tf-models-nightly 2.19.0.dev20250113__py2.py3-none-any.whl → 2.19.0.dev20250115__py2.py3-none-any.whl

official/vision/configs/retinanet.py

@@ -137,6 +137,7 @@ class RetinaNetHead(hyperparams.Config):
 @dataclasses.dataclass
 class DetectionGenerator(hyperparams.Config):
   apply_nms: bool = True
+  decode_boxes: bool = True
   pre_nms_top_k: int = 5000
   pre_nms_score_threshold: float = 0.05
   nms_iou_threshold: float = 0.5

official/vision/dataloaders/segmentation_input.py

@@ -157,9 +157,14 @@ class Parser(parser.Parser):
         dtype=tf.uint8,
     )
     image = tf.reshape(image, (height, width, self._image_feature.num_channels))
-    # Normalizes the image feature with mean and std values, which are divided
-    # by 255 because an uint8 image are re-scaled automatically. Images other
-    # than uint8 type will be wrongly normalized.
+    # Normalizes the image feature.
+    # The mean and stddev values are divided by 255 to ensure correct
+    # normalization, as the input `uint8` image is automatically converted to
+    # `float32` and rescaled to values in the range [0, 1] before the
+    # normalization happens (as a pre-processing step). So, we re-scale the
+    # mean and stddev values to the range [0, 1] beforehand.
+    # See `preprocess_ops.normalize_image` for details on the expected ranges
+    # for the image mean (`offset`) and stddev (`scale`).
     image = preprocess_ops.normalize_image(
         image,
         [mean / 255.0 for mean in self._image_feature.mean],

official/vision/ops/preprocess_ops.py

@@ -82,13 +82,28 @@ def normalize_image(
 ) -> tf.Tensor:
   """Normalizes the image to zero mean and unit variance.
 
-  If the input image dtype is float, it is expected to either have values in
-  [0, 1) and offset is MEAN_NORM, or have values in [0, 255] and offset is
-  MEAN_RGB.
+  This function normalizes the input image by subtracting the `offset`
+  and dividing by the `scale`.
+
+  **Important Note about Input Types and Normalization:**
+
+  * **Integer Images:** If the input `image` is an integer type (e.g., `uint8`),
+    the provided `offset` and `scale` values should be already **normalized**
+    to the range [0, 1]. This is because the function converts integer images to
+    float32 with values in the range [0, 1] before the normalization happens.
+
+  * **Float Images:** If the input `image` is a float type (e.g., `float32`),
+    the `offset` and `scale` values should be in the **same range** as the
+    image data.
+      - If the image has values in [0, 1], the `offset` and `scale` should
+        also be in [0, 1].
+      - If the image has values in [0, 255], the `offset` and `scale` should
+        also be in [0, 255].
 
   Args:
-    image: A tf.Tensor in either (1) float dtype with values in range [0, 1) or
-      [0, 255], or (2) int type with values in range [0, 255].
+    image: A `tf.Tensor` in either:
+           (1) float dtype with values in range [0, 1) or [0, 255], or
+           (2) int type with values in range [0, 255].
     offset: A tuple of mean values to be subtracted from the image.
     scale: A tuple of normalization factors.
 

official/vision/serving/detection.py

@@ -136,6 +136,14 @@ class DetectionModule(export_base.ExportModule):
 
     return detections_dict
 
+  def _flatten_output(self, feature_map, feature_size=4):
+    flatten_outputs = []
+    for level_output in feature_map.values():
+      flatten_outputs.append(
+          tf.reshape(level_output, (self._batch_size, -1, feature_size))
+      )
+    return tf.concat(flatten_outputs, axis=1)
+
   def preprocess(
       self, images: tf.Tensor
   ) -> Tuple[tf.Tensor, Mapping[str, tf.Tensor], tf.Tensor]:
@@ -271,6 +279,18 @@ class DetectionModule(export_base.ExportModule):
         final_outputs['detection_outer_boxes'] = detections[
             'detection_outer_boxes'
         ]
+    elif (
+        isinstance(self.params.task.model, configs.retinanet.RetinaNet)
+        and not self.params.task.model.detection_generator.decode_boxes
+    ):
+      final_outputs = {
+          'raw_boxes': self._flatten_output(detections['box_outputs'], 4),
+          'raw_scores': tf.sigmoid(
+              self._flatten_output(
+                  detections['cls_outputs'], self.params.task.model.num_classes
+              )
+          ),
+      }
     else:
       # For RetinaNet model, apply export_config.
       if isinstance(self.params.task.model, configs.retinanet.RetinaNet):

