Merge pull request #3845 from kinaryml:image-embedder-python

PiperOrigin-RevId: 487950862

mediapipe/python/BUILD

@@ -89,6 +89,7 @@ cc_library(
     deps = [
         "//mediapipe/tasks/cc/vision/gesture_recognizer:gesture_recognizer_graph",
         "//mediapipe/tasks/cc/vision/image_classifier:image_classifier_graph",
+        "//mediapipe/tasks/cc/vision/image_embedder:image_embedder_graph",
         "//mediapipe/tasks/cc/vision/image_segmenter:image_segmenter_graph",
         "//mediapipe/tasks/cc/vision/object_detector:object_detector_graph",
     ] + select({

mediapipe/tasks/python/components/containers/BUILD

@@ -95,3 +95,12 @@ py_library(
         "//mediapipe/tasks/python/core:optional_dependencies",
     ],
 )
+
+py_library(
+    name = "embedding_result",
+    srcs = ["embedding_result.py"],
+    deps = [
+        "//mediapipe/tasks/cc/components/containers/proto:embeddings_py_pb2",
+        "//mediapipe/tasks/python/core:optional_dependencies",
+    ],
+)

mediapipe/tasks/python/components/containers/embedding_result.py

@@ -0,0 +1,125 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Embeddings data class."""
+
+import dataclasses
+from typing import Optional, List
+
+import numpy as np
+from mediapipe.tasks.cc.components.containers.proto import embeddings_pb2
+from mediapipe.tasks.python.core.optional_dependencies import doc_controls
+
+_FloatEmbeddingProto = embeddings_pb2.FloatEmbedding
+_QuantizedEmbeddingProto = embeddings_pb2.QuantizedEmbedding
+_EmbeddingProto = embeddings_pb2.Embedding
+_EmbeddingResultProto = embeddings_pb2.EmbeddingResult
+
+
+@dataclasses.dataclass
+class FloatEmbedding:
+  """Defines a dense floating-point embedding.
+
+  Attributes:
+    values: A NumPy array indicating the raw output of the embedding layer.
+  """
+
+  values: np.ndarray
+
+  @classmethod
+  @doc_controls.do_not_generate_docs
+  def create_from_pb2(cls, pb2_obj: _FloatEmbeddingProto) -> 'FloatEmbedding':
+    """Creates a `FloatEmbedding` object from the given protobuf object."""
+    return FloatEmbedding(values=np.array(pb2_obj.values, dtype=float))
+
+
+@dataclasses.dataclass
+class QuantizedEmbedding:
+  """Defines a dense scalar-quantized embedding.
+
+  Attributes:
+    values: A NumPy array indicating the raw output of the embedding layer.
+  """
+
+  values: np.ndarray
+
+  @classmethod
+  @doc_controls.do_not_generate_docs
+  def create_from_pb2(
+      cls, pb2_obj: _QuantizedEmbeddingProto) -> 'QuantizedEmbedding':
+    """Creates a `QuantizedEmbedding` object from the given protobuf object."""
+    return QuantizedEmbedding(
+        values=np.array(bytearray(pb2_obj.values), dtype=np.uint8))
+
+
+@dataclasses.dataclass
+class Embedding:
+  """Embedding result for a given embedder head.
+
+  Attributes:
+    embedding: The actual embedding, either floating-point or scalar-quantized.
+    head_index: The index of the embedder head that produced this embedding.
+      This is useful for multi-head models.
+    head_name: The name of the embedder head, which is the corresponding tensor
+      metadata name (if any). This is useful for multi-head models.
+  """
+
+  embedding: np.ndarray
+  head_index: Optional[int] = None
+  head_name: Optional[str] = None
+
+  @classmethod
+  @doc_controls.do_not_generate_docs
+  def create_from_pb2(cls, pb2_obj: _EmbeddingProto) -> 'Embedding':
+    """Creates a `Embedding` object from the given protobuf object."""
+
+    quantized_embedding = np.array(
+        bytearray(pb2_obj.quantized_embedding.values))
+    float_embedding = np.array(pb2_obj.float_embedding.values, dtype=float)
+
+    if not quantized_embedding:
+      return Embedding(
+          embedding=float_embedding,
+          head_index=pb2_obj.head_index,
+          head_name=pb2_obj.head_name)
+    else:
+      return Embedding(
+          embedding=quantized_embedding,
+          head_index=pb2_obj.head_index,
+          head_name=pb2_obj.head_name)
+
+
+@dataclasses.dataclass
+class EmbeddingResult:
+  """Embedding results for a given embedder model.
+
+  Attributes:
+    embeddings: A list of `Embedding` objects.
+    timestamp_ms: The optional timestamp (in milliseconds) of the start of the
+      chunk of data corresponding to these results. This is only used for
+      embedding extraction on time series (e.g. audio embedding). In these use
+      cases, the amount of data to process might exceed the maximum size that
+      the model can process: to solve this, the input data is split into
+      multiple chunks starting at different timestamps.
+  """
+
+  embeddings: List[Embedding]
+  timestamp_ms: Optional[int] = None
+
+  @classmethod
+  @doc_controls.do_not_generate_docs
+  def create_from_pb2(cls, pb2_obj: _EmbeddingResultProto) -> 'EmbeddingResult':
+    """Creates a `EmbeddingResult` object from the given protobuf object."""
+    return EmbeddingResult(embeddings=[
+        Embedding.create_from_pb2(embedding) for embedding in pb2_obj.embeddings
+    ])

