oriented-det 0.1.0__tar.gz → 0.2.0__tar.gz

{oriented_det-0.1.0 → oriented_det-0.2.0}/PKG-INFO

@@ -1,12 +1,12 @@
 Metadata-Version: 2.4
 Name: oriented-det
-Version: 0.1.0
+Version: 0.2.0
 Summary: A lightweight, modern PyTorch library for rotated object detection in aerial and satellite imagery
 Author: Jeff Faudi, DL4EO
 Maintainer: Jeff Faudi, DL4EO
 License-Expression: Apache-2.0
 Project-URL: Homepage, https://github.com/DL4EO/oriented-det
-Project-URL: Documentation, https://github.com/DL4EO/oriented-det#readme
+Project-URL: Documentation, https://dl4eo.github.io/oriented-det/
 Project-URL: Repository, https://github.com/DL4EO/oriented-det
 Project-URL: Issues, https://github.com/DL4EO/oriented-det/issues
 Project-URL: Bug Tracker, https://github.com/DL4EO/oriented-det/issues
@@ -25,6 +25,7 @@ Classifier: Topic :: Scientific/Engineering :: Image Recognition
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
+Requires-Dist: setuptools<82,>=65.0
 Requires-Dist: torch>=2.3.0
 Requires-Dist: torchvision>=0.18.0
 Requires-Dist: numpy<2
@@ -35,9 +36,11 @@ Requires-Dist: tensorboard>=2.10
 Requires-Dist: matplotlib>=3.5.0
 Requires-Dist: huggingface_hub>=0.20.0
 Requires-Dist: shapely>=2.0.0
+Provides-Extra: viewer
+Requires-Dist: gradio==6.8.0; extra == "viewer"
 Provides-Extra: export
 Requires-Dist: onnx>=1.14.0; extra == "export"
-Requires-Dist: onnxruntime>=1.16.0; extra == "export"
+Requires-Dist: onnxruntime-gpu[cuda,cudnn]>=1.16.0; extra == "export"
 Requires-Dist: onnx2tf>=1.22.0; extra == "export"
 Requires-Dist: tensorflow>=2.14.0; extra == "export"
 Provides-Extra: dev
@@ -65,7 +68,7 @@ Dynamic: license-file
 - **Geometry**: Rotated bounding boxes (rbox: cx, cy, w, h, angle), quadrilateral boxes (qbox), polygon ↔ rbox ↔ hbox conversions, angle normalization (le90, 0–180°), flip/rotate/scale transforms, visualization helpers
 - **IoU & NMS**: Rotated IoU and oriented NMS (CPU with optional GPU kernels when available); AABB pre-filtering; `obb_to_xyxy` / HBB conversion
 - **Datasets**: DOTA polygon loader (pattern, split file, or separate folders), image tiling, label filtering, ignore masks, oriented mAP evaluation
-- **Models**: **Oriented R-CNN** (horizontal RPN + MidpointOffset → oriented ROI), **Rotated Faster R-CNN** (oriented RPN + oriented ROI), **Rotated RetinaNet** (oriented anchors, focal loss); ResNet + FPN backbones; selective loading of external checkpoints where configs wire `checkpoint.load_from_checkpoint`
+- **Models**: **Oriented R-CNN** ([Xie et al., ICCV 2021](https://openaccess.thecvf.com/content/ICCV2021/html/Xie_Oriented_R-CNN_for_Object_Detection_ICCV_2021_paper.html); horizontal RPN + MidpointOffset → oriented RoIAlign + oriented ROI head), **Rotated Faster R-CNN** (Ren et al., NeurIPS 2015 two-stage baseline with horizontal RPN + horizontal RoIAlign + rotated ROI head; MMRotate reference), **Rotated RetinaNet** ([Lin et al., ICCV 2017](https://openaccess.thecvf.com/content_ICCV_2017/papers/Lin_Focal_Loss_for_ICCV_2017_paper.pdf); oriented anchors, sigmoid focal loss), **Rotated FCOS** (anchor-free single-stage; distance-angle coder, centerness, L1 / KFIoU / decoded rIoU); ResNet + FPN backbones; selective loading of external checkpoints where configs wire `checkpoint.load_from_checkpoint`
 - **Training**: JSON configs + **`odet train`**, mixed precision (AMP), gradient accumulation, checkpointing, best-metric tracking, TensorBoard, optional curriculum learning and profiling
 
