mayaku 1.0.0

Backend-portable detection / segmentation / keypoint library inspired by Detectron2.

Mayaku

Detectron2's R-CNN family, ported to where it actually deploys. A clean reimplementation of Faster R-CNN, Mask R-CNN, and Keypoint R-CNN that runs on Apple Silicon, exports cleanly to ONNX / CoreML / OpenVINO / TensorRT, and reproduces all 12 of Detectron2's model-zoo checkpoints to within ±0.1 AP. Pure Python — no custom CUDA kernels, no compiled extensions, no wheel chase. Apache 2.0.

The 12 D2 model-zoo configs ship pre-converted — mayaku predict --weights faster_rcnn_R_50_FPN_3x image.jpg runs on Apple Silicon the first time you call it. Same architectures, same numbers (±0.1 AP), with the deployment paths Detectron2 never had: .mlpackage, .onnx, .xml, .engine.

Why migrate from Detectron2?

	Detectron2	Mayaku
Apple Silicon (MPS) train + eval	not supported	first-class
pip install -e . on macOS	painful (custom CUDA kernels, ABI mismatches)	works out of the box
ONNX export	community-maintained, brittle, broken for keypoint	parity-tested (atol=1e-3, opset 17)
CoreML export	not first-class	shipped CLI verb, fp16/fp32, mlprogram
OpenVINO export	not first-class	shipped CLI verb
TensorRT export	not first-class	shipped CLI verb
Configuration	yacs / LazyConfig	Pydantic v2 (frozen, fail-at-load)
Last release	v0.6 (Nov 2021)	active

You don't have to retrain or convert anything. All 12 model-zoo checkpoints reproduce D2's published COCO val2017 numbers within ±0.1 AP and ship pre-converted and hosted — mayaku predict --weights <name> fetches on first use. Switch the runtime, keep the weights.

Small things that matter

• Channel order is RGB, not BGR. Detectron2 inherits Caffe2's BGR convention; Mayaku is RGB-native end-to-end (ADR 002). Feed a cv2.imread-shaped BGR array and Mayaku will run without complaint and produce wrong detections, no error raised. Either swap channels at the boundary (img = img[:, :, ::-1] or mayaku.utils.bgr_to_rgb) or — better — load with mayaku.utils.image.read_image (Pillow under the hood, RGB by default).

• Pixel mean / std are RGB-ordered. The defaults are [123.675, 116.280, 103.530] / [58.395, 57.120, 57.375] (matching torchvision's ImageNet stats in RGB). If you copy a D2 yacs config that pins PIXEL_MEAN = [103.53, 116.28, 123.675] (BGR), the model normalises with the channels swapped and you again silently get bad features. Don't override the defaults unless your dataset truly needs different stats — and if you do, write them in RGB order.

• Fine-tuned .pkl files need a one-time conversion. The 12 zoo checkpoints are already converted and hosted — you don't need to do anything for the standard configs. If you have your own fine-tuned .pkl from a D2 training run, D2's caffe2-flavoured key names (conv.norm.*, shortcut.*, fpn_lateral{N}) won't load_state_dict into Mayaku's torchvision-flavoured layout. Run python tools/convert_d2_checkpoint.py your_model.pkl -o your_model.pth once. See Bringing your own Detectron2 checkpoint below.

Detectron2 parity

All 12 checkpoints reproduce Detectron2's published numbers within ±0.1 AP. Maximum observed gap across the sweep: +0.08 AP.

Config	D2 published	Mayaku	Δ
faster_rcnn_R_50_FPN_3x	40.2	40.23	+0.03
faster_rcnn_R_101_FPN_3x	42.0	42.00	+0.00
faster_rcnn_X_101_32x8d_FPN_3x	43.0	43.07	+0.07
mask_rcnn_R_50_FPN_3x (box / mask)	41.0 / 37.2	40.98 / 37.17	-0.02 / -0.03
mask_rcnn_R_101_FPN_3x (box / mask)	42.9 / 38.6	42.93 / 38.63	+0.03 / +0.03
mask_rcnn_X_101_32x8d_FPN_3x (box / mask)	44.3 / 39.5	44.28 / 39.52	-0.02 / +0.02
keypoint_rcnn_R_50_FPN_3x (box / kpt)	55.4 / 65.5	55.45 / 65.49	+0.05 / -0.01
keypoint_rcnn_R_101_FPN_3x (box / kpt)	56.4 / 66.1	56.43 / 66.04	+0.03 / -0.06
keypoint_rcnn_X_101_32x8d_FPN_3x (box / kpt)	57.3 / 66.0	57.26 / 66.08	-0.04 / +0.08

Full per-checkpoint table including the 1x configs: docs/d2_parity_report.md. Reproduce the entire sweep with bash tools/convert_all_d2.sh (downloads each .pkl, converts, evaluates, regenerates the report).

