm records. The corresponding Antenna change is in RolnickLab/antenna#1407.

Verification

trapdata/tests/test_trainable_flag.py covers the default, the response default, a model that declares the flag, and a model that predates the attribute.

Note: these tests have not been run locally — the working tree had no Python environment and building one pulls in torch. The schema field itself was checked in isolation with pydantic. CI runs the rest.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}