birder 0.2.2__py3-none-any.whl → 0.2.3__py3-none-any.whl

birder/net/detection/yolo_anchors.py

@@ -0,0 +1,205 @@
+"""
+Shared YOLO anchor definitions and helpers.
+"""
+
+import json
+from collections.abc import Sequence
+from typing import Any
+from typing import Literal
+from typing import NotRequired
+from typing import TypedDict
+from typing import overload
+
+AnchorPair = tuple[float, float]
+AnchorGroup = list[AnchorPair]
+AnchorGroups = list[AnchorGroup]
+AnchorLike = AnchorGroups | AnchorGroup
+
+# Default anchors from yolo.cfg (COCO dataset), in grid units
+YOLO_V2_ANCHORS: AnchorGroup = [
+    (0.57273, 0.677385),
+    (1.87446, 2.06253),
+    (3.33843, 5.47434),
+    (7.88282, 3.52778),
+    (9.77052, 9.16828),
+]
+
+# Default anchors from YOLO v3 paper (sorted by area, small to large)
+# These values are in absolute pixels (width, height) computed using K-Means
+# on the COCO dataset with a reference input size of 416x416.
+YOLO_V3_ANCHORS: AnchorGroups = [
+    [(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)
+]
+
+# Default anchors from YOLO v4 (COCO), in pixels
+YOLO_V4_ANCHORS: AnchorGroups = [
+    [(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
+]
+
+# Default anchors from YOLO v4 Tiny (COCO), in pixels
+YOLO_V4_TINY_ANCHORS: AnchorGroups = [
+    [(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
+]
+
+
+class AnchorPreset(TypedDict):
+    anchors: AnchorLike
+    format: Literal["grid", "pixels"]
+    size: tuple[int, int]
+    strides: NotRequired[Sequence[int]]
+
+
+ANCHOR_PRESETS: dict[str, AnchorPreset] = {
+    "yolo_v2": {"anchors": YOLO_V2_ANCHORS, "format": "grid", "size": (416, 416), "strides": (32,)},
+    "yolo_v3": {"anchors": YOLO_V3_ANCHORS, "format": "pixels", "size": (416, 416)},
+    "yolo_v4": {"anchors": YOLO_V4_ANCHORS, "format": "pixels", "size": (608, 608)},
+    "yolo_v4_tiny": {"anchors": YOLO_V4_TINY_ANCHORS, "format": "pixels", "size": (416, 416)},
+}
+
+
+@overload
+def scale_anchors(anchors: AnchorGroup, from_size: tuple[int, int], to_size: tuple[int, int]) -> AnchorGroup: ...
+
+
+@overload
+def scale_anchors(anchors: AnchorGroups, from_size: tuple[int, int], to_size: tuple[int, int]) -> AnchorGroups: ...
+
+
+def scale_anchors(anchors: AnchorLike, from_size: tuple[int, int], to_size: tuple[int, int]) -> AnchorLike:
+    (anchor_groups, single) = _normalize_anchor_groups(anchors)
+
+    if from_size == to_size:
+        # Avoid aliasing default anchors in case they are mutated later
+        scaled: AnchorGroups = [list(group) for group in anchor_groups]
+        if single is True:
+            return scaled[0]
+
+        return scaled
+
+    scale_h = to_size[0] / from_size[0]
+    scale_w = to_size[1] / from_size[1]
+    scaled = [[(w * scale_w, h * scale_h) for (w, h) in group] for group in anchor_groups]
+
+    if single is True:
+        return scaled[0]
+
+    return scaled
+
+
+@overload
+def pixels_to_grid(anchors: AnchorGroup, strides: Sequence[int]) -> AnchorGroup: ...
+
+
+@overload
+def pixels_to_grid(anchors: AnchorGroups, strides: Sequence[int]) -> AnchorGroups: ...
+
+
+def pixels_to_grid(anchors: AnchorLike, strides: Sequence[int]) -> AnchorLike:
+    (anchor_groups, single) = _normalize_anchor_groups(anchors)
+    if len(anchor_groups) != len(strides):
+        raise ValueError("strides must provide one value per anchor scale")
+
+    converted: AnchorGroups = []
+    for group, stride in zip(anchor_groups, strides):
