birder 0.2.0__tar.gz → 0.2.1__tar.gz

{birder-0.2.0 → birder-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: birder
-Version: 0.2.0
+Version: 0.2.1
 Summary: An open-source computer vision framework for wildlife image analysis, featuring state-of-the-art models for species classification and detection.
 Author: Ofer Hasson
 License-Expression: Apache-2.0
@@ -48,7 +48,7 @@ Requires-Dist: black~=25.12.0; extra == "dev"
 Requires-Dist: build~=1.3.0; extra == "dev"
 Requires-Dist: bumpver~=2025.1131; extra == "dev"
 Requires-Dist: captum~=0.7.0; extra == "dev"
-Requires-Dist: coverage~=7.13.0; extra == "dev"
+Requires-Dist: coverage~=7.13.1; extra == "dev"
 Requires-Dist: debugpy; extra == "dev"
 Requires-Dist: flake8-pep585~=0.1.7; extra == "dev"
 Requires-Dist: flake8~=7.3.0; extra == "dev"

{birder-0.2.0 → birder-0.2.1}/birder/data/collators/detection.py

@@ -3,7 +3,9 @@ import random
 from typing import Any
 
 import torch
+from torchvision import tv_tensors
 from torchvision.transforms import v2
+from torchvision.transforms.v2 import functional as F
 
 BATCH_MULTISCALE_SIZES = (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
 
@@ -84,6 +86,16 @@ class BatchRandomResizeCollator(DetectionCollator):
         resized_images = []
         resized_targets = []
         for image, target in zip(images, targets):
+            if "boxes" in target:
+                boxes = target["boxes"]
+                if isinstance(boxes, tv_tensors.BoundingBoxes) is False:
+                    if boxes.numel() == 0:
+                        boxes = boxes.reshape(0, 4)
+                    boxes = tv_tensors.BoundingBoxes(
+                        boxes, format=tv_tensors.BoundingBoxFormat.XYXY, canvas_size=F.get_size(image)
+                    )
+                    target = {**target, "boxes": boxes}
+
             (image, target) = resize(image, target)
             resized_images.append(image)
             resized_targets.append(target)

{birder-0.2.0 → birder-0.2.1}/birder/data/transforms/detection.py

@@ -129,7 +129,7 @@ def get_birder_augment(
     return v2.Compose(transformations)  # type: ignore
 
 
-AugType = Literal["birder", "lsj", "multiscale", "ssd", "ssdlite", "detr"]
+AugType = Literal["birder", "lsj", "multiscale", "ssd", "ssdlite", "yolo", "detr"]
 
 
 def training_preset(
@@ -235,6 +235,34 @@ def training_preset(
             ]
         )
 
+    if aug_type == "yolo":
+        return v2.Compose(  # type: ignore
+            [
+                v2.ToImage(),
+                v2.RandomPhotometricDistort(),
+                (
+                    v2.RandomAffine(
+                        degrees=10.0,
+                        translate=(0.1, 0.1),
+                        scale=(0.5, 1.5),
+                        shear=2.0,
+                        interpolation=v2.InterpolationMode.BILINEAR,
+                        fill=fill_value,
+                    )
+                    if post_mosaic is False
+                    else v2.Identity()
+                ),
+                ResizeWithRandomInterpolation(
+                    target_size, max_size, interpolation=[v2.InterpolationMode.BILINEAR, v2.InterpolationMode.BICUBIC]
+                ),
+                v2.RandomHorizontalFlip(0.5),
+                v2.SanitizeBoundingBoxes(),
+                v2.ToDtype(torch.float32, scale=True),
+                v2.Normalize(mean=mean, std=std),
+                v2.ToPureTensor(),
+            ]
+        )
+
     if aug_type == "detr":
         return v2.Compose(  # type: ignore
             [

{birder-0.2.0 → birder-0.2.1}/birder/net/detection/yolo_v3.py

@@ -25,12 +25,26 @@ from birder.net.detection.base import ImageList
 # These values are in absolute pixels (width, height) computed using K-Means
 # on the COCO dataset with a reference input size of 416x416.
 DEFAULT_ANCHORS = [
-    [(10, 13), (16, 30), (33, 23)],  # Small objects (stride 8)
-    [(30, 61), (62, 45), (59, 119)],  # Medium objects (stride 16)
-    [(116, 90), (156, 198), (373, 326)],  # Large objects (stride 32)
+    [(10.0, 13.0), (16.0, 30.0), (33.0, 23.0)],  # Small objects (stride 8)
+    [(30.0, 61.0), (62.0, 45.0), (59.0, 119.0)],  # Medium objects (stride 16)
+    [(116.0, 90.0), (156.0, 198.0), (373.0, 326.0)],  # Large objects (stride 32)
 ]
 
