keras-hub-nightly 0.16.1.dev202410200345__py3-none-any.whl → 0.19.0.dev202412070351__py3-none-any.whl

keras_hub/src/models/pali_gemma/pali_gemma_presets.py

@@ -8,9 +8,7 @@ backbone_presets = {
                 "image size 224, mix fine tuned, text sequence " "length is 256"
             ),
             "params": 2923335408,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
         "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_224/3",
     },
@@ -20,9 +18,7 @@ backbone_presets = {
                 "image size 448, mix fine tuned, text sequence length is 512"
             ),
             "params": 2924220144,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
         "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_mix_448/3",
     },
@@ -32,9 +28,7 @@ backbone_presets = {
                 "image size 224, pre trained, text sequence length is 128"
             ),
             "params": 2923335408,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
         "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_224/3",
     },
@@ -44,9 +38,7 @@ backbone_presets = {
                 "image size 448, pre trained, text sequence length is 512"
             ),
             "params": 2924220144,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
         "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_448/3",
     },
@@ -56,10 +48,174 @@ backbone_presets = {
                 "image size 896, pre trained, text sequence length " "is 512"
             ),
             "params": 2927759088,
-            "official_name": "PaliGemma",
             "path": "pali_gemma",
-            "model_card": "https://www.kaggle.com/models/google/paligemma",
         },
         "kaggle_handle": "kaggle://keras/paligemma/keras/pali_gemma_3b_896/3",
     },
+    # PaliGemma2
+    "pali_gemma2_3b_ft_docci_448": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been fine-tuned on the DOCCI dataset "
+                "for improved descriptions with fine-grained details."
+            ),
+            "params": 3032979696,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_3b_ft_docci_448/1",
+    },
+    "pali_gemma2_10b_ft_docci_448": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, 27-layer for SigLIP-So400m vision "
+                "encoder and 42-layer Gemma2 9B lanuage model. This model has "
+                "been fine-tuned on the DOCCI dataset for improved "
+                "descriptions with fine-grained details."
+            ),
+            "params": 9663294192,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_10b_ft_docci_448/1",
+    },
+    "pali_gemma2_pt_3b_224": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 224, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 3032094960,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_3b_224/1",
+    },
+    "pali_gemma2_pt_3b_448": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 3032979696,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_3b_448/1",
+    },
+    "pali_gemma2_pt_3b_896": {
+        "metadata": {
+            "description": (
+                "3 billion parameter, image size 896, 27-layer for "
+                "SigLIP-So400m vision encoder and 26-layer Gemma2 2B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 3036518640,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_3b_896/1",
+    },
+    "pali_gemma2_pt_10b_224": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, image size 224, 27-layer for "
+                "SigLIP-So400m vision encoder and 42-layer Gemma2 9B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9662409456,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_224/1",
+    },
+    "pali_gemma2_pt_10b_448": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 42-layer Gemma2 9B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9663294192,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_448/1",
+    },
+    "pali_gemma2_pt_10b_896": {
+        "metadata": {
+            "description": (
+                "10 billion parameter, image size 896, 27-layer for "
+                "SigLIP-So400m vision encoder and 42-layer Gemma2 9B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9666833136,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_10b_896/1",
+    },
+    "pali_gemma2_pt_28b_224": {
+        "metadata": {
+            "description": (
+                "28 billion parameter, image size 224, 27-layer for "
+                "SigLIP-So400m vision encoder and 46-layer Gemma2 27B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9662409456,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_224/1",
+    },
+    "pali_gemma2_pt_28b_448": {
+        "metadata": {
+            "description": (
+                "28 billion parameter, image size 448, 27-layer for "
+                "SigLIP-So400m vision encoder and 46-layer Gemma2 27B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9663294192,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_448/1",
+    },
+    "pali_gemma2_pt_28b_896": {
+        "metadata": {
+            "description": (
+                "28 billion parameter, image size 896, 27-layer for "
+                "SigLIP-So400m vision encoder and 46-layer Gemma2 27B lanuage "
+                "model. This model has been pre-trained on a mixture of "
+                "datasets."
+            ),
+            "params": 9666833136,
+            "official_name": "PaliGemma2",
+            "path": "pali_gemma2",
+            "model_card": "https://www.kaggle.com/models/google/paligemma-2",
+        },
+        "kaggle_handle": "kaggle://keras/paligemma2/keras/pali_gemma2_pt_28b_896/1",
+    },
 }