+        converted.append([(w / stride, h / stride) for (w, h) in group])
+
+    if single is True:
+        return converted[0]
+
+    return converted
+
+
+@overload
+def grid_to_pixels(anchors: AnchorGroup, strides: Sequence[int]) -> AnchorGroup: ...
+
+
+@overload
+def grid_to_pixels(anchors: AnchorGroups, strides: Sequence[int]) -> AnchorGroups: ...
+
+
+def grid_to_pixels(anchors: AnchorLike, strides: Sequence[int]) -> AnchorLike:
+    (anchor_groups, single) = _normalize_anchor_groups(anchors)
+    if len(anchor_groups) != len(strides):
+        raise ValueError("strides must provide one value per anchor scale")
+
+    converted: AnchorGroups = []
+    for group, stride in zip(anchor_groups, strides):
+        converted.append([(w * stride, h * stride) for (w, h) in group])
+
+    if single is True:
+        return converted[0]
+
+    return converted
+
+
+def _normalize_anchor_groups(anchors: AnchorLike) -> tuple[AnchorGroups, bool]:
+    if len(anchors) > 0 and _is_anchor_pair(anchors[0]) is True:
+        return ([anchors], True)  # type: ignore[list-item]
+
+    return (anchors, False)  # type: ignore[return-value]
+
+
+def _is_anchor_pair(value: Any) -> bool:
+    if not isinstance(value, Sequence) or len(value) != 2:
+        return False
+
+    return all(isinstance(item, (float, int)) for item in value)
+
+
+def _resolve_anchors(
+    preset: str, *, anchor_format: str, model_size: tuple[int, int], model_strides: Sequence[int]
+) -> AnchorLike:
+    if preset.endswith(".json") is True:
+        with open(preset, "r", encoding="utf-8") as handle:
+            preset_spec = json.load(handle)
+    else:
+        if preset not in ANCHOR_PRESETS:
+            raise ValueError(f"Unknown anchor preset: {preset}")
+
+        preset_spec = ANCHOR_PRESETS[preset]
+
+    anchors = preset_spec["anchors"]
+    preset_size = tuple(preset_spec["size"])
+    preset_format = preset_spec["format"]
+    if preset_format == "grid":
+        if "strides" not in preset_spec:
+            raise ValueError("Preset is missing strides required for grid anchors")
+
+        preset_strides = preset_spec["strides"]
+        anchors = grid_to_pixels(anchors, preset_strides)
+
+    anchors = scale_anchors(anchors, preset_size, model_size)
+    if anchor_format == "pixels":
+        return anchors
+
+    if anchor_format == "grid":
+        return pixels_to_grid(anchors, model_strides)
+
+    raise ValueError(f"Unsupported anchor format: {anchor_format}")
+
+
+def resolve_anchor_group(
+    preset: str, *, anchor_format: str, model_size: tuple[int, int], model_strides: Sequence[int]
+) -> AnchorGroup:
+    anchors = _resolve_anchors(preset, anchor_format=anchor_format, model_size=model_size, model_strides=model_strides)
+    (anchor_groups, single) = _normalize_anchor_groups(anchors)
+    if single is False:
+        raise ValueError("Expected a single anchor group for this model")
+
+    return anchor_groups[0]
+
+
+def resolve_anchor_groups(
+    preset: str, *, anchor_format: str, model_size: tuple[int, int], model_strides: Sequence[int]
+) -> AnchorGroups:
+    anchors = _resolve_anchors(preset, anchor_format=anchor_format, model_size=model_size, model_strides=model_strides)
+    (anchor_groups, single) = _normalize_anchor_groups(anchors)
+    if single is True:
+        raise ValueError("Expected multiple anchor groups for this model")
+
+    return anchor_groups

