tf-models-nightly 2.11.0.dev20230322__py2.py3-none-any.whl → 2.11.0.dev20230323__py2.py3-none-any.whl

official/modeling/hyperparams/base_config.py

@@ -95,6 +95,20 @@ class Config(params_dict.ParamsDict):
   def BUILDER(self):
     return self._BUILDER
 
+  @classmethod
+  def _get_annotations(cls):
+    """Returns valid annotations.
+
+    Note: this is similar to dataclasses.__annotations__ except it also includes
+      annotations from its parent classes.
+    """
+    all_annotations = typing.get_type_hints(cls)
+    # Removes Config class annotation from the value, e.g., default_params,
+    # restrictions, etc.
+    for k in Config.__annotations__:
+      del all_annotations[k]
+    return all_annotations
+
   @classmethod
   def _isvalidsequence(cls, v):
     """Check if the input values are valid sequences.
@@ -175,9 +189,10 @@ class Config(params_dict.ParamsDict):
     if not subconfig_type:
       subconfig_type = Config
 
-    if k in cls.__annotations__:
+    annotations = cls._get_annotations()
+    if k in annotations:
       # Directly Config subtype.
-      type_annotation = cls.__annotations__[k]  # pytype: disable=invalid-annotation
+      type_annotation = annotations[k]
       i = 0
       # Loop for striping the Optional annotation.
       traverse_in = True
@@ -326,6 +341,9 @@ class Config(params_dict.ParamsDict):
   @classmethod
   def from_args(cls, *args, **kwargs):
     """Builds a config from the given list of arguments."""
+    # Note we intend to keep `__annotations__` instead of `_get_annotations`.
+    # Assuming a parent class of (a, b) with the sub-class of (c, d), the
+    # sub-class will take (c, d) for args, rather than starting from (a, b).
     attributes = list(cls.__annotations__.keys())
     default_params = {a: p for a, p in zip(attributes, args)}
     default_params.update(kwargs)

official/modeling/hyperparams/base_config_test.py

@@ -33,6 +33,7 @@ class DumpConfig2(base_config.Config):
   c: int = 2
   d: str = 'text'
   e: DumpConfig1 = DumpConfig1()
+  optional_e: Optional[DumpConfig1] = None
 
 
 @dataclasses.dataclass
@@ -348,6 +349,34 @@ class BaseConfigTest(parameterized.TestCase, tf.test.TestCase):
         ]),
         "['s', 1, 1.0, True, None, {}, [], (), {8: 9, (2,): (3, [4], {6: 7})}]")
 
+  def test_with_superclass_override(self):
+    config = DumpConfig2()
+    config.override({'optional_e': {'a': 2}})
+    self.assertEqual(
+        config.optional_e.as_dict(),
+        {
+            'a': 2,
+            'b': 'text',
+        },
+    )
+
+    # Previously, the following will fail. See b/274696969 for context.
+    config = DumpConfig3()
+    config.override({'optional_e': {'a': 2}})
+    self.assertEqual(
+        config.optional_e.as_dict(),
+        {
+            'a': 2,
+            'b': 'text',
+        },
+    )
+
+  def test_get_annotations_without_base_config_leak(self):
+    with self.assertRaisesRegex(
+        KeyError, "The key 'restrictions' does not exist"
+    ):
+      DumpConfig3().override({'restrictions': None})
+
   def test_with_restrictions(self):
     restrictions = ['e.a<c']
     config = DumpConfig2(restrictions=restrictions)

official/vision/ops/augment.py

@@ -389,82 +389,94 @@ def _fill_rectangle_video(image,
   return image
 
 
-def cutout_video(image: tf.Tensor, replace: int = 0) -> tf.Tensor:
+def cutout_video(video: tf.Tensor, replace: int = 0) -> tf.Tensor:
   """Apply cutout (https://arxiv.org/abs/1708.04552) to a video.
 
   This operation applies a random size 3D mask of zeros to a random location
-  within `image`. The mask is padded The pixel values filled in will be of the
+  within `video`. The mask is padded The pixel values filled in will be of the
   value `replace`. The location where the mask will be applied is randomly
-  chosen uniformly over the whole image. The size of the mask is randomly
+  chosen uniformly over the whole video. The size of the mask is randomly
   sampled uniformly from [0.25*height, 0.5*height], [0.25*width, 0.5*width],
   and [1, 0.25*depth], which represent the height, width, and number of frames
   of the input video tensor respectively.
 