official/vision/serving/detection_test.py

@@ -38,6 +38,7 @@ class DetectionExportTest(tf.test.TestCase, parameterized.TestCase):
       normalized_coordinates=False,
       nms_version='batched',
       output_intermediate_features=False,
+      decode_boxes=True,
   ):
     params = exp_factory.get_exp_config(experiment_name)
     params.task.model.outer_boxes_scale = outer_boxes_scale
@@ -48,6 +49,8 @@ class DetectionExportTest(tf.test.TestCase, parameterized.TestCase):
     params.task.model.detection_generator.nms_version = nms_version
     if output_intermediate_features:
       params.task.export_config.output_intermediate_features = True
+    if not decode_boxes:
+      params.task.model.detection_generator.decode_boxes = False
     detection_module = detection.DetectionModule(
         params,
         batch_size=1,
@@ -232,6 +235,41 @@ class DetectionExportTest(tf.test.TestCase, parameterized.TestCase):
         max_values.numpy(), tf.ones_like(max_values).numpy()
     )
 
+  @parameterized.parameters(
+      'retinanet_mobile_coco',
+      'retinanet_spinenet_coco',
+  )
+  def test_export_without_decoding_boxes(
+      self,
+      experiment_name,
+  ):
+    input_type = 'tflite'
+    tmp_dir = self.get_temp_dir()
+    module = self._get_detection_module(
+        experiment_name,
+        input_type=input_type,
+        apply_nms=False,
+        decode_boxes=False,
+    )
+
+    self._export_from_module(module, input_type, tmp_dir)
+
+    imported = tf.saved_model.load(tmp_dir)
+    detection_fn = imported.signatures['serving_default']
+
+    images = self._get_dummy_input(
+        input_type, batch_size=1, image_size=(640, 640)
+    )
+    outputs = detection_fn(tf.constant(images))
+
+    self.assertContainsSubset(
+        {
+            'raw_boxes',
+            'raw_scores',
+        },
+        outputs.keys(),
+    )
+
 
 if __name__ == '__main__':
   tf.test.main()

{tf_models_nightly-2.19.0.dev20250113.dist-info → tf_models_nightly-2.19.0.dev20250115.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tf-models-nightly
-Version: 2.19.0.dev20250113
+Version: 2.19.0.dev20250115
 Summary: TensorFlow Official Models
 Home-page: https://github.com/tensorflow/models
 Author: Google Inc.

{tf_models_nightly-2.19.0.dev20250113.dist-info → tf_models_nightly-2.19.0.dev20250115.dist-info}/RECORD

