mnmly/rfdetr-keypoint-preview-mlx-fp32

RF-DETR Keypoint (preview) — MLX, fp32

MLX-format conversion of Roboflow's RF-DETR keypoint-preview (GroupPose) model, for use with mlx-swift-rf-detr on Apple Silicon.

• Upstream: roboflow/rf-detr (Apache-2.0), checkpoint rf-detr-keypoint-preview-xlarge.pth
• Task: person keypoint detection — 1 class (person), 17 COCO keypoints (GroupPose)
• Resolution: 576 · decoder layers: 4 · queries: 100 · backbone: dinov2_windowed_small (hidden 256)
• Precision: fp32
• Files: config.json + preprocessor_config.json + model.safetensors
• Status: preview — tracks the upstream keypoint-preview line.

Conversion

Converted from the upstream PyTorch checkpoint with Scripts/convert_keypoint.py. Built person-only (num_classes=1); four dead keypoint_head.keypoint_proj.* keys are dropped. Weights keep their model. prefix and NCHW conv layout (the Swift loader transposes to NHWC at load).

Usage (Swift)

import MLXRFDETR

let dir = URL(fileURLWithPath: "rfdetr-keypoint-preview-mlx-fp32")   // this repo's downloaded files
let predictor = try MLXRFDETR.fromPretrained(dir, scoreThreshold: 0.3, nmsThreshold: 0.5)
let result = try predictor.predict(url: URL(fileURLWithPath: "image.jpg"))
// result carries per-instance keypoints for the GroupPose model.

License

Apache 2.0 — see LICENSE. Model architecture and weights © Roboflow, Inc. This repository redistributes a format-converted copy of the upstream Apache-2.0 weights; no weights were retrained or modified beyond the PyTorch→MLX serialization.