feat: support rfdetr-keypoint-preview custom weights upload

[mkaic]

Note

Code and PR description are LLM-written. Code has been reviewed and tested in staging by Kai

What

Enables uploading RF-DETR Keypoint Preview (rfdetr-keypoint-preview) models to the Roboflow platform through the existing custom-weights flows: version.deploy(), workspace.deploy_model(), and the roboflow upload_model CLI.

The platform side already supports this model type: the model registry (packages/shared/models/modelRegistry/rfdetrModels.js in roboflow) registers rfdetr-keypoint-preview as a keypoint-detection model, and the export job (train/export/rfdetr/export.py in roboflow-train, PR in flight) converts uploaded keypoint checkpoints — deriving keypoints_metadata.json from the checkpoint's args.num_keypoints_per_class, so the existing bundle format (weights.pt + class_names.txt) needs no changes. This PR removes the client-side gate that rejected the model type.

Changes (all in roboflow/util/model_processor.py)

• Register the type: rfdetr-keypoint-preview → RFDETRKeypointPreview in _RFDETR_MODEL_TYPE_TO_CLASS. The class ships in public rfdetr>=1.8.0, which is already the required minimum (RFDETR_MIN_VERSION) for the raw PyTorch-Lightning checkpoint path, so no dependency bump is needed.
• Task mapping: task_of_model_type now maps model types containing keypoint to TASK_POSE (the type string spells the task out instead of using the -pose suffix token). This makes project-type validation accept keypoint-detection projects and reject others.
• Checkpoint task detection: keypoint training checkpoints from rf-detr-internal carry args but no model_name, and their config sets segmentation_head=False alongside keypoint_head=True — _detect_rfdetr_task previously misdetected them as detection. It now checks args.keypoint_head first. Checkpoints with model_name (RFDETRKeypointPreview) were already classified as pose; they are now accepted rather than rejected.
• Variant grid check: added the keypoint-preview position-encoding grid (576 // 12 = 48, mirroring RFDETRKeypointPreviewConfig in rf-detr-internal/public rfdetr) to RFDETR_POSITIONAL_ENCODING_SIZE.

Cross-task mislabels stay rejected in both directions: a keypoint checkpoint uploaded as rfdetr-base and a detection checkpoint uploaded as rfdetr-keypoint-preview both raise TaskMismatchError before anything is uploaded.

Testing

• New unit tests: task mapping, keypoint_head args fallback (dict + namespace), project-type validation accept/reject, PE-grid match, full packaging flow, and both cross-task rejection directions.
• Full suite: 871 passed, 1 skipped locally.
• Manually exercised packaging with a real torch.saved keypoint-style checkpoint (args including keypoint_head, num_keypoints_per_class, class_names): produces roboflow_deploy.zip with weights.pt + background-prefixed class_names.txt, no warnings; uploading the same checkpoint as rfdetr-nano is rejected with an actionable TaskMismatchError.

How to test an upload

roboflow upload_model -p <keypoint-project> -v <version> -t rfdetr-keypoint-preview -m <dir-with-checkpoint> -f checkpoint_best_total.pth

🤖 Generated with Claude Code

[mkaic]

Note

This comment was NOT written by Kai. It was written by Kai's Claude and posted through Kai's GitHub account on this PR.

Closing as a duplicate of #496, which predates this PR and uses a more robust keypoint-checkpoint detection signal: args.num_keypoints_per_class works for checkpoints from both the public rfdetr package and rf-detr-internal, whereas this PR's args.keypoint_head only exists in rf-detr-internal's config — a keypoint checkpoint trained with the public package would have been misdetected as detection and rejected.