These numbers come from loading and evaluating D2's converged weights in Mayaku — not from training from scratch. Training-from-scratch parity (270k iters on COCO, ending at D2's published numbers) is on the roadmap, not measured.

Pre-converted models

Pass a model name instead of a path; the CLI fetches it on first use:

mayaku predict configs/detection/faster_rcnn_R_50_FPN_3x.yaml image.jpg \
    --weights faster_rcnn_R_50_FPN_3x

Or pre-stage with mayaku download <name> (all variants) or mayaku download <name> --target <variant>. Cached under ~/.cache/mayaku/v1/models/, SHA256-verified.

Variants: pth (PyTorch) · onnx (dynamic) · onnx-fixed (TRT-friendly) · coreml-fp16 (Apple Silicon) · openvino (Intel CPU/iGPU/Arc/NPU).

Detection

Model	pth	onnx	onnx-fixed	coreml-fp16	openvino
faster_rcnn_R_50_FPN_1x	⬇	⬇	⬇	⬇	xml · bin
faster_rcnn_R_50_FPN_3x	⬇	⬇	⬇	⬇	xml · bin
faster_rcnn_R_101_FPN_3x	⬇	⬇	⬇	⬇	xml · bin
faster_rcnn_X_101_32x8d_FPN_3x	⬇	⬇	⬇	⬇	xml · bin

Segmentation

Model	pth	onnx	onnx-fixed	coreml-fp16	openvino
mask_rcnn_R_50_FPN_1x	⬇	⬇	⬇	⬇	xml · bin
mask_rcnn_R_50_FPN_3x	⬇	⬇	⬇	⬇	xml · bin
mask_rcnn_R_101_FPN_3x	⬇	⬇	⬇	⬇	xml · bin
mask_rcnn_X_101_32x8d_FPN_3x	⬇	⬇	⬇	⬇	xml · bin

Keypoints

Model	pth	onnx	onnx-fixed	coreml-fp16	openvino
keypoint_rcnn_R_50_FPN_1x	⬇	⬇	⬇	⬇	xml · bin
keypoint_rcnn_R_50_FPN_3x	⬇	⬇	⬇	⬇	xml · bin
keypoint_rcnn_R_101_FPN_3x	⬇	⬇	⬇	⬇	xml · bin
keypoint_rcnn_X_101_32x8d_FPN_3x	⬇	⬇	⬇	⬇	xml · bin

coreml-fp16 is a zipped .mlpackage directory — mayaku download unpacks it; if you fetch directly, unzip it yourself.

TensorRT engines aren't hosted — they're tied to a specific GPU architecture and TRT version, so a hosted file would fail to load on mismatched hardware. Build locally on your CUDA host from the onnx-fixed artifact:

mayaku download faster_rcnn_R_50_FPN_3x --target onnx-fixed
mayaku export tensorrt configs/detection/faster_rcnn_R_50_FPN_3x.yaml \
    --weights faster_rcnn_R_50_FPN_3x \
    --output ~/faster_rcnn_R_50_FPN_3x.engine \
    --sample-height 800 --sample-width 1344

Machine-readable index: manifest.json (size + SHA256 per file; dtmfiles also publishes a .sha256 sidecar next to every artifact).

What you get

Detectors	Faster R-CNN · Mask R-CNN · Keypoint R-CNN (56×56 heatmaps)
Backbones	ResNet-50, ResNet-101, ResNeXt-101 (32×8d) — all with FPN
Backends	CUDA · MPS · CPU (single codebase, single wheel)
Export targets	mayaku export {onnx, coreml, openvino, tensorrt} — all four parity-tested. TensorRT needs a CUDA host.
Configuration	Pydantic v2, frozen + extra-forbidden, validated at load
Distributed	DDP via torchrun (nccl on CUDA, gloo elsewhere)
License	Apache 2.0

See docs/architecture.md for module layout and docs/portability.md for the per-backend matrix.

Install

pip install mayaku

That's enough for inference, training, evaluation, and mayaku download. Export targets are optional extras — install only what you need:

pip install mayaku[onnx]        # ONNX export (adds the onnx validation package)
pip install mayaku[coreml]      # CoreML export (macOS only)
pip install mayaku[openvino]    # OpenVINO export (Intel targets)
pip install mayaku[tensorrt]    # TensorRT export (CUDA Linux only)

onnxruntime is not bundled in [onnx] because the CPU and CUDA wheels are separate packages — installing the wrong one is the most common source of libcudart.so / .dll errors. Install the variant that matches your host:

pip install onnxruntime          # CPU · Apple Silicon · any host without NVIDIA GPU
pip install onnxruntime-gpu      # CUDA Linux / Windows — must match your CUDA version

If either is missing when needed, the CLI reports the correct install command for your hardware. You can combine extras freely: pip install mayaku[onnx,coreml]. The tensorrt extra carries a PEP 508 marker that makes it a no-op on macOS, so pip install mayaku[onnx,coreml,openvino,tensorrt] is safe to run on any host.

