keras-hub-nightly 0.19.0.dev202502060348__py3-none-any.whl → 0.19.0.dev202502080344__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- keras_hub/api/__init__.py +0 -1
- keras_hub/api/layers/__init__.py +3 -1
- keras_hub/api/models/__init__.py +10 -4
- keras_hub/src/{models/retinanet → layers/modeling}/anchor_generator.py +11 -18
- keras_hub/src/{models/retinanet → layers/modeling}/box_matcher.py +17 -4
- keras_hub/src/{models/retinanet → layers/modeling}/non_max_supression.py +84 -32
- keras_hub/src/layers/preprocessing/image_converter.py +25 -3
- keras_hub/src/models/{image_object_detector.py → object_detector.py} +12 -7
- keras_hub/src/models/{image_object_detector_preprocessor.py → object_detector_preprocessor.py} +29 -13
- keras_hub/src/models/retinanet/retinanet_image_converter.py +8 -40
- keras_hub/src/models/retinanet/retinanet_label_encoder.py +18 -16
- keras_hub/src/models/retinanet/retinanet_object_detector.py +28 -28
- keras_hub/src/models/retinanet/retinanet_object_detector_preprocessor.py +3 -3
- keras_hub/src/utils/tensor_utils.py +13 -0
- keras_hub/src/version_utils.py +1 -1
- {keras_hub_nightly-0.19.0.dev202502060348.dist-info → keras_hub_nightly-0.19.0.dev202502080344.dist-info}/METADATA +1 -1
- {keras_hub_nightly-0.19.0.dev202502060348.dist-info → keras_hub_nightly-0.19.0.dev202502080344.dist-info}/RECORD +19 -28
- keras_hub/api/bounding_box/__init__.py +0 -23
- keras_hub/src/bounding_box/__init__.py +0 -2
- keras_hub/src/bounding_box/converters.py +0 -606
- keras_hub/src/bounding_box/formats.py +0 -149
- keras_hub/src/bounding_box/iou.py +0 -251
- keras_hub/src/bounding_box/to_dense.py +0 -81
- keras_hub/src/bounding_box/to_ragged.py +0 -86
- keras_hub/src/bounding_box/utils.py +0 -181
- keras_hub/src/bounding_box/validate_format.py +0 -85
- {keras_hub_nightly-0.19.0.dev202502060348.dist-info → keras_hub_nightly-0.19.0.dev202502080344.dist-info}/WHEEL +0 -0
- {keras_hub_nightly-0.19.0.dev202502060348.dist-info → keras_hub_nightly-0.19.0.dev202502080344.dist-info}/top_level.txt +0 -0
keras_hub/api/__init__.py
CHANGED
keras_hub/api/layers/__init__.py
CHANGED
@@ -5,11 +5,14 @@ since your modifications would be overwritten.
|
|
5
5
|
"""
|
6
6
|
|
7
7
|
from keras_hub.src.layers.modeling.alibi_bias import AlibiBias
|
8
|
+
from keras_hub.src.layers.modeling.anchor_generator import AnchorGenerator
|
9
|
+
from keras_hub.src.layers.modeling.box_matcher import BoxMatcher
|
8
10
|
from keras_hub.src.layers.modeling.cached_multi_head_attention import (
|
9
11
|
CachedMultiHeadAttention,
|
10
12
|
)
|
11
13
|
from keras_hub.src.layers.modeling.f_net_encoder import FNetEncoder
|
12
14
|
from keras_hub.src.layers.modeling.masked_lm_head import MaskedLMHead
|
15
|
+
from keras_hub.src.layers.modeling.non_max_supression import NonMaxSuppression
|
13
16
|
from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
|
14
17
|
from keras_hub.src.layers.modeling.reversible_embedding import (
|
15
18
|
ReversibleEmbedding,
|
@@ -55,7 +58,6 @@ from keras_hub.src.models.pali_gemma.pali_gemma_image_converter import (
|
|
55
58
|
from keras_hub.src.models.resnet.resnet_image_converter import (
|
56
59
|
ResNetImageConverter,
|
57
60
|
)
|
58
|
-
from keras_hub.src.models.retinanet.anchor_generator import AnchorGenerator
|
59
61
|
from keras_hub.src.models.retinanet.retinanet_image_converter import (
|
60
62
|
RetinaNetImageConverter,
|
61
63
|
)
|
keras_hub/api/models/__init__.py
CHANGED
@@ -193,10 +193,6 @@ from keras_hub.src.models.image_classifier import ImageClassifier
|
|
193
193
|