 ## Installation
@@ -90,8 +93,10 @@ uv pip install -r requirements.txt
 uv pip install -e .
 ```
 
-- From PyPI (when published): `pip install oriented-det`
+- To auto-activate `.venv` when entering this repo, install the [`direnv`](https://direnv.net/) shell hook and run `direnv allow` from the repo root. The local `.envrc` is intentionally gitignored so each developer can opt in on their machine.
+- From PyPI: `pip install oriented-det`
 - For development and tests: `uv pip install -e ".[dev]"`
+- For the Gradio prediction viewer: `uv pip install -e ".[viewer]"` or `pip install "oriented-det[viewer]"`
 - For **macOS Apple Silicon** or **CPU-only**, see [Installation](docs/getting-started/installation.md).
 - Verify: `pytest tests/test_geometry.py tests/test_iou.py tests/test_nms.py`
 
@@ -101,13 +106,13 @@ After [installation](#installation):
 
 ```bash
 # Train on DOTA tiles (edit dataset paths in the config first)
-odet train --config configs/rotated_faster_rcnn/dota_le90_3x.json
+odet train --config configs/oriented_rcnn/dota_le90_1x.json
 
 # Or use the Makefile wrapper (same default config)
 make train
 ```
 
-Default 3× recipe: `configs/rotated_faster_rcnn/dota_le90_3x.json` (see [configs/rotated_faster_rcnn/README.md](configs/rotated_faster_rcnn/README.md)). 1× baseline: `dota_le90_1x.json`.
+Default starter recipe: `configs/oriented_rcnn/dota_le90_1x.json` (see [configs/oriented_rcnn/README.md](configs/oriented_rcnn/README.md)). Use `configs/oriented_rcnn/dota_le90_3x.json` when you want the longer Oriented R-CNN schedule.
 
 Programmatic APIs and a longer walkthrough: [Getting Started](docs/getting-started/installation.md). Config fields: [Configuration](docs/user-guide/configuration.md).
 
@@ -150,7 +155,7 @@ After that, copy-pasted commands and unmodified configs that reference `/path/to
 Full documentation is in the **docs/** folder and can be built and served with MkDocs:
 
 - **Build/serve**: `make docs` or `make docs-serve` (see [docs/README.md](docs/README.md)); or `uv pip install -e ".[docs]"` then `mkdocs serve`.
-- **Guides**: [Getting Started](docs/getting-started/installation.md), [User Guide](docs/user-guide/geometry.md), [API Reference](docs/api/geometry.md), [Examples](docs/examples/visualization.md).
+- **Guides**: [Getting Started](docs/getting-started/installation.md), [User Guide](docs/user-guide/geometry.md), [API Reference](docs/api/geometry.md), [Examples](docs/examples/inference.md).
 
 ## Documentation by folder
 
@@ -171,18 +176,18 @@ Install once: `uv pip install -e .`. Then:
 
 | Task | Command |
 |------|---------|
-| Train | `odet train --config configs/rotated_faster_rcnn/dota_le90_3x.json` or `make train` |
+| Train | `odet train --config configs/oriented_rcnn/dota_le90_1x.json` or `make train` |
 | Multi-GPU | `make train-multi-gpu` (`torchrun` + cuDNN on `LD_LIBRARY_PATH`) |
 | Tile DOTA | `odet tile-dota /path/to/dota/train` |
-| Val predictions | `odet preds --experiment-dir runs/rotated_faster_rcnn/<timestamp>` or `make preds` |
+| Val predictions | `odet preds --experiment-dir runs/oriented_rcnn/<timestamp>` or `make preds` |
 | Offline mAP | `make eval-val` or `make preds` then `make metrics` |
 
 DOTA configs: per-model `dota_le90_1x.json` / `dota_le90_3x.json` under [configs/](configs/). Run `odet --help` for all subcommands. Makefile shortcuts and script-level options: [tools/README.md](tools/README.md). Config reference: [docs/user-guide/configuration.md](docs/user-guide/configuration.md), [configs/config.schema.json](configs/config.schema.json), [configs/README.md](configs/README.md).
 
 ## Pretrained weights and evaluation
 
-- Place exported best checkpoints under **`pretrained/`** or use Hub slugs (`odet pretrained download oriented_rcnn_dota_le90_1x`). See [pretrained/README.md](pretrained/README.md) and [configs/README.md](configs/README.md#dota-pretrained-models-model-zoo).
-- **Tiled validation:** after training, run `make preds` then `make metrics` (or `odet preds` / `odet preds --metrics-from-json`). Large images use sliding-window inference when they exceed the model canvas; thresholds and overlap come from **`production.*`** in the experiment config.