Contributing:

git clone https://github.com/datamarkin/mayaku
pip install -e ".[dev,onnx]"

Bringing your own Detectron2 checkpoint

The 12 zoo checkpoints are already converted and hosted — you don't need to do anything for the standard configs. This section is only for the case where you have a fine-tuned D2 .pkl from your own training run and want to keep using those weights.

# Convert your fine-tuned D2 .pkl → Mayaku .pth (one-shot, no network)
python tools/convert_d2_checkpoint.py your_model_final.pkl -o your_model.pth

# Use the converted .pth like any other Mayaku checkpoint
mayaku predict configs/detection/faster_rcnn_R_50_FPN_3x.yaml image.jpg \
    --weights your_model.pth --device mps

Covers Faster / Mask / Keypoint R-CNN with R-50 / R-101 / X-101_32x8d FPN — the same architectures Mayaku ships. Head-specific rename rules are inert when the source .pkl doesn't contain them, so a Faster R-CNN checkpoint converts cleanly without flags. See tools/README.md for the full rename table and edge cases.

If your D2 setup uses anything Mayaku doesn't ship (DCN, Cascade, Panoptic, DETR, ViTDet…), the converter can't help — see Small things that matter for the missing pieces.

Quickstart — CLI

The mayaku console script bundles four subcommands. Every one takes a YAML config plus per-task arguments.

Train

mayaku train configs/detection/faster_rcnn_R_50_FPN_3x.yaml \
    --json /data/coco/annotations/instances_train2017.json \
    --images /data/coco/train2017 \
    --output runs/frcnn_r50 \
    --pretrained-backbone \   # ImageNet weights for the ResNet (recommended)
    --device cuda             # cpu / mps / cuda; default = auto

Pass --pretrained-backbone for fine-tuning. The schema's backbone.freeze_at=2 default freezes the ResNet stem + res2, which is only meaningful when those stages already carry useful features. For genuine from-scratch training, set model.backbone.freeze_at: 0 in the YAML and omit the flag.

--max-iter N overrides solver.max_iter for smoke runs; --log-period N controls per-iteration log frequency (default 20).

Evaluate

mayaku eval configs/detection/faster_rcnn_R_50_FPN_3x.yaml \
    --weights runs/frcnn_r50/model_final.pth \
    --json /data/coco/annotations/instances_val2017.json \
    --images /data/coco/val2017

Prints the COCO AP / AP50 / AP75 / APs / APm / APl dict for boxes, plus masks/keypoints when the meta-architecture asks for them.

Predict (single image)

mayaku predict configs/detection/faster_rcnn_R_50_FPN_3x.yaml path/to/image.jpg \
    --weights runs/frcnn_r50/model_final.pth \
    --output detections.json

Download

# List all available models and variants
mayaku download --list

# Download all variants of a model (pth + onnx + onnx-fixed + coreml-fp16 + openvino)
mayaku download faster_rcnn_R_50_FPN_3x

# Download a single variant
mayaku download faster_rcnn_R_50_FPN_3x --target coreml-fp16

# Pre-stage everything for an offline deployment machine
mayaku download --all

# Skip SHA256 verification (air-gapped or offline-cached environments)
mayaku download faster_rcnn_R_50_FPN_3x --no-verify

Artifacts are cached under ~/.cache/mayaku/v1/models/ (or $XDG_CACHE_HOME/mayaku/…). Passing a bare model name to --weights in eval and predict triggers the same fetch automatically on first use.

Export

# ONNX (required target, opset 17, dynamic batch + spatial axes by default)
mayaku export onnx configs/detection/faster_rcnn_R_50_FPN_3x.yaml \
    --weights runs/frcnn_r50/model_final.pth --output model.onnx

# Same surface for the other targets
mayaku export coreml ...    --output model.mlpackage
mayaku export openvino ...  --output model.xml      # .bin written alongside
mayaku export tensorrt ...  --output model.engine   # CUDA host required

Per-target details (what's in the exported graph, what stays Python-side, parity tolerances, runtime examples) live in docs/export/.

Quickstart — Python API

import torch
from mayaku.cli._factory import build_detector
from mayaku.config import load_yaml
from mayaku.inference import Predictor

cfg = load_yaml("configs/detection/faster_rcnn_R_50_FPN_3x.yaml")
model = build_detector(cfg).eval()
model.load_state_dict(
    torch.load("runs/frcnn_r50/model_final.pth", map_location="cpu",
               weights_only=True)
)

pred = Predictor.from_config(cfg, model)
instances = pred("path/to/image.jpg")
# -> mayaku.structures.Instances with .pred_boxes, .scores, .pred_classes
#    (and .pred_masks / .pred_keypoints when the architecture asks for them),
#    all in original-image pixel coordinates.