from keras_hub.src.models.image_classifier_preprocessor import (
|
194
194
|
ImageClassifierPreprocessor,
|
195
195
|
)
|
196
|
-
from keras_hub.src.models.image_object_detector import ImageObjectDetector
|
197
|
-
from keras_hub.src.models.image_object_detector_preprocessor import (
|
198
|
-
ImageObjectDetectorPreprocessor,
|
199
|
-
)
|
200
196
|
from keras_hub.src.models.image_segmenter import ImageSegmenter
|
201
197
|
from keras_hub.src.models.image_segmenter_preprocessor import (
|
202
198
|
ImageSegmenterPreprocessor,
|
@@ -232,6 +228,16 @@ from keras_hub.src.models.mobilenet.mobilenet_backbone import MobileNetBackbone
|
|
232
228
|
from keras_hub.src.models.mobilenet.mobilenet_image_classifier import (
|
233
229
|
MobileNetImageClassifier,
|
234
230
|
)
|
231
|
+
from keras_hub.src.models.object_detector import ObjectDetector
|
232
|
+
from keras_hub.src.models.object_detector import (
|
233
|
+
ObjectDetector as ImageObjectDetector,
|
234
|
+
)
|
235
|
+
from keras_hub.src.models.object_detector_preprocessor import (
|
236
|
+
ObjectDetectorPreprocessor,
|
237
|
+
)
|
238
|
+
from keras_hub.src.models.object_detector_preprocessor import (
|
239
|
+
ObjectDetectorPreprocessor as ImageObjectDetectorPreprocessor,
|
240
|
+
)
|
235
241
|
from keras_hub.src.models.opt.opt_backbone import OPTBackbone
|
236
242
|
from keras_hub.src.models.opt.opt_causal_lm import OPTCausalLM
|
237
243
|
from keras_hub.src.models.opt.opt_causal_lm_preprocessor import (
|
@@ -4,9 +4,7 @@ import keras
|
|
4
4
|
from keras import ops
|
5
5
|
|
6
6
|
from keras_hub.src.api_export import keras_hub_export
|
7
|
-
|
8
|
-
# TODO: https://github.com/keras-team/keras-hub/issues/1965
|
9
|
-
from keras_hub.src.bounding_box.converters import convert_format
|
7
|
+
from keras_hub.src.utils.tensor_utils import assert_bounding_box_support
|
10
8
|
|
11
9
|
|
12
10
|
@keras_hub_export("keras_hub.layers.AnchorGenerator")
|
@@ -56,7 +54,7 @@ class AnchorGenerator(keras.layers.Layer):
|
|
56
54
|
|
57
55
|
Example:
|
58
56
|
```python
|
59
|
-
anchor_generator = AnchorGenerator(
|
57
|
+
anchor_generator = keras_hub.layers.AnchorGenerator(
|
60
58
|
bounding_box_format='xyxy',
|
61
59
|
min_level=3,
|
62
60
|
max_level=7,
|
@@ -78,6 +76,9 @@ class AnchorGenerator(keras.layers.Layer):
|
|
78
76
|
anchor_size,
|
79
77
|
**kwargs,
|
80
78
|
):
|
79
|
+
# Check whether current version of keras support bounding box utils
|
80
|
+
assert_bounding_box_support(self.__class__.__name__)
|
81
|
+
|
81
82
|
super().__init__(**kwargs)
|
82
83
|
self.bounding_box_format = bounding_box_format
|
83
84
|
self.min_level = min_level
|
@@ -94,29 +95,23 @@ class AnchorGenerator(keras.layers.Layer):
|
|
94
95
|
image_shape = images_shape[1:-1]
|
95
96
|
else:
|
96
97
|
image_shape = images_shape[:-1]
|
97
|
-
|
98
98
|
image_shape = tuple(image_shape)
|
99
|
-
|
100
99
|
multilevel_anchors = {}
|
101
100
|
for level in range(self.min_level, self.max_level + 1):
|
102
101
|
# Calculate the feature map size for this level
|
103
102
|
feat_size_y = math.ceil(image_shape[0] / 2**level)
|
104
103
|
feat_size_x = math.ceil(image_shape[1] / 2**level)
|
105
|
-
|
106
104
|
# Calculate the stride (step size) for this level
|
107
105
|
stride_y = image_shape[0] // feat_size_y
|
108
106
|
stride_x = image_shape[1] // feat_size_x
|
109
|
-
|
110
107
|
# Generate anchor center points
|
111
108
|
# Start from stride/2 to center anchors on pixels
|
112
109
|
cx = ops.arange(0, feat_size_x, dtype="float32") * stride_x
|
113
110
|
cy = ops.arange(0, feat_size_y, dtype="float32") * stride_y
|
114
|
-
|
115
111
|