birder/net/detection/yolo_v2.py

@@ -17,18 +17,11 @@ from torch import nn
 from torchvision.ops import Conv2dNormActivation
 from torchvision.ops import boxes as box_ops
 
+from birder.model_registry import registry
 from birder.net.base import DetectorBackbone
 from birder.net.detection.base import DetectionBaseNet
 from birder.net.detection.base import ImageList
-
-# Default anchors from yolo.cfg (COCO dataset)
-DEFAULT_ANCHORS = [
-    (0.57273, 0.677385),
-    (1.87446, 2.06253),
-    (3.33843, 5.47434),
-    (7.88282, 3.52778),
-    (9.77052, 9.16828),
-]
+from birder.net.detection.yolo_anchors import resolve_anchor_group
 
 
 def decode_predictions(
@@ -102,8 +95,8 @@ def decode_predictions(
 class YOLOAnchorGenerator(nn.Module):
     def __init__(self, anchors: list[tuple[float, float]]) -> None:
         super().__init__()
-        self.anchors = anchors
-        self.num_anchors = len(anchors)
+        self.anchors = nn.Buffer(torch.tensor(anchors, dtype=torch.float32))
+        self.num_anchors: int = self.anchors.size(0)
 
     def num_anchors_per_location(self) -> int:
         return self.num_anchors
@@ -134,8 +127,7 @@ class YOLOAnchorGenerator(nn.Module):
         grid = torch.stack([grid_x, grid_y], dim=-1)
 
         # Scale anchors to feature map stride (anchors are in grid units)
-        anchors_tensor = torch.tensor(self.anchors, device=device, dtype=dtype)
-        anchors_tensor = anchors_tensor * torch.tensor([stride_w, stride_h], device=device, dtype=dtype)
+        anchors_tensor = self.anchors * torch.tensor([stride_w, stride_h], device=device, dtype=dtype)
 
         # Store strides as tensor
         stride = torch.tensor([stride_h, stride_w], device=device, dtype=dtype)
@@ -222,7 +214,6 @@ class YOLOHead(nn.Module):
 # pylint: disable=invalid-name
 class YOLO_v2(DetectionBaseNet):
     default_size = (416, 416)
-    auto_register = True
 
     def __init__(
         self,
@@ -234,7 +225,7 @@ class YOLO_v2(DetectionBaseNet):
         export_mode: bool = False,
     ) -> None:
         super().__init__(num_classes, backbone, config=config, size=size, export_mode=export_mode)
-        assert self.config is None, "config not supported"
+        assert self.config is not None, "must set config"
 
         self.num_classes = self.num_classes - 1
 
@@ -242,13 +233,19 @@ class YOLO_v2(DetectionBaseNet):
         nms_thresh = 0.45
         detections_per_img = 300
         mid_channels = 1024
-        self.ignore_thresh = 0.5
-        self.noobj_coeff = 0.5
-        self.coord_coeff = 5.0
-        self.obj_coeff = 1.0
-        self.cls_coeff = 1.0
+        ignore_thresh = 0.5
+        noobj_coeff = 0.5
+        coord_coeff = 5.0
+        obj_coeff = 1.0
+        cls_coeff = 1.0
+        anchor_spec = self.config["anchors"]
+
+        self.ignore_thresh = ignore_thresh
+        self.noobj_coeff = noobj_coeff
+        self.coord_coeff = coord_coeff
+        self.obj_coeff = obj_coeff
+        self.cls_coeff = cls_coeff
 
-        self.anchors = DEFAULT_ANCHORS
         self.score_thresh = score_thresh
         self.nms_thresh = nms_thresh
         self.detections_per_img = detections_per_img
@@ -258,7 +255,8 @@ class YOLO_v2(DetectionBaseNet):
 
         self.neck = YOLONeck(self.backbone.return_channels, mid_channels)
 
-        self.anchor_generator = YOLOAnchorGenerator(self.anchors)
+        anchors = resolve_anchor_group(anchor_spec, anchor_format="grid", model_size=self.size, model_strides=(32,))
+        self.anchor_generator = YOLOAnchorGenerator(anchors)
         num_anchors = self.anchor_generator.num_anchors_per_location()
         self.head = YOLOHead(self.neck.out_channels, num_anchors, self.num_classes)
 