+- Place exported best checkpoints under **`pretrained/`** or use Hub slugs (`odet pretrained download oriented_rcnn_dota_le90_3x` or `oriented_rcnn_dota_le90_1x`). See [pretrained/README.md](pretrained/README.md) and [configs/README.md](configs/README.md#dota-pretrained-models-model-zoo).
+- **Tiled validation:** after training, run `make preds` then `make metrics`. Published mAP reports: [`docs/eval-reports/`](docs/eval-reports/) (git). Raw detections for the viewer: gitignored [`predictions/`](predictions/).
 
 ## Important notes
 
@@ -192,9 +197,14 @@ DOTA configs: per-model `dota_le90_1x.json` / `dota_le90_3x.json` under [configs
 
 ## Roadmap
 
-- **v0.1** (current): Geometry, IoU/NMS, DOTA loader + tiler, three baseline detectors, config-based training, Hub pretrained weights
-- **Next**: HRSC2016 dataset support, expanded tutorials, hosted docs
-- **Future**: Additional oriented detectors (e.g. FAIR1M), optional fused CUDA kernels
+See **[docs/roadmap.md](docs/roadmap.md)** for the full public plan. Summary:
+
+- **v0.1** (shipped): Geometry, IoU/NMS, DOTA, three ResNet-FPN detectors, config training, Hub pretrained weights
+- **v0.1.1** (shipped): ProbIoU Faster R-CNN 1×/3× on Hub
+- **v0.2** (shipped): Rotated FCOS (anchor-free single-stage); Hub `rotated_fcos_dota_le90_3x_riou` (81.58% eval-val mAP50)
+- **v0.3**: HRSC2016 and FAIR1M dataset support
+- **v0.4**: RTMDet-R and native YOLO-OBB (AGPL-free production tier)
+- **v0.5+**: Swin-FPN backbone; optional fused CUDA kernels; hosted docs
 
 ## Contributing
 
@@ -202,7 +212,7 @@ Contributions are welcome. Run tests with `pytest`, format with `black` and `ruf
 
 ## Publishing to PyPI
 
-Version **0.1.0** — tag releases as **`v0.1.0`** (git) matching `version` in `pyproject.toml`.
+Version **0.2.0** — tag releases as **`v0.2.0`** (git) matching `version` in `pyproject.toml`.
 
 Configs: edit **`configs/`** at the repo root, then **`make sync-configs`** so **`oriented_det/configs/`** stays in sync (see **`oriented_det/configs/vendored_manifest.txt`**). CI runs **`make check-configs`**.
 
@@ -287,7 +297,7 @@ Create a [GitHub Release](https://github.com/DL4EO/oriented-det/releases) from t
 
 #### 6. After publish
 
-- Change install docs from “when published” to `pip install oriented-det`.
+- Verify `pip install oriented-det` in a clean environment.
 - Users still install **PyTorch** separately for their CUDA/CPU platform.
 
 ### Makefile targets

{oriented_det-0.1.0 → oriented_det-0.2.0}/README.md

@@ -7,7 +7,7 @@
 - **Geometry**: Rotated bounding boxes (rbox: cx, cy, w, h, angle), quadrilateral boxes (qbox), polygon ↔ rbox ↔ hbox conversions, angle normalization (le90, 0–180°), flip/rotate/scale transforms, visualization helpers
 - **IoU & NMS**: Rotated IoU and oriented NMS (CPU with optional GPU kernels when available); AABB pre-filtering; `obb_to_xyxy` / HBB conversion
 - **Datasets**: DOTA polygon loader (pattern, split file, or separate folders), image tiling, label filtering, ignore masks, oriented mAP evaluation
-- **Models**: **Oriented R-CNN** (horizontal RPN + MidpointOffset → oriented ROI), **Rotated Faster R-CNN** (oriented RPN + oriented ROI), **Rotated RetinaNet** (oriented anchors, focal loss); ResNet + FPN backbones; selective loading of external checkpoints where configs wire `checkpoint.load_from_checkpoint`
+- **Models**: **Oriented R-CNN** ([Xie et al., ICCV 2021](https://openaccess.thecvf.com/content/ICCV2021/html/Xie_Oriented_R-CNN_for_Object_Detection_ICCV_2021_paper.html); horizontal RPN + MidpointOffset → oriented RoIAlign + oriented ROI head), **Rotated Faster R-CNN** (Ren et al., NeurIPS 2015 two-stage baseline with horizontal RPN + horizontal RoIAlign + rotated ROI head; MMRotate reference), **Rotated RetinaNet** ([Lin et al., ICCV 2017](https://openaccess.thecvf.com/content_ICCV_2017/papers/Lin_Focal_Loss_for_ICCV_2017_paper.pdf); oriented anchors, sigmoid focal loss), **Rotated FCOS** (anchor-free single-stage; distance-angle coder, centerness, L1 / KFIoU / decoded rIoU); ResNet + FPN backbones; selective loading of external checkpoints where configs wire `checkpoint.load_from_checkpoint`
 - **Training**: JSON configs + **`odet train`**, mixed precision (AMP), gradient accumulation, checkpointing, best-metric tracking, TensorBoard, optional curriculum learning and profiling
 
 ## Installation
@@ -32,8 +32,10 @@ uv pip install -r requirements.txt
 uv pip install -e .
 ```
 