# Create a grid of anchor centers
|
116
112
|
cy_grid, cx_grid = ops.meshgrid(cy, cx, indexing="ij")
|
117
113
|
cy_grid = ops.reshape(cy_grid, (-1,))
|
118
114
|
cx_grid = ops.reshape(cx_grid, (-1,))
|
119
|
-
|
120
115
|
shifts = ops.stack((cx_grid, cy_grid, cx_grid, cy_grid), axis=1)
|
121
116
|
sizes = [
|
122
117
|
int(
|
@@ -124,7 +119,6 @@ class AnchorGenerator(keras.layers.Layer):
|
|
124
119
|
)
|
125
120
|
for scale in range(self.num_scales)
|
126
121
|
]
|
127
|
-
|
128
122
|
base_anchors = self.generate_base_anchors(
|
129
123
|
sizes=sizes, aspect_ratios=self.aspect_ratios
|
130
124
|
)
|
@@ -133,10 +127,12 @@ class AnchorGenerator(keras.layers.Layer):
|
|
133
127
|
|
134
128
|
anchors = shifts + base_anchors
|
135
129
|
anchors = ops.reshape(anchors, (-1, 4))
|
136
|
-
multilevel_anchors[f"P{level}"] =
|
137
|
-
|
138
|
-
|
139
|
-
|
130
|
+
multilevel_anchors[f"P{level}"] = (
|
131
|
+
keras.utils.bounding_boxes.convert_format(
|
132
|
+
anchors,
|
133
|
+
source="xyxy",
|
134
|
+
target=self.bounding_box_format,
|
135
|
+
)
|
140
136
|
)
|
141
137
|
return multilevel_anchors
|
142
138
|
|
@@ -145,10 +141,8 @@ class AnchorGenerator(keras.layers.Layer):
|
|
145
141
|
aspect_ratios = ops.convert_to_tensor(aspect_ratios)
|
146
142
|
h_ratios = ops.sqrt(aspect_ratios)
|
147
143
|
w_ratios = 1 / h_ratios
|
148
|
-
|
149
144
|
ws = ops.reshape(w_ratios[:, None] * sizes[None, :], (-1,))
|
150
145
|
hs = ops.reshape(h_ratios[:, None] * sizes[None, :], (-1,))
|
151
|
-
|
152
146
|
base_anchors = ops.stack([-1 * ws, -1 * hs, ws, hs], axis=1) / 2
|
153
147
|
base_anchors = ops.round(base_anchors)
|
154
148
|
return base_anchors
|
@@ -159,7 +153,6 @@ class AnchorGenerator(keras.layers.Layer):
|
|
159
153
|
image_height, image_width = input_shape[1:-1]
|
160
154
|
else:
|
161
155
|
image_height, image_width = input_shape[:-1]
|
162
|
-
|
163
156
|
for i in range(self.min_level, self.max_level + 1):
|
164
157
|
multilevel_boxes_shape[f"P{i}"] = (
|
165
158
|
int(
|
@@ -1,7 +1,11 @@
|
|
1
1
|
import keras
|
2
2
|
from keras import ops
|
3
3
|
|
4
|
+
from keras_hub.src.api_export import keras_hub_export
|
5
|
+
from keras_hub.src.utils.tensor_utils import assert_bounding_box_support
|
4
6
|
|
7
|
+
|
8
|
+
@keras_hub_export("keras_hub.layers.BoxMatcher")
|
5
9
|
class BoxMatcher(keras.layers.Layer):
|
6
10
|
"""Box matching logic based on argmax of highest value (e.g., IOU).
|
7
11
|
|
@@ -51,10 +55,16 @@ class BoxMatcher(keras.layers.Layer):
|
|
51
55
|
|
52
56
|
Example:
|
53
57
|
```python
|
54
|
-
|
55
|
-
|
56
|
-
|
57
|
-
|
58
|
+
positive_threshold = 0.5
|
59
|
+
negative_threshold = 0.4
|
60
|
+
|
61
|
+
matcher = keras_hub.layers.BoxMatcher(
|
62
|
+
thresholds=[negative_threshold, positive_threshold],
|
63
|
+
match_values=[-1, -2, 1],
|
64
|
+
)
|
65
|
+
match_indices, matched_values = matcher(sim_matrix)
|
66
|
+
positive_mask = ops.equal(matched_vals, 1)
|
67
|
+
ignore_mask = ops.equal(matched_vals, -2)
|
58
68
|
```
|
59
69
|
|
60
70
|
"""
|
@@ -66,6 +76,9 @@ class BoxMatcher(keras.layers.Layer):
|
|
66
76
|
force_match_for_each_col=False,
|
67
77
|
**kwargs,
|
68
78
|
):
|
79
|
+
# Check whether current version of keras support bounding box utils
|
80
|
+
assert_bounding_box_support(self.__class__.__name__)
|
81
|
+
|
69
82
|
super().__init__(**kwargs)
|
70
83
|
if sorted(thresholds) != thresholds:
|
71
84
|
raise ValueError(f"`threshold` must be sorted, got {thresholds}")
|
@@ -3,32 +3,52 @@ import math
|
|
3
3
|
import keras
|
4
4
|
from keras import ops
|
5
5
|
|
6
|
-
|
7
|
-
from keras_hub.src.