@@ -319,7 +317,7 @@ class YOLO_v2(DetectionBaseNet):
         device = predictions.device
         dtype = predictions.dtype
         (batch_size, _, H, W) = predictions.size()
-        num_anchors = len(self.anchors)
+        num_anchors = self.anchor_generator.num_anchors
 
         stride_h = stride[0]
         stride_w = stride[1]
@@ -423,7 +421,7 @@ class YOLO_v2(DetectionBaseNet):
 
         device = predictions.device
         (N, _, H, W) = predictions.size()
-        num_anchors = len(self.anchors)
+        num_anchors = self.anchor_generator.num_anchors
 
         predictions = predictions.view(N, num_anchors, 5 + self.num_classes, H, W)
         predictions = predictions.permute(0, 1, 3, 4, 2).contiguous()
@@ -552,3 +550,6 @@ class YOLO_v2(DetectionBaseNet):
             detections = self.postprocess_detections(decoded_predictions, images.image_sizes)
 
         return (detections, losses)
+
+
+registry.register_model_config("yolo_v2", YOLO_v2, config={"anchors": "yolo_v2"})

birder/net/detection/yolo_v3.py

@@ -17,32 +17,11 @@ from torch import nn
 from torchvision.ops import Conv2dNormActivation
 from torchvision.ops import boxes as box_ops
 
+from birder.model_registry import registry
 from birder.net.base import DetectorBackbone
 from birder.net.detection.base import DetectionBaseNet
 from birder.net.detection.base import ImageList
-
-# Default anchors from YOLO v3 paper (sorted by area, small to large)
-# 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 = [
-    [(10.0, 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]
+from birder.net.detection.yolo_anchors import resolve_anchor_groups
 
 
 def decode_predictions(
@@ -116,11 +95,20 @@ def decode_predictions(
 class YOLOAnchorGenerator(nn.Module):
     def __init__(self, anchors: list[list[tuple[float, float]]]) -> None:
         super().__init__()
-        self.anchors = anchors
-        self.num_scales = len(anchors)
+        self.anchors = nn.Buffer(torch.tensor(anchors, dtype=torch.float32))
+        self.num_scales = self.anchors.size(0)
 
     def num_anchors_per_location(self) -> list[int]:
-        return [len(a) for a in self.anchors]
+        return [a.size(0) for a in self.anchors]
+
+    def scale_anchors(self, from_size: tuple[int, int], to_size: tuple[int, int]) -> None:
+        if from_size == to_size:
+            return
+
+        scale_h = to_size[0] / from_size[0]
+        scale_w = to_size[1] / from_size[1]
+        self.anchors[..., 0].mul_(scale_w)
+        self.anchors[..., 1].mul_(scale_h)
 
     def forward(
         self, image_list: ImageList, feature_maps: list[torch.Tensor]
@@ -152,7 +140,7 @@ class YOLOAnchorGenerator(nn.Module):
             grid = torch.stack([grid_x, grid_y], dim=-1)
 
             # Select anchors for this scale
-            anchors_for_scale = torch.tensor(self.anchors[idx], device=device, dtype=dtype)
+            anchors_for_scale = self.anchors[idx]
 
             # Store strides as tensor
             strides = torch.tensor([stride_h, stride_w], device=device, dtype=dtype)
@@ -321,7 +309,6 @@ class YOLONeck(nn.Module):
 # pylint: disable=invalid-name
 class YOLO_v3(DetectionBaseNet):
     default_size = (416, 416)
-    auto_register = True
 
     def __init__(
         self,
@@ -333,20 +320,26 @@ class YOLO_v3(DetectionBaseNet):
         export_mode: bool = False,
     ) -> None:
         super().__init__(num_classes, backbone, config=config, size=size, export_mode=export_mode)
-        assert self.config is None, "config not supported"
+        assert self.config is not None, "must set config"
 
         self.num_classes = self.num_classes - 1
 
         score_thresh = 0.05
         nms_thresh = 0.45
         detections_per_img = 300
-        self.ignore_thresh = 0.5
-        self.noobj_coeff = 0.2
-        self.coord_coeff = 5.0
-        self.obj_coeff = 1.0
-        self.cls_coeff = 1.0
+        ignore_thresh = 0.5
+        noobj_coeff = 0.2
+        coord_coeff = 5.0
+        obj_coeff = 1.0
+        cls_coeff = 1.0
+        anchor_spec = self.config["anchors"]
+
+        self.ignore_thresh = ignore_thresh
+        self.noobj_coeff = noobj_coeff
+        self.coord_coeff = coord_coeff
+        self.obj_coeff = obj_coeff
+        self.cls_coeff = cls_coeff
 