-- From PyPI (when published): `pip install oriented-det`
+- To auto-activate `.venv` when entering this repo, install the [`direnv`](https://direnv.net/) shell hook and run `direnv allow` from the repo root. The local `.envrc` is intentionally gitignored so each developer can opt in on their machine.
+- From PyPI: `pip install oriented-det`
 - For development and tests: `uv pip install -e ".[dev]"`
+- For the Gradio prediction viewer: `uv pip install -e ".[viewer]"` or `pip install "oriented-det[viewer]"`
 - For **macOS Apple Silicon** or **CPU-only**, see [Installation](docs/getting-started/installation.md).
 - Verify: `pytest tests/test_geometry.py tests/test_iou.py tests/test_nms.py`
 
@@ -43,13 +45,13 @@ After [installation](#installation):
 
 ```bash
 # Train on DOTA tiles (edit dataset paths in the config first)
-odet train --config configs/rotated_faster_rcnn/dota_le90_3x.json
+odet train --config configs/oriented_rcnn/dota_le90_1x.json
 
 # Or use the Makefile wrapper (same default config)
 make train
 ```
 
-Default 3× recipe: `configs/rotated_faster_rcnn/dota_le90_3x.json` (see [configs/rotated_faster_rcnn/README.md](configs/rotated_faster_rcnn/README.md)). 1× baseline: `dota_le90_1x.json`.
+Default starter recipe: `configs/oriented_rcnn/dota_le90_1x.json` (see [configs/oriented_rcnn/README.md](configs/oriented_rcnn/README.md)). Use `configs/oriented_rcnn/dota_le90_3x.json` when you want the longer Oriented R-CNN schedule.
 
 Programmatic APIs and a longer walkthrough: [Getting Started](docs/getting-started/installation.md). Config fields: [Configuration](docs/user-guide/configuration.md).
 