   Args:
-    image: A video Tensor of type uint8.
+    video: A video Tensor of shape [T, H, W, C].
     replace: What pixel value to fill in the image in the area that has the
       cutout mask applied to it.
 
   Returns:
-    An video Tensor that is of type uint8.
+    A video Tensor with cutout applied.
   """
-  image_depth = tf.shape(image)[0]
-  image_height = tf.shape(image)[1]
-  image_width = tf.shape(image)[2]
+  tf.debugging.assert_shapes([
+      (video, ('T', 'H', 'W', 'C')),
+  ])
+
+  video_depth = tf.shape(video)[0]
+  video_height = tf.shape(video)[1]
+  video_width = tf.shape(video)[2]
 
   # Sample the center location in the image where the zero mask will be applied.
   cutout_center_height = tf.random.uniform(
-      shape=[], minval=0, maxval=image_height, dtype=tf.int32)
+      shape=[], minval=0, maxval=video_height, dtype=tf.int32
+  )
 
   cutout_center_width = tf.random.uniform(
-      shape=[], minval=0, maxval=image_width, dtype=tf.int32)
+      shape=[], minval=0, maxval=video_width, dtype=tf.int32
+  )
 
   cutout_center_depth = tf.random.uniform(
-      shape=[], minval=0, maxval=image_depth, dtype=tf.int32)
+      shape=[], minval=0, maxval=video_depth, dtype=tf.int32
+  )
 
   pad_size_height = tf.random.uniform(
       shape=[],
-      minval=tf.maximum(1, tf.cast(image_height / 4, tf.int32)),
-      maxval=tf.maximum(2, tf.cast(image_height / 2, tf.int32)),
-      dtype=tf.int32)
+      minval=tf.maximum(1, tf.cast(video_height / 4, tf.int32)),
+      maxval=tf.maximum(2, tf.cast(video_height / 2, tf.int32)),
+      dtype=tf.int32,
+  )
   pad_size_width = tf.random.uniform(
       shape=[],
-      minval=tf.maximum(1, tf.cast(image_width / 4, tf.int32)),
-      maxval=tf.maximum(2, tf.cast(image_width / 2, tf.int32)),
-      dtype=tf.int32)
+      minval=tf.maximum(1, tf.cast(video_width / 4, tf.int32)),
+      maxval=tf.maximum(2, tf.cast(video_width / 2, tf.int32)),
+      dtype=tf.int32,
+  )
   pad_size_depth = tf.random.uniform(
       shape=[],
       minval=1,
-      maxval=tf.maximum(2, tf.cast(image_depth / 4, tf.int32)),
-      dtype=tf.int32)
+      maxval=tf.maximum(2, tf.cast(video_depth / 4, tf.int32)),
+      dtype=tf.int32,
+  )
 
   lower_pad = tf.maximum(0, cutout_center_height - pad_size_height)
   upper_pad = tf.maximum(
-      0, image_height - cutout_center_height - pad_size_height)
+      0, video_height - cutout_center_height - pad_size_height
+  )
   left_pad = tf.maximum(0, cutout_center_width - pad_size_width)
-  right_pad = tf.maximum(0, image_width - cutout_center_width - pad_size_width)
+  right_pad = tf.maximum(0, video_width - cutout_center_width - pad_size_width)
   back_pad = tf.maximum(0, cutout_center_depth - pad_size_depth)
   forward_pad = tf.maximum(
-      0, image_depth - cutout_center_depth - pad_size_depth)
+      0, video_depth - cutout_center_depth - pad_size_depth
+  )
 
