[Draft] Serve the BioCLIP 2.5 + LogReg classifier as a pipeline

[mohamedelabbas1996]

Summary

This adds the BioCLIP 2.5 + logistic-regression species classifiers to the processing service as pipelines. A frozen BioCLIP 2.5 ViT-H/14 backbone carries a linear head fit as an sklearn LogisticRegression over L2-normalised embeddings, so a softmax over the head reproduces sklearn's predict_proba exactly. Two heads ship: one over the Newfoundland species list and one over the Panama list.

Only serving is included here. Retraining a head from verified labels builds on top of this and is kept for a separate change (RolnickLab/antenna#1407 and #167), so this PR is small and reviewable on its own.

List of Changes

#	What it does	How
1	Two BioCLIP 2.5 + LogReg classifiers	BioCLIP25NewfoundlandClassifier and BioCLIP25PanamaClassifier in trapdata/ml/models/bioclip.py, loading the backbone and a head from the Hub
2	They are offered as pipelines	bioclip_2_5_newfoundland and bioclip_2_5_panama in CLASSIFIER_CHOICES, with MothClassifierBioCLIP25* API wrappers
3	A classification carries the embedding it was made from	features on ClassificationResponse and ClassifierResult; the forward pass returns (logits, features)
4	open-clip-torch is available	added to pyproject.toml and uv.lock

Detailed Description

The backbone is frozen, so a crop's embedding never changes. The forward pass returns that embedding alongside the logits, and post_process_batch carries it through to ClassificationResponse.features. Storing the vector means the backbone never has to run over the same crop twice, which is what makes a later head-retrain cheap; a model that returns logits alone is unaffected and its features stays None.

The Newfoundland head is read from a public Hugging Face repository; the Panama head is read from a directory the deployment provides via BIOCLIP_PANAMA_HEAD_DIR. The label vocabulary for the Panama head travels inside the npz, so it is normalised through labels_from / BioCLIPVocabularyInHeadClassifier.

scripts/run_bioclip_on_remote_gpu.sh is the helper used to serve these on a remote GPU.

Verification

trapdata/tests/test_bioclip_classifier.py (6 tests) covers: both classifiers are registered under their slugs, they extend APIMothClassifier and the BioCLIP bases, they declare a backbone and a head, the forward pass returns an L2-normalised embedding matching the logits, post_process_batch keeps the embedding on the result, and ClassificationResponse carries features. These check the wiring without downloading the backbone or a head.

Run locally against this branch: the new tests and test_registry.py pass; black, isort, and flake8 are clean on the changed files.

[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.}