-        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
@@ -356,7 +349,10 @@ class YOLO_v3(DetectionBaseNet):
 
         self.neck = YOLONeck(self.backbone.return_channels)
 
-        self.anchor_generator = YOLOAnchorGenerator(self.anchors)
+        anchors = resolve_anchor_groups(
+            anchor_spec, anchor_format="pixels", model_size=self.size, model_strides=(8, 16, 32)
+        )
+        self.anchor_generator = YOLOAnchorGenerator(anchors)
         num_anchors = self.anchor_generator.num_anchors_per_location()
         self.head = YOLOHead(self.neck.out_channels, num_anchors, self.num_classes)
 
@@ -376,8 +372,7 @@ class YOLO_v3(DetectionBaseNet):
         super().adjust_size(new_size)
 
         if adjust_anchors is True:
-            self.anchors = scale_anchors(self.anchors, old_size, new_size)
-            self.anchor_generator.anchors = self.anchors
+            self.anchor_generator.scale_anchors(old_size, new_size)
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
@@ -435,7 +430,7 @@ class YOLO_v3(DetectionBaseNet):
 
         # Build flat list of all anchors with their scale indices
         all_anchors = torch.concat(anchors, dim=0)
-        anchors_per_scale = [len(self.anchors[i]) for i in range(num_scales)]
+        anchors_per_scale = self.anchor_generator.num_anchors_per_location()
         cumsum_anchors = torch.tensor([0] + anchors_per_scale, device=device).cumsum(0)
 
         # Get grid sizes and strides for each scale
@@ -586,6 +581,7 @@ class YOLO_v3(DetectionBaseNet):
         (target_tensors, obj_masks, noobj_masks) = self._build_targets(predictions, targets, anchors, strides)
 
         device = predictions[0].device
+        anchors_per_scale = self.anchor_generator.num_anchors_per_location()
         coord_loss = torch.tensor(0.0, device=device)
         obj_loss = torch.tensor(0.0, device=device)
         noobj_loss = torch.tensor(0.0, device=device)
@@ -594,7 +590,7 @@ class YOLO_v3(DetectionBaseNet):
         num_obj = 0
         for scale_idx, pred in enumerate(predictions):
             (N, _, H, W) = pred.size()
-            num_anchors_scale = len(self.anchors[scale_idx])
+            num_anchors_scale = anchors_per_scale[scale_idx]
 
             pred = pred.view(N, num_anchors_scale, 5 + self.num_classes, H, W)
             pred = pred.permute(0, 1, 3, 4, 2).contiguous()
@@ -730,3 +726,6 @@ class YOLO_v3(DetectionBaseNet):
             detections = self.postprocess_detections(decoded_predictions, images.image_sizes)
 
         return (detections, losses)
+
+
+registry.register_model_config("yolo_v3", YOLO_v3, config={"anchors": "yolo_v3"})

birder/net/detection/yolo_v4.py

@@ -17,18 +17,12 @@ from torch import nn
 from torchvision.ops import Conv2dNormActivation
 from torchvision.ops import boxes as box_ops
 
+from birder.model_registry import registry
 from birder.net.base import DetectorBackbone
 from birder.net.detection.base import DetectionBaseNet
+from birder.net.detection.yolo_anchors import resolve_anchor_groups
 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 = [
-    [(12.0, 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
 DEFAULT_SCALE_XY = [1.2, 1.1, 1.05]  # [small, medium, large]
@@ -59,7 +53,6 @@ def decode_predictions(
         Number of classes.
     scale_xy
         Scale factor for grid sensitivity elimination.
-        YOLOv4 uses 1.05-1.2 depending on scale level.
 
     Returns
     -------
@@ -378,7 +371,6 @@ class YOLONeck(nn.Module):
 # pylint: disable=invalid-name
 class YOLO_v4(DetectionBaseNet):
     default_size = (608, 608)
-    auto_register = True
 
     def __init__(
         self,
@@ -390,22 +382,26 @@ class YOLO_v4(DetectionBaseNet):
         export_mode: bool = False,
     ) -> None:
         super().__init__(num_classes, backbone, config=config, size=size, export_mode=export_mode)
-        assert self.config is None, "config not supported"
+        assert self.config is not None, "must set config"
 