keras_hub/src/models/pali_gemma/pali_gemma_vit.py

@@ -12,7 +12,7 @@ class PaliGemmaVitEmbeddings(keras.layers.Layer):
         dtype=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
+        super().__init__(dtype=dtype, **kwargs)
         self.hidden_dim = hidden_dim
         self.image_size = image_size
         self.patch_size = patch_size
@@ -72,7 +72,7 @@ class PaliGemmaVitAttention(keras.layers.Layer):
         dtype=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
+        super().__init__(dtype=dtype, **kwargs)
 
         self.hidden_dim = hidden_dim
         self.num_heads = num_heads
@@ -282,7 +282,7 @@ class PaliGemmaVitEncoder(keras.layers.Layer):
         dtype=None,
         **kwargs,
     ):
-        super().__init__(**kwargs)
+        super().__init__(dtype=dtype, **kwargs)
         self.hidden_dim = hidden_dim
         self.num_layers = num_layers
         self.num_heads = num_heads
@@ -311,25 +311,26 @@ class PaliGemmaVitEncoder(keras.layers.Layer):
             for i in range(self.num_layers)
         ]
 
-    def build(self, input_shape):
-        self.vision_embeddings.build(input_shape)
+    def build(self, inputs_shape):
+        self.vision_embeddings.build(inputs_shape)
         for block in self.resblocks:
             block.build([None, None, self.hidden_dim])
         self.encoder_layer_norm.build([None, None, self.hidden_dim])
         self.built = True
 
-    def call(
-        self,
-        x,
-        mask=None,
-    ):
-        x = self.vision_embeddings(x)
+    def call(self, inputs, mask=None):
+        x = self.vision_embeddings(inputs)
         for block in self.resblocks:
             x = block(x, mask=mask)
         x = self.encoder_layer_norm(x)
         return x
 
     def compute_output_shape(self, inputs_shape):
+        if inputs_shape is None:
+            # Fix the compatibility issue with Keras 3.1 where
+            # `compute_output_spec` fails to propagate `inputs_shape`
+            # correctly, causing it to be `None`.
+            inputs_shape = [None, None, None]
         return [inputs_shape[0], inputs_shape[1], self.hidden_dim]
 
     def get_config(self):

keras_hub/src/models/phi3/phi3_presets.py

@@ -12,9 +12,7 @@ backbone_presets = {
                 "reasoning-dense properties."
             ),
             "params": 3821079552,
-            "official_name": "Phi-3",
             "path": "phi3",
-            "model_card": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct",
         },
         "kaggle_handle": "kaggle://keras/phi3/keras/phi3_mini_4k_instruct_en",
     },
@@ -28,9 +26,7 @@ backbone_presets = {
                 "reasoning-dense properties."
             ),
             "params": 3821079552,
-            "official_name": "Phi-3",
             "path": "phi3",
-            "model_card": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct",
         },
         "kaggle_handle": "kaggle://keras/phi3/keras/phi3_mini_128k_instruct_en",
     },