@@ -1005,7 +1005,7 @@ official/vision/configs/image_classification.py,sha256=JiJUYKI82NtlNXEjATF0JzSmh
 official/vision/configs/image_classification_test.py,sha256=cdlkY5fPaqNZUc4-A2OOEUS1v3T-ekxqUc0Vm4IknBI,1867
 official/vision/configs/maskrcnn.py,sha256=yL8kggxXaCTIpSkcozAV2UudO7UqVcEh1_-rMd7oGDs,24236
 official/vision/configs/maskrcnn_test.py,sha256=Wfkbz30h2qxPcpuu6CEpQsf8I_2df6y10-4bRLsWlj8,1733
-official/vision/configs/retinanet.py,sha256=oCKinkh4IyPslmI1pakwi6dVziwjkZ2cIcpSoGRjqnM,17806
+official/vision/configs/retinanet.py,sha256=68GPsklbUThOlA4IeRM2gRBeoRMzNVX8wWYmTSDfz34,17834
 official/vision/configs/retinanet_test.py,sha256=ffS3QufQMLF8FZhKNmi7Yr1RDTnIyZ1XKQ9agr2EyW8,1699
 official/vision/configs/semantic_segmentation.py,sha256=xJbCWl46hTlUkw_S5UeKctdjuAZPTqE6UtSqQaLezYc,30808
 official/vision/configs/semantic_segmentation_test.py,sha256=va-ZG6CtBKcs0NicZe6WmJvHxPxxih7nB0orNtrRiEA,1867
@@ -1030,7 +1030,7 @@ official/vision/dataloaders/input_reader_factory.py,sha256=WpvSA8qyqAo3wkmme4WqX
 official/vision/dataloaders/maskrcnn_input.py,sha256=6QhSnKv7Q9qlKrQ6VnWsOxKsOq55MxBs9Rr47ZStUks,16803
 official/vision/dataloaders/parser.py,sha256=nMXnhigMa_ascSJ2OK88xi4HdE9xvfL3G4oMrHau-t4,2315
 official/vision/dataloaders/retinanet_input.py,sha256=0UgHSuhjxii0IpAODqc1eW3CIfMio6IUg5c5vCthFqg,18811
-official/vision/dataloaders/segmentation_input.py,sha256=Klg5KAChYZDRvqzZfyIzdPy54rTlWYZp2AotolD3WX8,12934
+official/vision/dataloaders/segmentation_input.py,sha256=PaoBn6QoONuIp-t5YChY2n_gRje-tEl5rSnfGI0b8Sc,13254
 official/vision/dataloaders/tf_example_decoder.py,sha256=9yCT6uSLMpmw50w7zdaRR_BXy6vIvliLZntrYAgzD18,8647
 official/vision/dataloaders/tf_example_decoder_test.py,sha256=iRPq780PvET64Ooa-1m5ahMp8lAObpC3l8ez8aATr1A,12610
 official/vision/dataloaders/tf_example_label_map_decoder.py,sha256=zquhN__lZv-S4rEo595Sveno64c-lNUxlf6-z1w7J3c,2579
@@ -1159,7 +1159,7 @@ official/vision/ops/iou_similarity_test.py,sha256=x5jlcMqMCUYC5cRgdbR0VlAW67AoXo
 official/vision/ops/mask_ops.py,sha256=X0hHlL0vUKl1Jt2LWbZDDulpS1CSJPSg-2VeqXb0oRM,10250
 official/vision/ops/mask_ops_test.py,sha256=leC0GTvdNeT0XyDnRmZwS1JSJ6NDMOpJcqNlUDpMwQI,2816
 official/vision/ops/nms.py,sha256=zUzJLuL-vk5feTK1MEOI49qmK8VxGVGl8GIMzN702yk,8106
-official/vision/ops/preprocess_ops.py,sha256=1NcE_iFPyzREdql0CRKGobFieCA6niSJcXV3Uqw635k,42588
+official/vision/ops/preprocess_ops.py,sha256=n6sgzbXJppln8z0q-uqdIMLNMP44VAEeyP9w8xDRUy0,43301
 official/vision/ops/preprocess_ops_3d.py,sha256=o0ivTBK2BwaiPp_NhVG_yBc10VUySxfE7eKQkL7RNaU,16762
 official/vision/ops/preprocess_ops_3d_test.py,sha256=LA3-Ue4FTjdsr5Kl_BnpAMNcfikWylMisD2GrBTOLzA,9309
 official/vision/ops/preprocess_ops_test.py,sha256=nuNRdVL9PETUDHic-2pm7PTTtr4KN79CoDw6_M8Zw9w,17638