@@ -92,7 +94,7 @@ After that, copy-pasted commands and unmodified configs that reference `/path/to
 Full documentation is in the **docs/** folder and can be built and served with MkDocs:
 
 - **Build/serve**: `make docs` or `make docs-serve` (see [docs/README.md](docs/README.md)); or `uv pip install -e ".[docs]"` then `mkdocs serve`.
-- **Guides**: [Getting Started](docs/getting-started/installation.md), [User Guide](docs/user-guide/geometry.md), [API Reference](docs/api/geometry.md), [Examples](docs/examples/visualization.md).
+- **Guides**: [Getting Started](docs/getting-started/installation.md), [User Guide](docs/user-guide/geometry.md), [API Reference](docs/api/geometry.md), [Examples](docs/examples/inference.md).
 
 ## Documentation by folder
 
@@ -113,18 +115,18 @@ Install once: `uv pip install -e .`. Then:
 
 | Task | Command |
 |------|---------|
-| Train | `odet train --config configs/rotated_faster_rcnn/dota_le90_3x.json` or `make train` |
+| Train | `odet train --config configs/oriented_rcnn/dota_le90_1x.json` or `make train` |
 | Multi-GPU | `make train-multi-gpu` (`torchrun` + cuDNN on `LD_LIBRARY_PATH`) |
 | Tile DOTA | `odet tile-dota /path/to/dota/train` |
-| Val predictions | `odet preds --experiment-dir runs/rotated_faster_rcnn/<timestamp>` or `make preds` |
+| Val predictions | `odet preds --experiment-dir runs/oriented_rcnn/<timestamp>` or `make preds` |
 | Offline mAP | `make eval-val` or `make preds` then `make metrics` |
 
 DOTA configs: per-model `dota_le90_1x.json` / `dota_le90_3x.json` under [configs/](configs/). Run `odet --help` for all subcommands. Makefile shortcuts and script-level options: [tools/README.md](tools/README.md). Config reference: [docs/user-guide/configuration.md](docs/user-guide/configuration.md), [configs/config.schema.json](configs/config.schema.json), [configs/README.md](configs/README.md).
 
 ## Pretrained weights and evaluation
 
-- Place exported best checkpoints under **`pretrained/`** or use Hub slugs (`odet pretrained download oriented_rcnn_dota_le90_1x`). See [pretrained/README.md](pretrained/README.md) and [configs/README.md](configs/README.md#dota-pretrained-models-model-zoo).
-- **Tiled validation:** after training, run `make preds` then `make metrics` (or `odet preds` / `odet preds --metrics-from-json`). Large images use sliding-window inference when they exceed the model canvas; thresholds and overlap come from **`production.*`** in the experiment config.
+- Place exported best checkpoints under **`pretrained/`** or use Hub slugs (`odet pretrained download oriented_rcnn_dota_le90_3x` or `oriented_rcnn_dota_le90_1x`). See [pretrained/README.md](pretrained/README.md) and [configs/README.md](configs/README.md#dota-pretrained-models-model-zoo).
+- **Tiled validation:** after training, run `make preds` then `make metrics`. Published mAP reports: [`docs/eval-reports/`](docs/eval-reports/) (git). Raw detections for the viewer: gitignored [`predictions/`](predictions/).
 
 ## Important notes
 
@@ -134,9 +136,14 @@ DOTA configs: per-model `dota_le90_1x.json` / `dota_le90_3x.json` under [configs
 
 ## Roadmap
 
-- **v0.1** (current): Geometry, IoU/NMS, DOTA loader + tiler, three baseline detectors, config-based training, Hub pretrained weights
-- **Next**: HRSC2016 dataset support, expanded tutorials, hosted docs
-- **Future**: Additional oriented detectors (e.g. FAIR1M), optional fused CUDA kernels
+See **[docs/roadmap.md](docs/roadmap.md)** for the full public plan. Summary:
+
+- **v0.1** (shipped): Geometry, IoU/NMS, DOTA, three ResNet-FPN detectors, config training, Hub pretrained weights
+- **v0.1.1** (shipped): ProbIoU Faster R-CNN 1×/3× on Hub
+- **v0.2** (shipped): Rotated FCOS (anchor-free single-stage); Hub `rotated_fcos_dota_le90_3x_riou` (81.58% eval-val mAP50)
+- **v0.3**: HRSC2016 and FAIR1M dataset support
+- **v0.4**: RTMDet-R and native YOLO-OBB (AGPL-free production tier)
+- **v0.5+**: Swin-FPN backbone; optional fused CUDA kernels; hosted docs
 