|
8
|
-
from keras_hub.src.bounding_box import utils
|
9
|
-
from keras_hub.src.bounding_box import validate_format
|
6
|
+
from keras_hub.src.api_export import keras_hub_export
|
7
|
+
from keras_hub.src.utils.tensor_utils import assert_bounding_box_support
|
10
8
|
|
11
9
|
EPSILON = 1e-8
|
12
10
|
|
13
11
|
|
12
|
+
@keras_hub_export("keras_hub.layers.NonMaxSuppression")
|
14
13
|
class NonMaxSuppression(keras.layers.Layer):
|
15
14
|
"""A Keras layer that decodes predictions of an object detection model.
|
16
15
|
|
17
16
|
Args:
|
18
|
-
bounding_box_format: The format of bounding boxes of input dataset.
|
19
|
-
Refer
|
20
|
-
|
21
|
-
for more details on supported bounding box formats.
|
17
|
+
bounding_box_format: str. The format of bounding boxes of input dataset.
|
18
|
+
Refer `keras.utils.bounding_boxes.convert_format` args for more
|
19
|
+
details on supported bounding box formats.
|
22
20
|
from_logits: boolean, True means input score is logits, False means
|
23
21
|
confidence.
|
24
|
-
iou_threshold:
|
22
|
+
iou_threshold: float. Value in the range [0, 1] representing the
|
25
23
|
minimum IoU threshold for two boxes to be considered
|
26
24
|
same for suppression. Defaults to 0.5.
|
27
|
-
confidence_threshold:
|
25
|
+
confidence_threshold: float. Value in the range [0, 1]. All boxes with
|
28
26
|
confidence below this value will be discarded, defaults to 0.5.
|
29
|
-
max_detections: the maximum detections to consider after nms is
|
30
|
-
A large number may trigger significant memory overhead,
|
27
|
+
max_detections: int. the maximum detections to consider after nms is
|
28
|
+
applied. A large number may trigger significant memory overhead,
|
31
29
|
defaults to 100.
|
30
|
+
|
31
|
+
Example:
|
32
|
+
```python
|
33
|
+
boxes = np.random.uniform(low=0, high=1, size=(2, 5, 4))
|
34
|
+
classes = np.expand_dims(
|
35
|
+
np.array(
|
36
|
+
[[0.1, 0.1, 0.4, 0.5, 0.9], [0.7, 0.5, 0.3, 0.0, 0.0]],
|
37
|
+
"float32",
|
38
|
+
),
|
39
|
+
axis=-1,
|
40
|
+
)
|
41
|
+
|
42
|
+
nms = keras_hub.layers.NonMaxSuppression(
|
43
|
+
bounding_box_format="yxyx",
|
44
|
+
from_logits=False,
|
45
|
+
iou_threshold=1.0,
|
46
|
+
confidence_threshold=0.1,
|
47
|
+
max_detections=1,
|
48
|
+
)
|
49
|
+
|
50
|
+
nms_outputs = nms(boxes, classes)
|
51
|
+
```
|
32
52
|
"""
|
33
53
|
|
34
54
|
def __init__(
|
@@ -40,6 +60,8 @@ class NonMaxSuppression(keras.layers.Layer):
|
|
40
60
|
max_detections=100,
|
41
61
|
**kwargs,
|
42
62
|
):
|
63
|
+
# Check whether current version of keras support bounding box utils
|
64
|
+
assert_bounding_box_support(self.__class__.__name__)
|
43
65
|
super().__init__(**kwargs)
|
44
66
|
self.bounding_box_format = bounding_box_format
|
45
67
|
self.from_logits = from_logits
|
@@ -49,7 +71,10 @@ class NonMaxSuppression(keras.layers.Layer):
|
|
49
71
|
self.built = True
|
50
72
|
|
51
73
|
def call(
|
52
|
-
self,
|
74
|
+
self,
|
75
|
+
box_prediction,
|
76
|
+
class_prediction,
|
77
|
+
images=None,
|
53
78
|
):
|
54
79
|
"""Accepts images and raw scores, returning bounding box predictions.
|
55
80
|
|
@@ -59,15 +84,24 @@ class NonMaxSuppression(keras.layers.Layer):
|
|
59
84
|
class_prediction: Dense Tensor of shape [batch, boxes, num_classes].
|
60
85
|
"""
|
61
86
|
target_format = "yxyx"
|
62
|
-
|
63
|
-
|
87
|
+
height, width = None, None
|
88
|
+
|
89
|
+
if "rel" in self.bounding_box_format and images is None:
|
90
|
+
raise ValueError(
|
91
|
+
"`images` cannot be None when using relative "
|
92
|
+
"bounding box format."
