YoloNAS-R

LICENSE.YOLONAS-R.md

@@ -0,0 +1,16 @@
+# YOLO-NAS-R License
+
+These model weights or any components comprising the model and the associated documentation (the "Software") is licensed to you by Deci.AI, Inc. ("Deci") under the following terms:
+© 2023 – Deci.AI, Inc. 
+
+Subject to your full compliance with all of the terms herein, Deci hereby grants you a non-exclusive, revocable, non-sublicensable, non-transferable worldwide and limited right and license to use the Software. If you are using the Deci platform for model optimization, your use of the Software is subject to the Terms of Use available here (the "Terms of Use"). 
+
+You shall not, without Deci's prior written consent: 
+(i) resell, lease, sublicense or distribute the Software to any person; 
+(ii) use the Software to provide third parties with managed services or provide remote access to the Software to any person or compete with Deci in any way; 
+(iii) represent that you possess any proprietary interest in the Software; 
+(iv) directly or indirectly, take any action to contest Deci's intellectual property rights or infringe them in any way; 
+(V) reverse-engineer, decompile, disassemble, alter, enhance, improve, add to, delete from, or otherwise modify, or derive (or attempt to derive) the technology or source code underlying any part of the Software; 
+(vi) use the Software (or any part thereof) in any illegal, indecent, misleading, harmful, abusive, harassing and/or disparaging manner or for any such purposes. Except as provided under the terms of any separate agreement between you and Deci, including the Terms of Use to the extent applicable, you may not use the Software for any commercial use, including in connection with any models used in a production environment. 
+
+DECI PROVIDES THE SOFTWARE "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS OF THE SOFTWARE BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 

src/super_gradients/training/datasets/data_formats/obb/cxcywhr.py

@@ -0,0 +1,58 @@
+import cv2
+import numpy as np
+
+
+def cxcywhr_to_poly(boxes: np.ndarray) -> np.ndarray:
+    """
+    Convert oriented bounding boxes in CX-CY-W-H-R format to a polygon format
+    :param boxes: [N,...,5] oriented bounding boxes in CX-CY-W-H-R format
+    :return: [N,...,4, 2] oriented bounding boxes in polygon format
+    """
+    shape = boxes.shape
+    if shape[-1] != 5:
+        raise ValueError(f"Expected last dimension to be 5, got {shape[-1]}")
+
+    flat_rboxes = boxes.reshape(-1, 5).astype(np.float32)
+    polys = np.zeros((flat_rboxes.shape[0], 4, 2), dtype=np.float32)
+    for i, box in enumerate(flat_rboxes):
+        cx, cy, w, h, r = box
+        rect = ((cx, cy), (w, h), np.rad2deg(r))
+        poly = cv2.boxPoints(rect)
+        polys[i] = poly
+
+    return polys.reshape(*shape[:-1], 4, 2)
+
+
+def poly_to_cxcywhr(poly: np.ndarray) -> np.ndarray:
+    shape = poly.shape
+    if shape[-2:] != (4, 2):
+        raise ValueError(f"Expected last two dimensions to be (4, 2), got {shape[-2:]}")
+
+    flat_polys = poly.reshape(-1, 4, 2)
+    rboxes = np.zeros((flat_polys.shape[0], 5), dtype=np.float32)
+    for i, poly in enumerate(flat_polys):
+        hull = cv2.convexHull(np.reshape(poly, [-1, 2]).astype(np.float32))
+        rect = cv2.minAreaRect(hull)
+        cx, cy = rect[0]
+        w, h = rect[1]
+        angle = rect[2]
+        if h > w:
+            w, h = h, w
+            angle += 90
+        angle = np.deg2rad(angle)
+        rboxes[i] = [cx, cy, w, h, angle]
+
+    return rboxes.reshape(*shape[:-2], 5)
+
+
+def poly_to_xyxy(poly: np.ndarray) -> np.ndarray:
+    """
+    Convert oriented bounding boxes in polygon format to XYXY format
+    :param poly: [..., 4, 2]
+    :return: [..., 4]
+    """
+    x1 = poly[..., :, 0].min(axis=-1)
+    y1 = poly[..., :, 1].min(axis=-1)
+    x2 = poly[..., :, 0].max(axis=-1)
+    y2 = poly[..., :, 1].max(axis=-1)
+    return np.stack([x1, y1, x2, y2], axis=-1)

src/super_gradients/training/models/model_factory.py

@@ -23,7 +23,7 @@
 )
 from super_gradients.common.abstractions.abstract_logger import get_logger
 from super_gradients.training.utils.sg_trainer_utils import get_callable_param_names
-from super_gradients.training.processing.processing import get_pretrained_processing_params
+from super_gradients.training.processing import get_pretrained_processing_params
 
 logger = get_logger(__name__)
 

src/super_gradients/training/pretrained_models.py

@@ -59,6 +59,9 @@
     "yolo_nas_pose_s_coco_pose": "https://sghub.deci.ai/models/yolo_nas_pose_s_coco_pose.pth",
     "yolo_nas_pose_m_coco_pose": "https://sghub.deci.ai/models/yolo_nas_pose_m_coco_pose.pth",
     "yolo_nas_pose_l_coco_pose": "https://sghub.deci.ai/models/yolo_nas_pose_l_coco_pose.pth",
+    "yolo_nas_r_s_dota2": "https://sghub.deci.ai/models/yolo_nas_r_s_dota2.pth",
+    "yolo_nas_r_m_dota2": "https://sghub.deci.ai/models/yolo_nas_r_m_dota2.pth",
+    "yolo_nas_r_l_dota2": "https://sghub.deci.ai/models/yolo_nas_r_l_dota2.pth",
 }
 
 PRETRAINED_NUM_CLASSES = {
@@ -69,6 +72,7 @@
     "coco": 80,
     "coco_pose": 17,
     "cifar10": 10,
+    "dota2": 18,
 }
 
 DATASET_LICENSES = {
@@ -79,4 +83,5 @@
     "coco_pose": "https://cocodataset.org/#termsofuse",
     "cityscapes": "https://www.cs.toronto.edu/~kriz/cifar.html",
     "objects365": "https://www.objects365.org/download.html",
+    "dota2": "https://captain-whu.github.io/DOTA/dataset.html",
 }

src/super_gradients/training/processing/__init__.py

@@ -1,4 +1,5 @@
 from .processing import (
+    Processing,
     StandardizeImage,
     DetectionRescale,
     DetectionLongestMaxSizeRescale,
@@ -14,8 +15,10 @@
     SegmentationPadShortToCropSize,
     SegmentationPadToDivisible,
 )
+from .defaults import get_pretrained_processing_params
 
 __all__ = [
+    "Processing",
     "StandardizeImage",
     "DetectionRescale",
     "DetectionLongestMaxSizeRescale",
@@ -30,4 +33,5 @@
     "SegmentationResize",
     "SegmentationPadShortToCropSize",
     "SegmentationPadToDivisible",
+    "get_pretrained_processing_params",
 ]