keras_hub/src/models/resnet/resnet_presets.py

@@ -8,9 +8,7 @@ backbone_presets = {
                 "at a 224x224 resolution."
             ),
             "params": 11186112,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/2110.00476",
         },
         "kaggle_handle": "kaggle://keras/resnetv1/keras/resnet_18_imagenet/2",
     },
@@ -21,9 +19,7 @@ backbone_presets = {
                 "at a 224x224 resolution."
             ),
             "params": 23561152,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/2110.00476",
         },
         "kaggle_handle": "kaggle://keras/resnetv1/keras/resnet_50_imagenet/2",
     },
@@ -34,9 +30,7 @@ backbone_presets = {
                 "at a 224x224 resolution."
             ),
             "params": 42605504,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/2110.00476",
         },
         "kaggle_handle": "kaggle://keras/resnetv1/keras/resnet_101_imagenet/2",
     },
@@ -47,9 +41,7 @@ backbone_presets = {
                 "at a 224x224 resolution."
             ),
             "params": 58295232,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/2110.00476",
         },
         "kaggle_handle": "kaggle://keras/resnetv1/keras/resnet_152_imagenet/2",
     },
@@ -60,9 +52,7 @@ backbone_presets = {
                 "dataset at a 224x224 resolution."
             ),
             "params": 23561152,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/2110.00476",
         },
         "kaggle_handle": "kaggle://keras/resnetv2/keras/resnet_v2_50_imagenet/2",
     },
@@ -73,9 +63,7 @@ backbone_presets = {
                 "dataset at a 224x224 resolution."
             ),
             "params": 42605504,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/2110.00476",
         },
         "kaggle_handle": "kaggle://keras/resnetv2/keras/resnet_v2_101_imagenet/2",
     },
@@ -87,11 +75,9 @@ backbone_presets = {
                 "resolution."
             ),
             "params": 11722824,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_18_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_18_imagenet",
     },
     "resnet_vd_34_imagenet": {
         "metadata": {
@@ -101,11 +87,9 @@ backbone_presets = {
                 "resolution."
             ),
             "params": 21838408,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_34_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_34_imagenet",
     },
     "resnet_vd_50_imagenet": {
         "metadata": {
@@ -115,11 +99,9 @@ backbone_presets = {
                 "resolution."
             ),
             "params": 25629512,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_50_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_50_imagenet",
     },
     "resnet_vd_50_ssld_imagenet": {
         "metadata": {
@@ -129,11 +111,9 @@ backbone_presets = {
                 "resolution with knowledge distillation."
             ),
             "params": 25629512,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_50_ssld_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_50_ssld_imagenet",
     },
     "resnet_vd_50_ssld_v2_imagenet": {
         "metadata": {
@@ -143,11 +123,9 @@ backbone_presets = {
                 "resolution with knowledge distillation and AutoAugment."
             ),
             "params": 25629512,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_50_ssld_v2_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_50_ssld_v2_imagenet",
     },
     "resnet_vd_50_ssld_v2_fix_imagenet": {
         "metadata": {
@@ -158,11 +136,9 @@ backbone_presets = {
                 "additional fine-tuning of the classification head."
             ),
             "params": 25629512,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_50_ssld_v2_fix_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_50_ssld_v2_fix_imagenet",
     },
     "resnet_vd_101_imagenet": {
         "metadata": {
@@ -172,11 +148,9 @@ backbone_presets = {
                 "resolution."
             ),
             "params": 44673864,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_101_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_101_imagenet",
     },
     "resnet_vd_101_ssld_imagenet": {
         "metadata": {
@@ -186,11 +160,9 @@ backbone_presets = {
                 "resolution with knowledge distillation."
             ),
             "params": 44673864,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_101_ssld_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_101_ssld_imagenet",
     },
     "resnet_vd_152_imagenet": {
         "metadata": {
@@ -200,11 +172,9 @@ backbone_presets = {
                 "resolution."
             ),
             "params": 60363592,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_152_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_152_imagenet",
     },
     "resnet_vd_200_imagenet": {
         "metadata": {
@@ -214,10 +184,8 @@ backbone_presets = {
                 "resolution."
             ),
             "params": 74933064,
-            "official_name": "ResNet",
             "path": "resnet",
-            "model_card": "https://arxiv.org/abs/1812.01187",
         },
-        "kaggle_handle": "kaggle://kerashub/resnetvd/keras/resnet_vd_200_imagenet",
+        "kaggle_handle": "kaggle://keras/resnet_vd/keras/resnet_vd_200_imagenet",
     },
 }

keras_hub/src/models/retinanet/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.retinanet.retinanet_backbone import RetinaNetBackbone
+from keras_hub.src.models.retinanet.retinanet_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, RetinaNetBackbone)

keras_hub/src/models/retinanet/anchor_generator.py

@@ -3,9 +3,13 @@ import math
 import keras
 from keras import ops
 
+from keras_hub.src.api_export import keras_hub_export
+
+# TODO: https://github.com/keras-team/keras-hub/issues/1965
 from keras_hub.src.bounding_box.converters import convert_format
 
 
+@keras_hub_export("keras_hub.layers.AnchorGenerator")
 class AnchorGenerator(keras.layers.Layer):
     """Generates anchor boxes for object detection tasks.
 
@@ -81,6 +85,7 @@ class AnchorGenerator(keras.layers.Layer):
         self.num_scales = num_scales
         self.aspect_ratios = aspect_ratios
         self.anchor_size = anchor_size
+        self.num_base_anchors = num_scales * len(aspect_ratios)
         self.built = True
 
     def call(self, inputs):
@@ -92,60 +97,61 @@ class AnchorGenerator(keras.layers.Layer):
 
         image_shape = tuple(image_shape)
 
-        multilevel_boxes = {}
+        multilevel_anchors = {}
         for level in range(self.min_level, self.max_level + 1):
-            boxes_l = []
             # Calculate the feature map size for this level
             feat_size_y = math.ceil(image_shape[0] / 2**level)
             feat_size_x = math.ceil(image_shape[1] / 2**level)
 
             # Calculate the stride (step size) for this level
-            stride_y = ops.cast(image_shape[0] / feat_size_y, "float32")
-            stride_x = ops.cast(image_shape[1] / feat_size_x, "float32")
+            stride_y = image_shape[0] // feat_size_y
+            stride_x = image_shape[1] // feat_size_x
 