 ## Contributing
 
@@ -144,7 +151,7 @@ Contributions are welcome. Run tests with `pytest`, format with `black` and `ruf
 
 ## Publishing to PyPI
 
-Version **0.1.0** — tag releases as **`v0.1.0`** (git) matching `version` in `pyproject.toml`.
+Version **0.2.0** — tag releases as **`v0.2.0`** (git) matching `version` in `pyproject.toml`.
 
 Configs: edit **`configs/`** at the repo root, then **`make sync-configs`** so **`oriented_det/configs/`** stays in sync (see **`oriented_det/configs/vendored_manifest.txt`**). CI runs **`make check-configs`**.
 
@@ -229,7 +236,7 @@ Create a [GitHub Release](https://github.com/DL4EO/oriented-det/releases) from t
 
 #### 6. After publish
 
-- Change install docs from “when published” to `pip install oriented-det`.
+- Verify `pip install oriented-det` in a clean environment.
 - Users still install **PyTorch** separately for their CUDA/CPU platform.
 
 ### Makefile targets

oriented_det-0.2.0/export/contract.json

@@ -0,0 +1,73 @@
+{
+  "version": 3,
+  "description": "Inference export contract for oriented-det (ONNX + TensorFlow/Keras detect bundle).",
+  "input": {
+    "name": "images",
+    "dtype": "float32",
+    "layout": "NCHW",
+    "value_range": "[0, 1] RGB matching training preprocess (see TrainingExperimentConfig.preprocessing)",
+    "spatial": "Fixed H×W at export time; batch size 1 for pre-NMS detect modes."
+  },
+  "export_modes": {
+    "backbone": {
+      "models": ["rotated_retinanet", "oriented_rcnn", "rotated_faster_rcnn", "rotated_fcos"],
+      "onnx_outputs": "Tuple of FPN feature maps in stable key order (see BackboneExportWrapper).",
+      "postprocess": "None in ONNX; heads and decoding remain PyTorch or separate TF code."
+    },
+    "retinanet_heads": {
+      "models": ["rotated_retinanet"],
+      "onnx_outputs": "Interleaved per level: level{i}_cls_logits, level{i}_bbox_pred (i from 0).",
+      "postprocess": "Anchor grid, decode_oriented_boxes, score filter, rotated NMS — not in this ONNX."
+    },
+    "faster_rcnn_pre_nms": {
+      "models": ["rotated_faster_rcnn"],
+      "onnx_outputs": {
+        "pre_nms_boxes": "[P, 5] oriented boxes (cx,cy,w,h,angle), P = max_pre_nms_candidates",
+        "pre_nms_scores": "[P]",
+        "pre_nms_labels": "[P] 1-based foreground ids",
+        "pre_nms_count": "scalar int64 valid count <= P"
+      },
+      "postprocess": "Exact CPU rotated NMS + production score_threshold in TF/Keras bundle (export/scripts/build_faster_rcnn_savedmodel.py)."
+    },
+    "oriented_rcnn_pre_nms": {
+      "models": ["oriented_rcnn"],
+      "onnx_outputs": {
+        "pre_nms_boxes": "[P, 5] oriented boxes (cx,cy,w,h,angle), P = max_pre_nms_candidates",
+        "pre_nms_scores": "[P]",
+        "pre_nms_labels": "[P] 1-based foreground ids",
+        "pre_nms_count": "scalar int64 valid count <= P"
+      },
+      "postprocess": "Same Keras detect bundle as faster_rcnn_pre_nms (ORT core + exact rotated NMS)."
+    },
+    "rotated_fcos_pre_nms": {
+      "models": ["rotated_fcos"],
+      "onnx_outputs": {
+        "pre_nms_boxes": "[P, 5] oriented boxes (cx,cy,w,h,angle), P = max_pre_nms_candidates",
+        "pre_nms_scores": "[P]",
+        "pre_nms_labels": "[P] 1-based foreground ids",
+        "pre_nms_count": "scalar int64 valid count <= P"
+      },
+      "postprocess": "Same Keras detect bundle as faster_rcnn_pre_nms (ORT core + exact rotated NMS)."
+    }
+  },
+  "detect_bundle": {
+    "files": ["keras_model.keras", "model.onnx", "export_meta.json"],
+    "outputs": {
+      "detections": "[max_detections, 7] as (cx,cy,w,h,angle,score,label)",
+      "num_detections": "scalar int32"
+    },
+    "load": "export.tf_serving_model.load_keras_detect_model(bundle_dir_or_keras_path)"
+  },
+  "nms": {
+    "in_onnx": false,
+    "in_keras_bundle": true,
+    "pytorch_reference": "oriented_det.ops.rotated_ops rotated_nms (force_cpu for deploy parity)"
+  },
+  "notes": [
+    "faster_rcnn_pre_nms and oriented_rcnn_pre_nms pad RPN proposals to fixed P before ROI align for traceable ONNX.",
+    "oriented_rcnn_pre_nms uses midpoint RPN + OrientedROIAlign (masked all-N grid_sample under ONNX export).",
+    "rotated_fcos_pre_nms decodes DistanceAnglePointCoder + per-level topk, then pads to nms_pre × FPN levels. No ROI.",
+    "onnx2tf SavedModel for this graph may fail on ScatterND; the detect bundle uses ONNX Runtime inside Keras.",
+    "TFLite full detect is not supported while NMS remains in Python."
+  ]
+}