@@ -1168,8 +1168,8 @@ official/vision/ops/spatial_transform_ops.py,sha256=PVEJGAn0ygtsrid84vD5GgV0jsjy
 official/vision/ops/target_gather.py,sha256=Ir3X76yXYEVFSYX5h-yfS8SMkY37GYuypBP2C8ykggo,3965
 official/vision/ops/target_gather_test.py,sha256=yiTZg7w4HYh19pw9yIDppvenscp8MSBGlhDko180l08,2561
 official/vision/serving/__init__.py,sha256=YlRnCfAvBWlB7gn-Dz32dxVIDBCzxH8PhdM14kHxPgs,702
-official/vision/serving/detection.py,sha256=33aMP4oYpWAro0YyD_PGx3i0CmztQ84PmSQTa3HFynU,11022
-official/vision/serving/detection_test.py,sha256=s3ueCplPII8Am7tPSWcUv9KUcAqh1AWb6X91_B6qZIM,9165
+official/vision/serving/detection.py,sha256=VvX3s6CqLlkveGla7EoWWnR7iDn1edJLTbtxhnuSUOs,11759
+official/vision/serving/detection_test.py,sha256=TZTVfJusX8Z4ycQSmp72e0yiOVhdRGVHMW2Lm2v-mYc,10119
 official/vision/serving/export_base.py,sha256=ah8Cuh_cxpptCpRAjptvA00a-yWgd-KokLk3UBLdVt4,7363
 official/vision/serving/export_base_v2.py,sha256=GYIqt-xaOv4UztAKjx-acD-9i2pjftMw46DWRMy7Bsk,2741
 official/vision/serving/export_base_v2_test.py,sha256=khuY14W9Oi8LGlv_CvMwFiNnFbgpPVzvvD_hugf1_lk,2880
@@ -1248,9 +1248,9 @@ tensorflow_models/tensorflow_models_test.py,sha256=nc6A9K53OGqF25xN5St8EiWvdVbda
 tensorflow_models/nlp/__init__.py,sha256=4tA5Pf4qaFwT-fIFOpX7x7FHJpnyJT-5UgOeFYTyMlc,807
 tensorflow_models/uplift/__init__.py,sha256=mqfa55gweOdpKoaQyid4A_4u7xw__FcQeSIF0k_pYmI,999
 tensorflow_models/vision/__init__.py,sha256=zBorY_v5xva1uI-qxhZO3Qh-Dii-Suq6wEYh6hKHDfc,833
-tf_models_nightly-2.19.0.dev20250113.dist-info/AUTHORS,sha256=1dG3fXVu9jlo7bul8xuix5F5vOnczMk7_yWn4y70uw0,337
-tf_models_nightly-2.19.0.dev20250113.dist-info/LICENSE,sha256=WxeBS_DejPZQabxtfMOM_xn8qoZNJDQjrT7z2wG1I4U,11512
-tf_models_nightly-2.19.0.dev20250113.dist-info/METADATA,sha256=98oAYc-UwtRS9Kwdkb2nDHskoLI7cbmVB9v1EzAU1-4,1432
-tf_models_nightly-2.19.0.dev20250113.dist-info/WHEEL,sha256=kGT74LWyRUZrL4VgLh6_g12IeVl_9u9ZVhadrgXZUEY,110
-tf_models_nightly-2.19.0.dev20250113.dist-info/top_level.txt,sha256=gum2FfO5R4cvjl2-QtP-S1aNmsvIZaFFT6VFzU0f4-g,33
-tf_models_nightly-2.19.0.dev20250113.dist-info/RECORD,,
+tf_models_nightly-2.19.0.dev20250115.dist-info/AUTHORS,sha256=1dG3fXVu9jlo7bul8xuix5F5vOnczMk7_yWn4y70uw0,337
+tf_models_nightly-2.19.0.dev20250115.dist-info/LICENSE,sha256=WxeBS_DejPZQabxtfMOM_xn8qoZNJDQjrT7z2wG1I4U,11512
+tf_models_nightly-2.19.0.dev20250115.dist-info/METADATA,sha256=7frOorbCz6svv8ljwlZWhIKX4BCUs19IRwJ6jBxPKbk,1432
+tf_models_nightly-2.19.0.dev20250115.dist-info/WHEEL,sha256=kGT74LWyRUZrL4VgLh6_g12IeVl_9u9ZVhadrgXZUEY,110
+tf_models_nightly-2.19.0.dev20250115.dist-info/top_level.txt,sha256=gum2FfO5R4cvjl2-QtP-S1aNmsvIZaFFT6VFzU0f4-g,33
+tf_models_nightly-2.19.0.dev20250115.dist-info/RECORD,,