         self.num_classes = self.num_classes - 1
 
         score_thresh = 0.05
         nms_thresh = 0.45
         detections_per_img = 300
-        self.ignore_thresh = 0.7
-
-        # Loss coefficients
-        self.noobj_coeff = 0.25
-        self.coord_coeff = 3.0
-        self.obj_coeff = 1.0
-        self.cls_coeff = 1.0
-
-        self.anchors = scale_anchors(DEFAULT_ANCHORS, self.default_size, self.size)
+        ignore_thresh = 0.7
+        noobj_coeff = 0.25
+        coord_coeff = 3.0
+        obj_coeff = 1.0
+        cls_coeff = 1.0
+        label_smoothing = 0.1
+        anchor_spec = self.config["anchors"]
+
+        self.ignore_thresh = ignore_thresh
+        self.noobj_coeff = noobj_coeff
+        self.coord_coeff = coord_coeff
+        self.obj_coeff = obj_coeff
+        self.cls_coeff = cls_coeff
         self.scale_xy = DEFAULT_SCALE_XY
         self.score_thresh = score_thresh
         self.nms_thresh = nms_thresh
@@ -414,13 +410,16 @@ class YOLO_v4(DetectionBaseNet):
         self.backbone.return_channels = self.backbone.return_channels[-3:]
         self.backbone.return_stages = self.backbone.return_stages[-3:]
 
-        self.label_smoothing = 0.1
+        self.label_smoothing = label_smoothing
         self.smooth_positive = 1.0 - self.label_smoothing
         self.smooth_negative = self.label_smoothing / self.num_classes
 
         self.neck = YOLONeck(self.backbone.return_channels)
 
-        self.anchor_generator = YOLOAnchorGenerator(self.anchors)
+        anchors = resolve_anchor_groups(
+            anchor_spec, anchor_format="pixels", model_size=self.size, model_strides=(8, 16, 32)
+        )
+        self.anchor_generator = YOLOAnchorGenerator(anchors)
         num_anchors = self.anchor_generator.num_anchors_per_location()
         self.head = YOLOHead(self.neck.out_channels, num_anchors, self.num_classes)
 
@@ -441,8 +440,7 @@ class YOLO_v4(DetectionBaseNet):
         super().adjust_size(new_size)
 
         if adjust_anchors is True:
-            self.anchors = scale_anchors(self.anchors, old_size, new_size)
-            self.anchor_generator = YOLOAnchorGenerator(self.anchors)
+            self.anchor_generator.scale_anchors(old_size, new_size)
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
@@ -500,7 +498,7 @@ class YOLO_v4(DetectionBaseNet):
 
         # Build flat list of all anchors with their scale indices
         all_anchors = torch.concat(anchors, dim=0)
-        anchors_per_scale = [len(self.anchors[i]) for i in range(num_scales)]
+        anchors_per_scale = self.anchor_generator.num_anchors_per_location()
         cumsum_anchors = torch.tensor([0] + anchors_per_scale, device=device).cumsum(0)
 
         # Get grid sizes and strides for each scale
@@ -651,6 +649,7 @@ class YOLO_v4(DetectionBaseNet):
         (target_tensors, obj_masks, noobj_masks) = self._build_targets(predictions, targets, anchors, strides)
 
         device = predictions[0].device
+        anchors_per_scale = self.anchor_generator.num_anchors_per_location()
         coord_loss = torch.tensor(0.0, device=device)
         obj_loss = torch.tensor(0.0, device=device)
         noobj_loss = torch.tensor(0.0, device=device)
@@ -659,7 +658,7 @@ class YOLO_v4(DetectionBaseNet):
         num_obj = 0
         for scale_idx, pred in enumerate(predictions):
             (N, _, H, W) = pred.size()
-            num_anchors_scale = len(self.anchors[scale_idx])
+            num_anchors_scale = anchors_per_scale[scale_idx]
             stride_h = strides[scale_idx][0]
             stride_w = strides[scale_idx][1]
             scale_xy = self.scale_xy[scale_idx]
@@ -829,3 +828,6 @@ class YOLO_v4(DetectionBaseNet):
             detections = self.postprocess_detections(decoded_predictions, images.image_sizes)
 
         return (detections, losses)
+
+
+registry.register_model_config("yolo_v4", YOLO_v4, config={"anchors": "yolo_v4"})