|
93
|
+
)
|
64
94
|
|
65
|
-
|
95
|
+
if "rel" in self.bounding_box_format:
|
96
|
+
target_format = "rel_" + target_format
|
97
|
+
height, width, _ = ops.shape(images)
|
98
|
+
|
99
|
+
box_prediction = keras.utils.bounding_boxes.convert_format(
|
66
100
|
box_prediction,
|
67
101
|
source=self.bounding_box_format,
|
68
102
|
target=target_format,
|
69
|
-
|
70
|
-
|
103
|
+
height=height,
|
104
|
+
width=width,
|
71
105
|
)
|
72
106
|
if self.from_logits:
|
73
107
|
class_prediction = ops.sigmoid(class_prediction)
|
@@ -95,17 +129,17 @@ class NonMaxSuppression(keras.layers.Layer):
|
|
95
129
|
class_prediction, ops.expand_dims(idx, axis=-1), axis=1
|
96
130
|
)
|
97
131
|
|
98
|
-
box_prediction =
|
132
|
+
box_prediction = keras.utils.bounding_boxes.convert_format(
|
99
133
|
box_prediction,
|
100
134
|
source=target_format,
|
101
135
|
target=self.bounding_box_format,
|
102
|
-
|
103
|
-
|
136
|
+
height=height,
|
137
|
+
width=width,
|
104
138
|
)
|
105
139
|
bounding_boxes = {
|
106
140
|
"boxes": box_prediction,
|
107
141
|
"confidence": confidence_prediction,
|
108
|
-
"
|
142
|
+
"labels": ops.argmax(class_prediction, axis=-1),
|
109
143
|
"num_detections": valid_det,
|
110
144
|
}
|
111
145
|
|
@@ -519,22 +553,40 @@ def mask_invalid_detections(bounding_boxes):
|
|
519
553
|
returned value will also return `tf.RaggedTensor` representations.
|
520
554
|
"""
|
521
555
|
# ensure we are complying with Keras bounding box format.
|
522
|
-
|
523
|
-
|
556
|
+
if (
|
557
|
+
not isinstance(bounding_boxes, dict)
|
558
|
+
or "labels" not in bounding_boxes
|
559
|
+
or "boxes" not in bounding_boxes
|
560
|
+
):
|
524
561
|
raise ValueError(
|
525
|
-
"`
|
526
|
-
"
|
527
|
-
"
|
528
|
-
"to `bounding_box.mask_invalid_detections()`."
|
562
|
+
"Expected `bounding_boxes` agurment to be a "
|
563
|
+
"dict with keys 'boxes' and 'labels'. Received: "
|
564
|
+
f"bounding_boxes={bounding_boxes}"
|
529
565
|
)
|
566
|
+
|
530
567
|
if "num_detections" not in bounding_boxes:
|
531
568
|
raise ValueError(
|
532
569
|
"`bounding_boxes` must have key 'num_detections' "
|
533
|
-
"to be used with `
|
570
|
+
"to be used with `mask_invalid_detections()`."
|
534
571
|
)
|
535
572
|
|
536
573
|
boxes = bounding_boxes.get("boxes")
|
537
|
-
|
574
|
+
labels = bounding_boxes.get("labels")
|
575
|
+
if isinstance(boxes, list):
|
576
|
+
if not isinstance(labels, list):
|
577
|
+
raise ValueError(
|
578
|
+
"If `bounding_boxes['boxes']` is a list, then "
|
579
|
+
"`bounding_boxes['labels']` must also be a list."
|
580
|
+
f"Received: bounding_boxes['labels']={labels}"
|
581
|
+
)
|
582
|
+
if len(boxes) != len(labels):
|
583
|
+
raise ValueError(
|
584
|
+
"If `bounding_boxes['boxes']` and "
|
585
|
+
"`bounding_boxes['labels']` are both lists, "
|
586
|
+
"they must have the same length. Received: "
|
587
|
+
f"len(bounding_boxes['boxes'])={len(boxes)} and "
|
588
|
+
f"len(bounding_boxes['labels'])={len(labels)} and "
|
589
|
+
)
|
538
590
|
confidence = bounding_boxes.get("confidence", None)
|
539
591
|
num_detections = bounding_boxes.get("num_detections")
|
540
592
|
|
@@ -545,7 +597,7 @@ def mask_invalid_detections(bounding_boxes):
|
|
545
597
|
)
|
546
598
|
mask = mask < num_detections[:, None]
|
547
599
|
|
548
|
-
|
600
|
+
labels = ops.where(mask, labels, -ops.ones_like(labels))
|
549
601
|
|
550
602
|
if confidence is not None:
|
551
603
|
confidence = ops.where(mask, confidence, -ops.ones_like(confidence))
|
@@ -558,7 +610,7 @@ def mask_invalid_detections(bounding_boxes):
|
|
558
610
|
result = bounding_boxes.copy()
|
559
611
|
|
560
612
|
result["boxes"] = boxes
|
561
|
-
result["
|
613
|
+
result["labels"] = labels
|
562
614
|
if confidence is not None:
|
563
615
|
result["confidence"] = confidence
|
564
616
|
|
@@ -14,6 +14,7 @@ from keras_hub.src.utils.preset_utils import find_subclass
|
|
14
14
|
from keras_hub.src.utils.preset_utils import get_preset_loader
|
15
15
|
from keras_hub.src.utils.preset_utils import get_preset_saver
|
16
16
|
from keras_hub.src.utils.python_utils import classproperty
|
17
|
+
from keras_hub.src.utils.tensor_utils import check_bounding_box_support
|
17
18
|
from keras_hub.src.utils.tensor_utils import preprocessing_function
|
18
19
|
|
19
20
|
|
@@ -64,6 +65,12 @@ class ImageConverter(PreprocessingLayer):
|
|
64
65
|
interpolation: String, the interpolation method.