   cutout_shape = [
-      image_depth - (back_pad + forward_pad),
-      image_height - (lower_pad + upper_pad),
-      image_width - (left_pad + right_pad),
+      video_depth - (back_pad + forward_pad),
+      video_height - (lower_pad + upper_pad),
+      video_width - (left_pad + right_pad),
   ]
   padding_dims = [[back_pad, forward_pad],
                   [lower_pad, upper_pad],
                   [left_pad, right_pad]]
   mask = tf.pad(
-      tf.zeros(cutout_shape, dtype=image.dtype),
-      padding_dims,
-      constant_values=1)
+      tf.zeros(cutout_shape, dtype=video.dtype), padding_dims, constant_values=1
+  )
   mask = tf.expand_dims(mask, -1)
-  mask = tf.tile(mask, [1, 1, 1, 3])
-  image = tf.where(
-      tf.equal(mask, 0),
-      tf.ones_like(image, dtype=image.dtype) * replace, image)
-  return image
+  num_channels = tf.shape(video)[-1]
+  mask = tf.tile(mask, [1, 1, 1, num_channels])
+  video = tf.where(
+      tf.equal(mask, 0), tf.ones_like(video, dtype=video.dtype) * replace, video
+  )
+  return video
 
 
 def gaussian_noise(

official/vision/ops/augment_test.py

@@ -21,6 +21,7 @@ from __future__ import print_function
 import random
 from absl.testing import parameterized
 
+import numpy as np
 import tensorflow as tf
 
 from official.vision.ops import augment
@@ -85,6 +86,14 @@ class TransformsTest(parameterized.TestCase, tf.test.TestCase):
       image = tf.zeros(shape, dtype=dtype)
       self.assertAllEqual(image, augment.rotate(image, degrees))
 
+  def test_cutout_video(self, dtype):
+    for num_channels in (1, 2, 3):
+      video = tf.ones((2, 2, 2, num_channels), dtype=dtype)
+      video = augment.cutout_video(video)
+
+      num_zeros = np.sum(video == 0)
+      self.assertGreater(num_zeros, 0)
+
 
 class AutoaugmentTest(tf.test.TestCase, parameterized.TestCase):
 

official/vision/serving/export_tflite_lib.py

@@ -85,16 +85,21 @@ def representative_dataset(
     yield [image]
 
 
-def convert_tflite_model(saved_model_dir: str,
-                         quant_type: Optional[str] = None,
-                         params: Optional[cfg.ExperimentConfig] = None,
-                         task: Optional[base_task.Task] = None,
-                         calibration_steps: Optional[int] = 2000,
-                         denylisted_ops: Optional[List[str]] = None) -> 'bytes':
+def convert_tflite_model(
+    saved_model_dir: Optional[str] = None,
+    model: Optional[tf.keras.Model] = None,
+    quant_type: Optional[str] = None,
+    params: Optional[cfg.ExperimentConfig] = None,
+    task: Optional[base_task.Task] = None,
+    calibration_steps: Optional[int] = 2000,
+    denylisted_ops: Optional[List[str]] = None,
+) -> 'bytes':
   """Converts and returns a TFLite model.
 
   Args:
     saved_model_dir: The directory to the SavedModel.
+    model: An optional tf.keras.Model instance. If `saved_model_dir` is not
+      available, convert this model to TFLite.
     quant_type: The post training quantization (PTQ) method. It can be one of
       `default` (dynamic range), `fp16` (float16), `int8` (integer wih float
       fallback), `int8_full` (integer only) and None (no quantization).
@@ -111,9 +116,16 @@ def convert_tflite_model(saved_model_dir: str,
 
   Raises:
     ValueError: If `representative_dataset_path` is not present if integer
-    quantization is requested.
+      quantization is requested, or both `saved_model_dir` or `model` are not
+      provided.
   """
-  converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
+  if saved_model_dir:
+    converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
+  elif model is not None:
+    converter = tf.lite.TFLiteConverter.from_keras_model(model)
+  else:
+    raise ValueError('Either `saved_model_dir` or `model` must be specified.')
+
   if quant_type:
     if quant_type.startswith('int8'):
       converter.optimizations = [tf.lite.Optimize.DEFAULT]

{tf_models_nightly-2.11.0.dev20230322.dist-info → tf_models_nightly-2.11.0.dev20230323.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tf-models-nightly
-Version: 2.11.0.dev20230322
+Version: 2.11.0.dev20230323
 Summary: TensorFlow Official Models
 Home-page: https://github.com/tensorflow/models
 Author: Google Inc.