oriented_det-0.2.0/export/ort_runtime.py

@@ -0,0 +1,144 @@
+"""ONNX Runtime device selection and session cache for TF export inference."""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, List, Optional, Sequence, Tuple, Union
+
+_ORT_DEVICE_OVERRIDE: Optional[str] = None
+_SESSION_CACHE: Dict[Tuple[str, Tuple[str, ...]], object] = {}
+_TF_GPUS_HIDDEN: bool = False
+
+
+def set_ort_device(device: Optional[str]) -> None:
+    """Set runtime ORT device for this process (``cpu``, ``cuda``, ``auto``)."""
+    global _ORT_DEVICE_OVERRIDE
+    _ORT_DEVICE_OVERRIDE = device.lower().strip() if device else None
+
+
+def get_ort_device() -> str:
+    """Resolved ORT device string (override, env, or default ``cpu``)."""
+    if _ORT_DEVICE_OVERRIDE:
+        return _ORT_DEVICE_OVERRIDE
+    return (os.environ.get("ORIENTED_DET_ORT_DEVICE") or "cpu").lower().strip()
+
+
+def ort_providers_for_device(device: Optional[str] = None) -> List[str]:
+    """Map device string to ONNX Runtime ``providers`` list."""
+    import onnxruntime as ort
+
+    d = (device if device is not None else get_ort_device()).lower().strip()
+    if d in ("cuda", "gpu"):
+        if "CUDAExecutionProvider" not in ort.get_available_providers():
+            raise RuntimeError(
+                "ORT device=cuda requested but CUDAExecutionProvider is not available. "
+                "Install onnxruntime-gpu matching your CUDA driver "
+                '(e.g. pip install "oriented-det[export]" / onnxruntime-gpu[cuda,cudnn]).'
+            )
+        return ["CUDAExecutionProvider", "CPUExecutionProvider"]
+    if d == "auto":
+        if "CUDAExecutionProvider" in ort.get_available_providers():
+            return ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        return ["CPUExecutionProvider"]
+    if d != "cpu":
+        raise ValueError(f"Unknown ORT device {device!r}; use cpu, cuda, or auto.")
+    return ["CPUExecutionProvider"]
+
+
+def hide_tensorflow_gpus() -> bool:
+    """Hide all GPUs from TensorFlow so ORT can own CUDA memory.
+
+    Call **before** TF creates CUDA contexts (ideally before ``import tensorflow``
+    finishes initializing devices). When ORT uses ``CUDAExecutionProvider``, TF
+    otherwise claims every visible GPU and ORT then OOMs on ROI Align temps
+    (e.g. ``/roi_align/Transpose_2`` ~287 MB that should fit on an A100).
+
+    Returns:
+        True if GPUs are hidden from TF (or none were visible).
+    """
+    global _TF_GPUS_HIDDEN
+    if _TF_GPUS_HIDDEN:
+        return True
+    try:
+        import tensorflow as tf
+
+        gpus = tf.config.list_physical_devices("GPU")
+        if gpus:
+            tf.config.set_visible_devices([], "GPU")
+        _TF_GPUS_HIDDEN = True
+        return True