|
65
66
|
Supports `"bilinear"`, `"nearest"`, `"bicubic"`,
|
66
67
|
`"lanczos3"`, `"lanczos5"`. Defaults to `"bilinear"`.
|
68
|
+
bounding_box_format: A string specifying the format of the bounding
|
69
|
+
boxes, one of `"xyxy"`, `"rel_xyxy"`, `"xywh"`, `"center_xywh"`,
|
70
|
+
`"yxyx"`, `"rel_yxyx"`. Specifies the format of the bounding boxes
|
71
|
+
which will be resized to `image_size` along with the image. To pass
|
72
|
+
bounding boxed to this layer, pass a dict with keys `"images"` and
|
73
|
+
`"bounding_boxes"` when calling the layer.
|
67
74
|
data_format: String, either `"channels_last"` or `"channels_first"`.
|
68
75
|
The ordering of the dimensions in the inputs. `"channels_last"`
|
69
76
|
corresponds to inputs with shape `(batch, height, width, channels)`
|
@@ -100,6 +107,7 @@ class ImageConverter(PreprocessingLayer):
|
|
100
107
|
crop_to_aspect_ratio=True,
|
101
108
|
pad_to_aspect_ratio=False,
|
102
109
|
interpolation="bilinear",
|
110
|
+
bounding_box_format="yxyx",
|
103
111
|
data_format=None,
|
104
112
|
**kwargs,
|
105
113
|
):
|
@@ -121,6 +129,9 @@ class ImageConverter(PreprocessingLayer):
|
|
121
129
|
|
122
130
|
# Create the `Resizing` layer here even if it's not being used. That
|
123
131
|
# allows us to make `image_size` a settable property.
|
132
|
+
resizing_kwargs = {}
|
133
|
+
if check_bounding_box_support():
|
134
|
+
resizing_kwargs["bounding_box_format"] = bounding_box_format
|
124
135
|
self.resizing = keras.layers.Resizing(
|
125
136
|
height=image_size[0] if image_size else None,
|
126
137
|
width=image_size[1] if image_size else None,
|
@@ -130,12 +141,14 @@ class ImageConverter(PreprocessingLayer):
|
|
130
141
|
data_format=data_format,
|
131
142
|
dtype=self.dtype_policy,
|
132
143
|
name="resizing",
|
144
|
+
**resizing_kwargs,
|
133
145
|
)
|
134
146
|
self.scale = scale
|
135
147
|
self.offset = offset
|
136
148
|
self.crop_to_aspect_ratio = crop_to_aspect_ratio
|
137
149
|
self.pad_to_aspect_ratio = pad_to_aspect_ratio
|
138
150
|
self.interpolation = interpolation
|
151
|
+
self.bounding_box_format = bounding_box_format
|
139
152
|
self.data_format = standardize_data_format(data_format)
|
140
153
|
|
141
154
|
@property
|
@@ -154,14 +167,22 @@ class ImageConverter(PreprocessingLayer):
|
|
154
167
|
|
155
168
|
@preprocessing_function
|
156
169
|
def call(self, inputs):
|
157
|
-
x = inputs
|
158
170
|
if self.image_size is not None:
|
159
|
-
|
171
|
+
inputs = self.resizing(inputs)