{tf_models_nightly-2.11.0.dev20230322.dist-info → tf_models_nightly-2.11.0.dev20230323.dist-info}/RECORD

@@ -205,8 +205,8 @@ official/modeling/activations/sigmoid_test.py,sha256=yl_7jCvV__5bPAbkvyQJcCnDGW6
 official/modeling/activations/swish.py,sha256=NWUn8lzdNMdX78oe5kylXYjMEmrIeSwuKWKpkBhXwG4,2291
 official/modeling/activations/swish_test.py,sha256=MHf5CL6mdKMiODt-i_m15fMREHhPshVdsWe9A2v8uS0,1411
 official/modeling/hyperparams/__init__.py,sha256=llk_04oD2rd3nnWRmsQAQ19l6OBKMPhQcjh_7RxN4wM,846
-official/modeling/hyperparams/base_config.py,sha256=VqNUjpf6kgt1vLLzU33_STyV_dZXJHYBvSaP9IYknq0,11838
-official/modeling/hyperparams/base_config_test.py,sha256=4jCujgFhbxQUPvhOOnWa9HLysYuT58VyqRbMeFJn7iY,11814
+official/modeling/hyperparams/base_config.py,sha256=k6LSNEJOpGg0int5gGR-6SrU9ibs1w8iLhsHMNeNpvY,12509
+official/modeling/hyperparams/base_config_test.py,sha256=7IAWD_a1KjRUp6T9P8mDCCPEwHcM9WNeeYPhMrDEP2I,12593
 official/modeling/hyperparams/oneof.py,sha256=N_KhdFtTF1xvciQvsiNkh5q3nVcqE0S0dtqITK2GLug,1870
 official/modeling/hyperparams/oneof_test.py,sha256=1UAEMkGi6CFz_lKal60UbH7tzdz3096PYz123Th_pbg,1882
 official/modeling/hyperparams/params_dict.py,sha256=gyP6R3OYMNQcE9FyNU2GVwqCMLDcGzMiiSBhc_wwYS0,16412
@@ -971,8 +971,8 @@ official/vision/ops/anchor.py,sha256=ETLOiNx3CPbsj1yLqgMXU-YXPi6zSwXF6yP6RbEpQeM
 official/vision/ops/anchor_generator.py,sha256=hf5pZSKYVM2iSI6TlLYIxEOwKeBGvGiy-Xo3MR47hRU,7234
 official/vision/ops/anchor_generator_test.py,sha256=fa8PkMR_weA2TGiK_tVfeIZFV4bSUp9nGXuUFTaGfVg,5286
 official/vision/ops/anchor_test.py,sha256=Q8-7crDckzndyRMNmfeZXz88MTSYmLQpg3ELbhcdkak,7623
-official/vision/ops/augment.py,sha256=-Oevj7iFf--AsoiN03r8kiAILQ_1cjJojSDcNqB_jlA,91189
-official/vision/ops/augment_test.py,sha256=aUGf0ABjm2OSu5v-TZLmsomzSkRQk8ez9x5PQ8PlLvY,18896
+official/vision/ops/augment.py,sha256=pNa5xFsD5Pgv6YeyzCrvNXiEAttvKRkW_4YFP2F6xyg,91331
+official/vision/ops/augment_test.py,sha256=K7OwJSZ8C9iWYQJ7lxnk5qI8rIYTkkZhFI_q3zJw82g,19168
 official/vision/ops/box_matcher.py,sha256=AeDSdM8F6atkD71gniF0my1P0VsV93UcpeMayWVMOA4,9057
 official/vision/ops/box_matcher_test.py,sha256=N96ra6RHzRBgumomNfx0xZuOvEFWqQGeu77vZr7NK9w,2428
 official/vision/ops/box_ops.py,sha256=7oSB7dBJgVLS3Ny9hHAU6tjcExRpPjY-xmtadrKtgxQ,34487
@@ -1004,7 +1004,7 @@ official/vision/serving/export_saved_model_lib_v2.py,sha256=BlPSuA0OOFPIkm_iSRBt