 
+def scale_anchors(
+    anchors: list[list[tuple[float, float]]],
+    from_size: tuple[int, int],
+    to_size: tuple[int, int],
+) -> list[list[tuple[float, float]]]:
+    if from_size == to_size:
+        # Avoid aliasing default anchors in case they are mutated later.
+        return [list(scale) for scale in anchors]
+
+    scale_h = to_size[0] / from_size[0]
+    scale_w = to_size[1] / from_size[1]
+    return [[(w * scale_w, h * scale_h) for (w, h) in scale] for scale in anchors]
+
+
 def decode_predictions(
     predictions: torch.Tensor,
     anchors: torch.Tensor,
@@ -100,7 +114,7 @@ def decode_predictions(
 
 
 class YOLOAnchorGenerator(nn.Module):
-    def __init__(self, anchors: list[list[tuple[int, int]]]) -> None:
+    def __init__(self, anchors: list[list[tuple[float, float]]]) -> None:
         super().__init__()
         self.anchors = anchors
         self.num_scales = len(anchors)
@@ -332,7 +346,7 @@ class YOLO_v3(DetectionBaseNet):
         self.obj_coeff = 1.0
         self.cls_coeff = 1.0
 
-        self.anchors = DEFAULT_ANCHORS
+        self.anchors = scale_anchors(DEFAULT_ANCHORS, self.default_size, self.size)
         self.score_thresh = score_thresh
         self.nms_thresh = nms_thresh
         self.detections_per_img = detections_per_img
@@ -354,6 +368,15 @@ class YOLO_v3(DetectionBaseNet):
         num_anchors = self.anchor_generator.num_anchors_per_location()
         self.head = YOLOHead(self.neck.out_channels, num_anchors, self.num_classes)
 
+    def adjust_size(self, new_size: tuple[int, int]) -> None:
+        if new_size == self.size:
+            return
+
+        old_size = self.size
+        super().adjust_size(new_size)
+        self.anchors = scale_anchors(self.anchors, old_size, new_size)
+        self.anchor_generator.anchors = self.anchors
+
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
             param.requires_grad = False
@@ -682,6 +705,13 @@ class YOLO_v3(DetectionBaseNet):
         neck_features = self.neck(features)
         predictions = self.head(neck_features)
         (anchors, grids, strides) = self.anchor_generator(images, neck_features)
+        if self.dynamic_size is True:
+            image_size = (images.tensors.shape[-2], images.tensors.shape[-1])
+            if image_size[0] != self.size[0] or image_size[1] != self.size[1]:
+                scale_w = image_size[1] / self.size[1]
+                scale_h = image_size[0] / self.size[0]
+                scale_tensor = torch.tensor([scale_w, scale_h], device=anchors[0].device, dtype=anchors[0].dtype)
+                anchors = [anchor * scale_tensor for anchor in anchors]
 
         losses: dict[str, torch.Tensor] = {}
         detections: list[dict[str, torch.Tensor]] = []

{birder-0.2.0 → birder-0.2.1}/birder/net/detection/yolo_v4.py

@@ -21,12 +21,13 @@ from birder.net.base import DetectorBackbone
 from birder.net.detection.base import DetectionBaseNet
 from birder.net.detection.yolo_v3 import YOLOAnchorGenerator
 from birder.net.detection.yolo_v3 import YOLOHead
+from birder.net.detection.yolo_v3 import scale_anchors
 
 # Default anchors from YOLO v4 (COCO)
 DEFAULT_ANCHORS = [
-    [(12, 16), (19, 36), (40, 28)],  # Small
-    [(36, 75), (76, 55), (72, 146)],  # Medium
-    [(142, 110), (192, 243), (459, 401)],  # Large
+    [(12.0, 16.0), (19.0, 36.0), (40.0, 28.0)],  # Small
+    [(36.0, 75.0), (76.0, 55.0), (72.0, 146.0)],  # Medium
+    [(142.0, 110.0), (192.0, 243.0), (459.0, 401.0)],  # Large
 ]
 
 # Scale factors per detection scale to eliminate grid sensitivity
@@ -410,7 +411,7 @@ class YOLO_v4(DetectionBaseNet):
         self.obj_coeff = 1.0
         self.cls_coeff = 1.0
 