|
172
|
+
# Allow dictionary input for handling bounding boxes.
|
173
|
+
if isinstance(inputs, dict):
|
174
|
+
x = inputs["images"]
|
175
|
+
else:
|
176
|
+
x = inputs
|
160
177
|
if self.scale is not None:
|
161
178
|
x = x * self._expand_non_channel_dims(self.scale, x)
|
162
179
|
if self.offset is not None:
|
163
180
|
x = x + self._expand_non_channel_dims(self.offset, x)
|
164
|
-
|
181
|
+
if isinstance(inputs, dict):
|
182
|
+
inputs["images"] = x
|
183
|
+
else:
|
184
|
+
inputs = x
|
185
|
+
return inputs
|
165
186
|
|
166
187
|
def _expand_non_channel_dims(self, value, inputs):
|
167
188
|
unbatched = len(ops.shape(inputs)) == 3
|
@@ -192,6 +213,7 @@ class ImageConverter(PreprocessingLayer):
|
|
192
213
|
"interpolation": self.interpolation,
|
193
214
|
"crop_to_aspect_ratio": self.crop_to_aspect_ratio,
|
194
215
|
"pad_to_aspect_ratio": self.pad_to_aspect_ratio,
|
216
|
+
"bounding_box_format": self.bounding_box_format,
|
195
217
|
}
|
196
218
|
)
|
197
219
|
return config
|
@@ -4,20 +4,25 @@ from keras_hub.src.api_export import keras_hub_export
|
|
4
4
|
from keras_hub.src.models.task import Task
|
5
5
|
|
6
6
|
|
7
|
-
@keras_hub_export(
|
8
|
-
|
7
|
+
@keras_hub_export(
|
8
|
+
[
|
9
|
+
"keras_hub.models.ObjectDetector",
|
10
|
+
"keras_hub.models.ImageObjectDetector",
|
11
|
+
]
|
12
|
+
)
|
13
|
+
class ObjectDetector(Task):
|
9
14
|
"""Base class for all image object detection tasks.
|
10
15
|
|
11
|
-
The `
|
16
|
+
The `ObjectDetector` tasks wrap a `keras_hub.models.Backbone` and
|
12
17
|
a `keras_hub.models.Preprocessor` to create a model that can be used for
|
13
|
-
object detection. `
|
18
|
+
object detection. `ObjectDetector` tasks take an additional
|
14
19
|
`num_classes` argument, controlling the number of predicted output classes.
|
15
20
|
|
16
21
|
To fine-tune with `fit()`, pass a dataset containing tuples of `(x, y)`
|
17
22
|
labels where `x` is a string and `y` is dictionary with `boxes` and
|
18
23
|
`classes`.
|
19
24
|
|
20
|
-
All `
|
25
|
+
All `ObjectDetector` tasks include a `from_preset()` constructor which
|
21
26
|
can be used to load a pre-trained config and weights.
|
22
27
|
"""
|
23
28
|
|
@@ -29,9 +34,9 @@ class ImageObjectDetector(Task):
|
|
29
34
|
metrics=None,
|
30
35
|
**kwargs,
|
31
36
|
):
|
32
|
-
"""Configures the `
|
37
|
+
"""Configures the `ObjectDetector` task for training.
|
33
38
|
|
34
|
-
The `
|
39
|
+
The `ObjectDetector` task extends the default compilation signature
|
35
40
|
of `keras.Model.compile` with defaults for `optimizer`, `loss`, and
|
36
41
|
`metrics`. To override these defaults, pass any value
|
37
42
|
to these arguments during compilation.
|
keras_hub/src/models/{image_object_detector_preprocessor.py → object_detector_preprocessor.py}
RENAMED
@@ -5,20 +5,25 @@ from keras_hub.src.models.preprocessor import Preprocessor
|
|
5
5
|
from keras_hub.src.utils.tensor_utils import preprocessing_function
|
6
6
|
|
7
7
|
|
8
|
-
@keras_hub_export(
|
9
|
-
|
8
|
+
@keras_hub_export(
|
9
|
+
[
|
10
|
+
"keras_hub.models.ObjectDetectorPreprocessor",
|
11
|
+
"keras_hub.models.ImageObjectDetectorPreprocessor",
|
12
|
+
]
|
13
|
+
)
|
14
|
+
class ObjectDetectorPreprocessor(Preprocessor):
|
10
15
|
"""Base class for object detector preprocessing layers.
|
11
16
|
|
12
|
-
`
|
17
|
+
`ObjectDetectorPreprocessor` tasks wraps a
|
13
18
|
`keras_hub.layers.Preprocessor` to create a preprocessing layer for
|
14
19
|
object detection tasks. It is intended to be paired with a
|
15
20
|
`keras_hub.models.ImageObjectDetector` task.
|
16
21
|
|
17
|
-
All `
|
22
|
+
All `ObjectDetectorPreprocessor` take three inputs, `x`, `y`, and
|
18
23
|
`sample_weight`. `x`, the first input, should always be included. It can
|
19
24
|
be a image or batch of images. See examples below. `y` and `sample_weight`
|
20
|
-
are optional inputs that will be passed through unaltered. Usually, `y`
|
21
|
-
be the a dict of `{"boxes": Tensor(batch_size, num_boxes, 4),
|
25
|
+
are optional inputs that will be passed through unaltered. Usually, `y`
|
26
|
+
will be the a dict of `{"boxes": Tensor(batch_size, num_boxes, 4),
|
22
27
|
"classes": (batch_size, num_boxes)}.
|
23
28
|
|
24
29
|
The layer will returns either `x`, an `(x, y)` tuple if labels were
|
@@ -26,18 +31,18 @@ class ImageObjectDetectorPreprocessor(Preprocessor):
|
|
26
31
|
were provided. `x` will be the input images after all model preprocessing
|
27
32
|
has been applied.