             # Generate anchor center points
             # Start from stride/2 to center anchors on pixels
-            cx = ops.arange(stride_x / 2, image_shape[1], stride_x)
-            cy = ops.arange(stride_y / 2, image_shape[0], stride_y)
+            cx = ops.arange(0, feat_size_x, dtype="float32") * stride_x
+            cy = ops.arange(0, feat_size_y, dtype="float32") * stride_y
 
             # Create a grid of anchor centers
-            cx_grid, cy_grid = ops.meshgrid(cx, cy)
-
-            for scale in range(self.num_scales):
-                for aspect_ratio in self.aspect_ratios:
-                    # Calculate the intermediate scale factor
-                    intermidate_scale = 2 ** (scale / self.num_scales)
-                    # Calculate the base anchor size for this level and scale
-                    base_anchor_size = (
-                        self.anchor_size * 2**level * intermidate_scale
-                    )
-                    # Adjust anchor dimensions based on aspect ratio
-                    aspect_x = aspect_ratio**0.5
-                    aspect_y = aspect_ratio**-0.5
-                    half_anchor_size_x = base_anchor_size * aspect_x / 2.0
-                    half_anchor_size_y = base_anchor_size * aspect_y / 2.0
-
-                    # Generate anchor boxes (y1, x1, y2, x2 format)
-                    boxes = ops.stack(
-                        [
-                            cy_grid - half_anchor_size_y,
-                            cx_grid - half_anchor_size_x,
-                            cy_grid + half_anchor_size_y,
-                            cx_grid + half_anchor_size_x,
-                        ],
-                        axis=-1,
-                    )
-                    boxes_l.append(boxes)
-            # Concat anchors on the same level to tensor shape HxWx(Ax4)
-            boxes_l = ops.concatenate(boxes_l, axis=-1)
-            boxes_l = ops.reshape(boxes_l, (-1, 4))
-            # Convert to user defined
-            multilevel_boxes[f"P{level}"] = convert_format(
-                boxes_l,
-                source="yxyx",
+            cy_grid, cx_grid = ops.meshgrid(cy, cx, indexing="ij")
+            cy_grid = ops.reshape(cy_grid, (-1,))
+            cx_grid = ops.reshape(cx_grid, (-1,))
+
+            shifts = ops.stack((cx_grid, cy_grid, cx_grid, cy_grid), axis=1)
+            sizes = [
+                int(
+                    2**level * self.anchor_size * 2 ** (scale / self.num_scales)
+                )
+                for scale in range(self.num_scales)
+            ]
+
+            base_anchors = self.generate_base_anchors(
+                sizes=sizes, aspect_ratios=self.aspect_ratios
+            )
+            shifts = ops.reshape(shifts, (-1, 1, 4))
+            base_anchors = ops.reshape(base_anchors, (1, -1, 4))
+
+            anchors = shifts + base_anchors
+            anchors = ops.reshape(anchors, (-1, 4))
+            multilevel_anchors[f"P{level}"] = convert_format(
+                anchors,
+                source="xyxy",
                 target=self.bounding_box_format,
             )
-        return multilevel_boxes
+        return multilevel_anchors
+
+    def generate_base_anchors(self, sizes, aspect_ratios):
+        sizes = ops.convert_to_tensor(sizes, dtype="float32")
+        aspect_ratios = ops.convert_to_tensor(aspect_ratios)
+        h_ratios = ops.sqrt(aspect_ratios)
+        w_ratios = 1 / h_ratios
+
+        ws = ops.reshape(w_ratios[:, None] * sizes[None, :], (-1,))
+        hs = ops.reshape(h_ratios[:, None] * sizes[None, :], (-1,))
+
+        base_anchors = ops.stack([-1 * ws, -1 * hs, ws, hs], axis=1) / 2
+        base_anchors = ops.round(base_anchors)
+        return base_anchors
 
     def compute_output_shape(self, input_shape):
         multilevel_boxes_shape = {}
@@ -156,18 +162,11 @@ class AnchorGenerator(keras.layers.Layer):
 
         for i in range(self.min_level, self.max_level + 1):
             multilevel_boxes_shape[f"P{i}"] = (
-                (image_height // 2 ** (i))
-                * (image_width // 2 ** (i))
-                * self.anchors_per_location,
+                int(
+                    math.ceil(image_height / 2 ** (i))
+                    * math.ceil(image_width // 2 ** (i))
+                    * self.num_base_anchors
+                ),
                 4,
             )
         return multilevel_boxes_shape
-
-    @property
-    def anchors_per_location(self):
-        """
-        The `anchors_per_location` property returns the number of anchors
-        generated per pixel location, which is equal to
-        `num_scales * len(aspect_ratios)`.
-        """
-        return self.num_scales * len(self.aspect_ratios)