-        self.anchors = DEFAULT_ANCHORS
+        self.anchors = scale_anchors(DEFAULT_ANCHORS, self.default_size, self.size)
         self.scale_xy = DEFAULT_SCALE_XY
         self.score_thresh = score_thresh
         self.nms_thresh = nms_thresh
@@ -438,6 +439,15 @@ class YOLO_v4(DetectionBaseNet):
         num_anchors = self.anchor_generator.num_anchors_per_location()
         self.head = YOLOHead(self.neck.out_channels, num_anchors, self.num_classes)
 
+    def adjust_size(self, new_size: tuple[int, int]) -> None:
+        if new_size == self.size:
+            return
+
+        old_size = self.size
+        super().adjust_size(new_size)
+        self.anchors = scale_anchors(self.anchors, old_size, new_size)
+        self.anchor_generator = YOLOAnchorGenerator(self.anchors)
+
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
             param.requires_grad = False
@@ -799,6 +809,13 @@ class YOLO_v4(DetectionBaseNet):
         neck_features = self.neck(features)
         predictions = self.head(neck_features)
         (anchors, grids, strides) = self.anchor_generator(images, neck_features)
+        if self.dynamic_size is True:
+            image_size = (images.tensors.shape[-2], images.tensors.shape[-1])
+            if image_size[0] != self.size[0] or image_size[1] != self.size[1]:
+                scale_w = image_size[1] / self.size[1]
+                scale_h = image_size[0] / self.size[0]
+                scale_tensor = torch.tensor([scale_w, scale_h], device=anchors[0].device, dtype=anchors[0].dtype)
+                anchors = [anchor * scale_tensor for anchor in anchors]
 
         losses: dict[str, torch.Tensor] = {}
         detections: list[dict[str, torch.Tensor]] = []

{birder-0.2.0 → birder-0.2.1}/birder/net/detection/yolo_v4_tiny.py

@@ -18,12 +18,13 @@ from torchvision.ops import Conv2dNormActivation
 from birder.net.base import DetectorBackbone
 from birder.net.detection.yolo_v3 import YOLOAnchorGenerator
 from birder.net.detection.yolo_v3 import YOLOHead
+from birder.net.detection.yolo_v3 import scale_anchors
 from birder.net.detection.yolo_v4 import YOLO_v4
 
 # Default anchors from YOLO v4 Tiny (COCO)
 DEFAULT_ANCHORS = [
-    [(10, 14), (23, 27), (37, 58)],  # Medium
-    [(81, 82), (135, 169), (344, 319)],  # Large
+    [(10.0, 14.0), (23.0, 27.0), (37.0, 58.0)],  # Medium
+    [(81.0, 82.0), (135.0, 169.0), (344.0, 319.0)],  # Large
 ]
 
 # Scale factors per detection scale to eliminate grid sensitivity
@@ -118,7 +119,7 @@ class YOLO_v4_Tiny(YOLO_v4):
         self.obj_coeff = 1.0
         self.cls_coeff = 1.0
 
-        self.anchors = DEFAULT_ANCHORS
+        self.anchors = scale_anchors(DEFAULT_ANCHORS, self.default_size, self.size)
         self.scale_xy = DEFAULT_SCALE_XY
         self.score_thresh = score_thresh
         self.nms_thresh = nms_thresh

{birder-0.2.0 → birder-0.2.1}/birder/scripts/predict.py

@@ -629,8 +629,15 @@ def validate_args(args: argparse.Namespace) -> None:
 
     if args.wds is False and len(args.data_path) == 0:
         raise cli.ValidationError("Must provide at least one data source, --data-path or --wds")
-    if args.wds is True and len(args.data_path) > 1:
-        raise cli.ValidationError(f"--wds can have at most 1 --data-path, got {len(args.data_path)}")
+    if args.wds is True:
+        if args.wds_info is None and len(args.data_path) == 0:
+            raise cli.ValidationError("--wds requires a data path unless --wds-info is provided")
+        if len(args.data_path) > 1:
+            raise cli.ValidationError(f"--wds can have at most 1 --data-path, got {len(args.data_path)}")
+        if args.wds_info is None and len(args.data_path) == 1:
+            data_path = args.data_path[0]
+            if "://" in data_path and args.wds_size is None:
+                raise cli.ValidationError("--wds-size is required for remote --data-path")
     if args.wds is True and args.hierarchical is True:
         raise cli.ValidationError("--wds cannot be used with --hierarchical")
     if args.wds is True and args.ignore_dir_names is True:

{birder-0.2.0 → birder-0.2.1}/birder/scripts/train_detection.py

@@ -936,7 +936,7 @@ def validate_args(args: argparse.Namespace) -> None:
         if args.aug_type in {"multiscale", "detr"}:
             raise cli.ValidationError(
                 f"--batch-multiscale not supported with --aug-type {args.aug_type}, "
-                "use a fixed-size aug type (e.g. birder, ssd, ssdlite)"
+                "use a fixed-size aug type (e.g. birder, ssd, ssdlite, yolo)"
             )
     if args.mosaic_stop_epoch is not None:
         if args.mosaic_stop_epoch <= 0:

{birder-0.2.0 → birder-0.2.1}/birder/tools/show_det_iterator.py

@@ -314,7 +314,7 @@ def main(args: argparse.Namespace) -> None:
         if args.aug_type in {"multiscale", "detr"}:
             raise cli.ValidationError(
                 f"--batch-multiscale not supported with --aug-type {args.aug_type}, "
-                "use a fixed-size aug type (e.g. birder, ssd, ssdlite)"
+                "use a fixed-size aug type (e.g. birder, ssd, ssdlite, yolo)"
             )
 
     show_det_iterator(args)

birder-0.2.1/birder/version.py

@@ -0,0 +1 @@
+__version__ = "v0.2.1"

{birder-0.2.0 → birder-0.2.1}/birder.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: birder
-Version: 0.2.0
+Version: 0.2.1
 Summary: An open-source computer vision framework for wildlife image analysis, featuring state-of-the-art models for species classification and detection.
 Author: Ofer Hasson
 License-Expression: Apache-2.0
@@ -48,7 +48,7 @@ Requires-Dist: black~=25.12.0; extra == "dev"
 Requires-Dist: build~=1.3.0; extra == "dev"
 Requires-Dist: bumpver~=2025.1131; extra == "dev"
 Requires-Dist: captum~=0.7.0; extra == "dev"
-Requires-Dist: coverage~=7.13.0; extra == "dev"
+Requires-Dist: coverage~=7.13.1; extra == "dev"
 Requires-Dist: debugpy; extra == "dev"
 Requires-Dist: flake8-pep585~=0.1.7; extra == "dev"
 Requires-Dist: flake8~=7.3.0; extra == "dev"

{birder-0.2.0 → birder-0.2.1}/birder.egg-info/requires.txt

@@ -24,7 +24,7 @@ black~=25.12.0
 build~=1.3.0
 bumpver~=2025.1131
 captum~=0.7.0
-coverage~=7.13.0
+coverage~=7.13.1
 debugpy
 flake8-pep585~=0.1.7
 flake8~=7.3.0

{birder-0.2.0 → birder-0.2.1}/requirements/_requirements-dev.txt

@@ -4,7 +4,7 @@ black~=25.12.0
 build~=1.3.0
 bumpver~=2025.1131
 captum~=0.7.0
-coverage~=7.13.0
+coverage~=7.13.1
 debugpy
 flake8-pep585~=0.1.7
 flake8~=7.3.0

{birder-0.2.0 → birder-0.2.1}/tests/test_collators.py

@@ -29,3 +29,22 @@ class TestTransforms(unittest.TestCase):
 
         self.assertEqual(size_list[0], (10, 10))
         self.assertEqual(size_list[1], (12, 12))
+
+    def test_batch_random_resize_collator_scales_boxes(self) -> None:
+        collator = detection.BatchRandomResizeCollator(0, (32, 32))
+        collator.sizes = [20]
+
+        image = torch.zeros((3, 10, 20))
+        boxes = torch.tensor([[2.0, 1.0, 10.0, 5.0]], dtype=torch.float32)
+        labels = torch.tensor([1], dtype=torch.int64)
+        batch = [(image, {"boxes": boxes, "labels": labels})]
+
+        (images, targets, masks, size_list) = collator(batch)
+
+        # Collator pads to size_divisible=32
+        self.assertSequenceEqual(images.size(), (1, 3, 32, 32))
+        self.assertEqual(size_list[0], (20, 20))
+        self.assertTrue(torch.all(masks[:, :20, :20] == False))  # pylint: disable=singleton-comparison # noqa: E712
+
+        expected = torch.tensor([[2.0, 2.0, 10.0, 10.0]], dtype=torch.float32)
+        self.assertTrue(torch.allclose(targets[0]["boxes"], expected))

birder-0.2.0/birder/version.py

@@ -1 +0,0 @@
-__version__ = "v0.2.0"