|
28
33
|
|
29
|
-
All `
|
30
|
-
constructor which can be used to load a pre-trained config and
|
31
|
-
You can call the `from_preset()` constructor directly on
|
32
|
-
which case the correct class for your model will be
|
33
|
-
instantiated.
|
34
|
+
All `ObjectDetectorPreprocessor` tasks include a `from_preset()`
|
35
|
+
constructor which can be used to load a pre-trained config and
|
36
|
+
vocabularies. You can call the `from_preset()` constructor directly on
|
37
|
+
this base class, in which case the correct class for your model will be
|
38
|
+
automatically instantiated.
|
34
39
|
|
35
40
|
Args:
|
36
41
|
image_converter: Preprocessing pipeline for images.
|
37
42
|
|
38
43
|
Examples.
|
39
44
|
```python
|
40
|
-
preprocessor = keras_hub.models.
|
45
|
+
preprocessor = keras_hub.models.ObjectDetectorPreprocessor.from_preset(
|
41
46
|
"retinanet_resnet50",
|
42
47
|
)
|
43
48
|
"""
|
@@ -52,6 +57,17 @@ class ImageObjectDetectorPreprocessor(Preprocessor):
|
|
52
57
|
|
53
58
|
@preprocessing_function
|
54
59
|
def call(self, x, y=None, sample_weight=None):
|
55
|
-
if
|
60
|
+
if y is None:
|
56
61
|
x = self.image_converter(x)
|
62
|
+
else:
|
63
|
+
# Pass bounding boxes through image converter in the dictionary
|
64
|
+
# with keys format standardized by core Keras.
|
65
|
+
output = self.image_converter(
|
66
|
+
{
|
67
|
+
"images": x,
|
68
|
+
"bounding_boxes": y,
|
69
|
+
}
|
70
|
+
)
|
71
|
+
x = output["images"]
|
72
|
+
y = output["bounding_boxes"]
|
57
73
|
return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
|
@@ -1,7 +1,6 @@
|
|
1
1
|
from keras_hub.src.api_export import keras_hub_export
|
2
2
|
from keras_hub.src.layers.preprocessing.image_converter import ImageConverter
|
3
3
|
from keras_hub.src.models.retinanet.retinanet_backbone import RetinaNetBackbone
|
4
|
-
from keras_hub.src.utils.tensor_utils import preprocessing_function
|
5
4
|
|
6
5
|
|
7
6
|
@keras_hub_export("keras_hub.layers.RetinaNetImageConverter")
|
@@ -10,44 +9,13 @@ class RetinaNetImageConverter(ImageConverter):
|
|
10
9
|
|
11
10
|
def __init__(
|
12
11
|
self,
|
13
|
-
|
14
|
-
scale=None,
|
15
|
-
offset=None,
|
16
|
-
norm_mean=[0.485, 0.456, 0.406],
|
17
|
-
norm_std=[0.229, 0.224, 0.225],
|
12
|
+
*args,
|
18
13
|
**kwargs,
|
19
14
|
):
|
20
|
-
|
21
|
-
|
22
|
-
|
23
|
-
|
24
|
-
|
25
|
-
|
26
|
-
|
27
|
-
|
28
|
-
@preprocessing_function
|
29
|
-
def call(self, inputs):
|
30
|
-
# TODO: https://github.com/keras-team/keras-hub/issues/1965
|
31
|
-
x = inputs
|
32
|
-
# Rescaling Image
|
33
|
-
if self.scale is not None:
|
34
|
-
x = x * self._expand_non_channel_dims(self.scale, x)
|
35
|
-
if self.offset is not None:
|
36
|
-
x = x + self._expand_non_channel_dims(self.offset, x)
|
37
|
-
# By default normalize using imagenet mean and std
|
38
|
-
if self.norm_mean:
|
39
|
-
x = x - self._expand_non_channel_dims(self.norm_mean, x)
|
40
|
-
if self.norm_std:
|
41
|
-
x = x / self._expand_non_channel_dims(self.norm_std, x)
|
42
|
-
|
43
|
-
return x
|
44
|
-
|
45
|
-
def get_config(self):
|
46
|
-
config = super().get_config()
|
47
|
-
config.update(
|
48
|
-
{
|
49
|
-
"norm_mean": self.norm_mean,
|
50
|
-
"norm_std": self.norm_std,
|
51
|
-
}
|
52
|
-
)
|
53
|
-
return config
|
15
|
+
# TODO: update presets and remove these old config options. They were
|
16
|
+
# never needed.
|
17
|
+
if "norm_mean" in kwargs:
|
18
|
+
kwargs["offset"] = [-x for x in kwargs.pop("norm_mean")]
|
19
|
+
if "norm_std" in kwargs:
|
20
|
+
kwargs["scale"] = [1.0 / x for x in kwargs.pop("norm_std")]
|
21
|
+
super().__init__(*args, **kwargs)
|