For a complete fine-tuning script (custom dataset discovery, COCO eval hook, Mac-friendly defaults), see examples/finetune.py. For training your own loop, see docs/architecture.md §"Engine" — mayaku.engine.SimpleTrainer / AMPTrainer are usable directly.

Configuration

Configs are pydantic v2 models (see mayaku.config.schemas); YAML is the on-disk encoding. Defaults match the Detectron2 3x schedule modulo the overrides recorded in the decision log (RGB channel order, no rotated boxes, no deformable conv, device="auto").

Minimal example:

model:
  meta_architecture: faster_rcnn
  backbone:
    name: resnet50
    norm: FrozenBN
  roi_heads:
    num_classes: 80
solver:
  max_iter: 90000
  steps: [60000, 80000]
  base_lr: 0.02
input:
  min_size_train: [640, 672, 704, 736, 768, 800]
  max_size_train: 1333

Configs are frozen and reject unknown fields — you'll get a validation error at load time, not a silent default mid-training.

Deployment & throughput

Mayaku ships CoreML, ONNX, OpenVINO, and TensorRT exports as deployment artifacts for non-PyTorch targets — iOS apps, macOS apps, Linux CPU servers, Windows ML stacks, edge devices, INT8 quantization workflows. The artifacts are the value, not raw throughput on a developer machine.

Empirical pattern, measured across three platforms in benchmarks/:

• GPU-available targets (CUDA, Apple Silicon MPS): PyTorch eager is essentially optimal for the R-CNN graph shape — multi-output FPN with stride-2 lateral connections defeats the deployment runtimes' fusion templates. Use the exports when you need the artifact format for a non-PyTorch target, not for a speed gain on a developer box.
• CPU-only Intel targets: OpenVINO genuinely beats PyTorch CPU by 2.65× on R-CNN R-50 FPN. The one row where the deployment-runtime claim is real, not just rhetorical. Useful for embedded servers, edge boxes, virtualised CPU instances.
• Backbone-only feature extraction (any platform): ~6× over framework eager on Mac/CUDA, ~3× on Intel CPU. Classifier-shaped graphs are exactly what these runtimes are designed for.

Full per-platform numbers and analysis in docs/vs_detectron2.md §"A note on export throughput" and ADRs 004 (CoreML), 005 (ONNX/TRT/OpenVINO).

What's deliberately not shipped

• Deformable convolution (ADR 001) — the portability cost wasn't worth the marginal AP. D2's strongest DCN-cascade configs sit ~2–4 AP above what Mayaku can reach; we don't try to match them.
• BGR channel-order configurability (ADR 002) — RGB-native via PIL throughout.
• Out-of-scope architectures: rotated boxes, panoptic FPN, Cascade R-CNN, RetinaNet, DETR, ViTDet, PointRend, DensePose, and the rest of D2's projects/. Mayaku ships exactly three meta-architectures. For a researcher comparing detector families, D2 is still the right tool.
• cv2-based image pipeline — PIL only. Closes a known ~1–2 AP recipe gap on JPEG inputs against D2; a cv2 path is on the roadmap.

For the full honest comparison (including where D2 is still better), see docs/vs_detectron2.md.

Roadmap

• DINOv2 backbones — ViT ladder (S / B / L / g) for stronger pretrained init, replacing the current ResNet/ResNeXt-only backbone surface. Not yet shipped.
• Training-from-scratch parity validation — current parity numbers come from loading + evaluating D2's converged weights, not training to them.
• Opt-in cv2 image pipeline — closes the recipe-level AP gap for users who want bit-tighter D2 parity.
• CI — ruff + mypy + pytest matrix across CPU / MPS / CUDA on every push. Currently testing is manual.
• Expanded examples/ — minimal scripts for inference, fine-tuning, and per-target export.

Documentation

• docs/architecture.md — module layout end-to-end.
• docs/portability.md — backend matrix, op fallbacks, AMP rules, MPS quirks, distributed launchers.
• docs/extending.md — adding a backbone / head / dataset / augmentation / exporter.
• docs/vs_detectron2.md — capability matrix, parity, non-goals, deployment-throughput analysis.
• docs/export/ — per-target export recipes (ONNX, CoreML, OpenVINO, TensorRT).
• docs/decisions/ — architecture decision records.

Development

ruff check src tests
mypy
MAYAKU_DEVICE=cpu pytest          # also: mps, cuda

MAYAKU_DEVICE selects which backend the test suite runs on. Tests marked cuda / mps / multi_gpu / tensorrt auto-skip when the active backend or the optional dependency isn't available. There is no CI yet — testing is manual, cross-machine.

License

Apache 2.0 — see LICENSE.