+    except Exception as exc:  # pragma: no cover - depends on TF init order
+        print(
+            f"Warning: could not hide GPUs from TensorFlow ({exc}). "
+            "ORT CUDA may OOM if TF already initialized CUDA. "
+            "Configure ORT device before importing tensorflow / tf_serving_model."
+        )
+        return False
+
+
+def prepare_tensorflow_for_ort(providers: Sequence[str]) -> None:
+    """If ORT will use CUDA, keep TensorFlow on CPU only."""
+    if "CUDAExecutionProvider" in providers:
+        hide_tensorflow_gpus()
+
+
+def _cuda_provider_options() -> Dict[str, Any]:
+    """ORT CUDA EP options (device id, arena, optional mem cap)."""
+    opts: Dict[str, Any] = {
+        "device_id": int(os.environ.get("ORIENTED_DET_ORT_CUDA_DEVICE_ID", "0")),
+        "arena_extend_strategy": "kSameAsRequested",
+        "cudnn_conv_algo_search": os.environ.get(
+            "ORIENTED_DET_ORT_CUDNN_CONV_ALGO", "HEURISTIC"
+        ),
+    }
+    mem = os.environ.get("ORIENTED_DET_ORT_GPU_MEM_LIMIT")
+    if mem:
+        opts["gpu_mem_limit"] = int(mem)
+    return opts
+
+
+def _session_providers(
+    providers: Sequence[str],
+) -> List[Union[str, Tuple[str, Dict[str, Any]]]]:
+    """Expand provider names into ORT session provider list (with CUDA options)."""
+    out: List[Union[str, Tuple[str, Dict[str, Any]]]] = []
+    for p in providers:
+        if p == "CUDAExecutionProvider":
+            out.append((p, _cuda_provider_options()))
+        else:
+            out.append(p)
+    return out
+
+
+def configure_ort_device(device: Optional[str] = None) -> List[str]:
+    """Apply device override, hide TF GPUs when needed, return provider names."""
+    if device is not None:
+        set_ort_device(device)
+    providers = ort_providers_for_device()
+    prepare_tensorflow_for_ort(providers)
+    return providers
+
+
+def clear_ort_session_cache() -> None:
+    """Drop cached ORT sessions (tests)."""
+    _SESSION_CACHE.clear()
+
+
+def get_ort_session(onnx_path: str, device: Optional[str] = None):
+    """Return a cached ``onnxruntime.InferenceSession`` for ``onnx_path``."""
+    import onnxruntime as ort
+
+    providers = ort_providers_for_device(device)
+    prepare_tensorflow_for_ort(providers)
+    session_providers = _session_providers(providers)
+    # Cache key uses provider *names* only (options are env-stable per process).
+    key = (str(onnx_path), tuple(providers))
+    if key not in _SESSION_CACHE:
+        so = ort.SessionOptions()
+        so.enable_mem_pattern = True
+        _SESSION_CACHE[key] = ort.InferenceSession(
+            str(onnx_path),
+            sess_options=so,
+            providers=session_providers,
+        )
+    return _SESSION_CACHE[key]