Validate task type before model upload

[leeclemnet]

Summary

Uploading a segmentation .pt with model_type="yolov11" (or any other detect↔seg/pose/cls/obb mismatch) previously succeeded silently and crashed at first inference with ValueError: not enough values to unpack (expected 2, got 1). Same failure mode for rfdetr-* vs rfdetr-seg-*. The .pt always knew its task — nothing in the SDK or the server conversion was reading it.

This change adds client-side task validation on two axes:

1. .pt ↔ model_type reconciliation inside model_processor.process():

   • Ultralytics: detects task via type(model_instance).__name__ (SegmentationModel / PoseModel / ClassificationModel / OBBModel / DetectionModel) — works across v8/v10/v11/v12/yolo26.
   • RF-DETR: detects via args.segmentation_head today, with an args.task string path already wired in for future rfdetr-pose/obb/cls variants.
   • YOLO: auto-appends the correct suffix when the user omitted it (so model_type="yolov11" + seg .pt → "yolov11-seg"). Raises ValueError on conflict.
   • RF-DETR: raises on conflict (no auto-append, model_type needs explicit size).

2. model_type ↔ project type cross-check inside Version.deploy(). Catches uploads like rfdetr-seg-medium to an object-detection project. Not added to Workspace.deploy_model() since that path takes a project_ids list and would need extra fetches.

A single substring scanner (_task_from_substring) drives the class-name check, the rfdetr args.task path, and the deploy model_type parse — keeping the logic future-proof (rfdetr-pose-medium, rfdetr-obb-large, etc. resolve correctly with zero SDK change when they ship).

Also adds TYPE_CLASSIFICATION to roboflow/config.py; the typo'd TYPE_CLASSICATION is retained as a backwards-compat alias.

Changes

• roboflow/util/model_processor.py — new helpers (_task_from_substring, task_of_model_type, _detect_task_from_pt); _process_yolo and _process_rfdetr reconcile model_type against the loaded .pt; process() now returns (zip_file_name, model_type) so callers can use the corrected value.
• roboflow/core/version.py — unpack tuple, call new _validate_against_project_type.
• roboflow/core/workspace.py — unpack tuple.
• roboflow/config.py — add TYPE_CLASSIFICATION.
• tests/util/test_model_processor.py (new) — 15 cases covering task_of_model_type and _detect_task_from_pt including future-proofing for rfdetr-pose/obb.
• tests/test_version.py — 11 cases covering _validate_against_project_type.

Test plan

• python -m unittest — 446/446 pass
• ruff check roboflow — clean
• ruff format roboflow — applied
• Manual: deploy a YOLOv11l-seg .pt with model_type="yolov11" on a segmentation project → model_artifacts.json contains "model_type": "yolov11-seg", upload URL uses modelType=yolov11-seg, inference returns masks
• Manual: deploy same seg .pt with model_type="yolov11-pose" → ValueError at client
• Manual: deploy model_type="yolov11-seg" on a detection project → ValueError at client
• Manual: regression — detection .pt + model_type="yolov11" on a detection project deploys cleanly
• Manual (rfdetr): rfdetr seg .pt + model_type="rfdetr-medium" → ValueError (reads args.segmentation_head=True)
• Manual (rfdetr): model_type="rfdetr-seg-medium" on a detection project → ValueError

🤖 Generated with Claude Code