 official/vision/serving/export_tfhub.py,sha256=KPxiKl1oN47fwqFeICwXcLW7rUN8435OjJ3VxNqZtPw,3500
 official/vision/serving/export_tfhub_lib.py,sha256=uU4lKFcZB7NqZjdA_k-AmI0bxqgHHz6kaVBtMHdiAs4,2870
 official/vision/serving/export_tflite.py,sha256=0yF12gNHXOr1oRIkrjFRtB80inviz4MoBIxppl_j-4A,5105
-official/vision/serving/export_tflite_lib.py,sha256=nJ1NxDuVRa-KhG4dneNhqNttyM9ZoKI79LHCIwXdepw,6556
+official/vision/serving/export_tflite_lib.py,sha256=_otzNiP7kjpr2TP3CBnj94s3zYF3DffeEgoCem6axrw,6907
 official/vision/serving/export_tflite_lib_test.py,sha256=ysXdFoMSp46rKJHvNw9wlCMp-heAHhY68BKaMMXIF1A,7185
 official/vision/serving/export_utils.py,sha256=36KfyIgF0-HxPaOJsWT4IH4heKkFWYnDVB-iX_mwm9I,4871
 official/vision/serving/image_classification.py,sha256=V20LDXOIWTSpUlmuUBwBh9Sa-ipaRo8kgwW4qVFiJZk,2858
@@ -1069,9 +1069,9 @@ tensorflow_models/__init__.py,sha256=021FKgqdPz3ds1xxfV67FWL7e5ECQ7WHbo67D37vAQI
 tensorflow_models/tensorflow_models_test.py,sha256=3oRV5seq-V1La0eY0IFpGLD7AKkiemylW8GyvZIRtmo,1385
 tensorflow_models/nlp/__init__.py,sha256=ro-1L0G8Z1wby8D1Jbaa3No-n73tiNEx7C4f8pAUNlk,807
 tensorflow_models/vision/__init__.py,sha256=3qeLW_6HkgH5hEclFog2DIRu1FSzOr3JynyM23zGhu8,833
-tf_models_nightly-2.11.0.dev20230322.dist-info/AUTHORS,sha256=1dG3fXVu9jlo7bul8xuix5F5vOnczMk7_yWn4y70uw0,337
-tf_models_nightly-2.11.0.dev20230322.dist-info/LICENSE,sha256=WxeBS_DejPZQabxtfMOM_xn8qoZNJDQjrT7z2wG1I4U,11512
-tf_models_nightly-2.11.0.dev20230322.dist-info/METADATA,sha256=nDQJeoSeIrhNpJ_zLaHs0TTJh6nTufa15zPzqMjIZUk,1426
-tf_models_nightly-2.11.0.dev20230322.dist-info/WHEEL,sha256=kGT74LWyRUZrL4VgLh6_g12IeVl_9u9ZVhadrgXZUEY,110
-tf_models_nightly-2.11.0.dev20230322.dist-info/top_level.txt,sha256=gum2FfO5R4cvjl2-QtP-S1aNmsvIZaFFT6VFzU0f4-g,33
-tf_models_nightly-2.11.0.dev20230322.dist-info/RECORD,,
+tf_models_nightly-2.11.0.dev20230323.dist-info/AUTHORS,sha256=1dG3fXVu9jlo7bul8xuix5F5vOnczMk7_yWn4y70uw0,337
+tf_models_nightly-2.11.0.dev20230323.dist-info/LICENSE,sha256=WxeBS_DejPZQabxtfMOM_xn8qoZNJDQjrT7z2wG1I4U,11512
+tf_models_nightly-2.11.0.dev20230323.dist-info/METADATA,sha256=rK5fmRrHIaf7Z3A655AJ9uUtsrcNQ9UAhxSzvS7mdE8,1426
+tf_models_nightly-2.11.0.dev20230323.dist-info/WHEEL,sha256=kGT74LWyRUZrL4VgLh6_g12IeVl_9u9ZVhadrgXZUEY,110
+tf_models_nightly-2.11.0.dev20230323.dist-info/top_level.txt,sha256=gum2FfO5R4cvjl2-QtP-S1aNmsvIZaFFT6VFzU0f4-g,33
+tf_models_nightly-2.11.0.dev20230323.dist-info/RECORD,,