mediapipe/tasks/python/components/processors/BUILD

@@ -28,3 +28,12 @@ py_library(
         "//mediapipe/tasks/python/core:optional_dependencies",
     ],
 )
+
+py_library(
+    name = "embedder_options",
+    srcs = ["embedder_options.py"],
+    deps = [
+        "//mediapipe/tasks/cc/components/processors/proto:embedder_options_py_pb2",
+        "//mediapipe/tasks/python/core:optional_dependencies",
+    ],
+)

mediapipe/tasks/python/components/processors/embedder_options.py

@@ -0,0 +1,68 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Embedder options data class."""
+
+import dataclasses
+from typing import Any, Optional
+
+from mediapipe.tasks.cc.components.processors.proto import embedder_options_pb2
+from mediapipe.tasks.python.core.optional_dependencies import doc_controls
+
+_EmbedderOptionsProto = embedder_options_pb2.EmbedderOptions
+
+
+@dataclasses.dataclass
+class EmbedderOptions:
+  """Shared options used by all embedding extraction tasks.
+
+  Attributes:
+    l2_normalize: Whether to normalize the returned feature vector with L2 norm.
+      Use this option only if the model does not already contain a native
+      L2_NORMALIZATION TF Lite Op. In most cases, this is already the case and
+      L2 norm is thus achieved through TF Lite inference.
+    quantize: Whether the returned embedding should be quantized to bytes via
+      scalar quantization. Embeddings are implicitly assumed to be unit-norm and
+      therefore any dimension is guaranteed to have a value in [-1.0, 1.0]. Use
+      the l2_normalize option if this is not the case.
+  """
+
+  l2_normalize: Optional[bool] = None
+  quantize: Optional[bool] = None
+
+  @doc_controls.do_not_generate_docs
+  def to_pb2(self) -> _EmbedderOptionsProto:
+    """Generates a EmbedderOptions protobuf object."""
+    return _EmbedderOptionsProto(
+        l2_normalize=self.l2_normalize, quantize=self.quantize)
+
+  @classmethod
+  @doc_controls.do_not_generate_docs
+  def create_from_pb2(cls, pb2_obj: _EmbedderOptionsProto) -> 'EmbedderOptions':
+    """Creates a `EmbedderOptions` object from the given protobuf object."""
+    return EmbedderOptions(
+        l2_normalize=pb2_obj.l2_normalize, quantize=pb2_obj.quantize)
+
+  def __eq__(self, other: Any) -> bool:
+    """Checks if this object is equal to the given object.
+
+    Args:
+      other: The object to be compared with.
+
+    Returns:
+      True if the objects are equal.
+    """
+    if not isinstance(other, EmbedderOptions):
+      return False
+
+    return self.to_pb2().__eq__(other.to_pb2())

mediapipe/tasks/python/components/utils/BUILD

@@ -0,0 +1,30 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Placeholder for internal Python strict library compatibility macro.
+
+# Placeholder for internal Python strict library and test compatibility macro.
+
+package(default_visibility = ["//mediapipe/tasks:internal"])
+
+licenses(["notice"])
+
+py_library(
+    name = "cosine_similarity",
+    srcs = ["cosine_similarity.py"],
+    deps = [
+        "//mediapipe/tasks/python/components/containers:embedding_result",
+        "//mediapipe/tasks/python/components/processors:embedder_options",
+    ],
+)

mediapipe/tasks/python/components/utils/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

mediapipe/tasks/python/components/utils/cosine_similarity.py

@@ -0,0 +1,67 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Cosine similarity utilities."""
+
+import numpy as np
+
+from mediapipe.tasks.python.components.containers import embedding_result
+from mediapipe.tasks.python.components.processors import embedder_options
+
+_Embedding = embedding_result.Embedding
+_EmbedderOptions = embedder_options.EmbedderOptions
+
+
+def _compute_cosine_similarity(u, v):
+  """Computes cosine similarity between two embeddings."""
+
+  if len(u.embedding) <= 0:
+    raise ValueError("Cannot compute cosing similarity on empty embeddings.")
+
+  norm_u = np.linalg.norm(u.embedding)
+  norm_v = np.linalg.norm(v.embedding)
+
+  if norm_u <= 0 or norm_v <= 0:
+    raise ValueError(
+        "Cannot compute cosine similarity on embedding with 0 norm.")
+
+  return np.dot(u.embedding, v.embedding.T) / (norm_u * norm_v)
+
+
+def cosine_similarity(u: _Embedding, v: _Embedding) -> float:
+  """Utility function to compute cosine similarity between two embedding.
+
+  May return an InvalidArgumentError if e.g. the feature vectors are of
+  different types (quantized vs. float), have different sizes, or have an
+  L2-norm of 0.
+
+  Args:
+    u: An embedding.
+    v: An embedding.
+
+  Returns:
+    Cosine similarity value.
+  """
+  if len(u.embedding) != len(v.embedding):
+    raise ValueError(f"Cannot compute cosine similarity between embeddings "
+                     f"of different sizes "
+                     f"({len(u.embedding)} vs. {len(v.embedding)}).")
+
+  if u.embedding.dtype == float and v.embedding.dtype == float:
+    return _compute_cosine_similarity(u, v)
+
+  if u.embedding.dtype == np.uint8 and v.embedding.dtype == np.uint8:
+    return _compute_cosine_similarity(u, v)
+
+  raise ValueError("Cannot compute cosine similarity between quantized and "
+                   "float embeddings.")

mediapipe/tasks/python/test/vision/BUILD

@@ -58,6 +58,26 @@ py_test(
     ],
 )
 
+py_test(
+    name = "image_embedder_test",
+    srcs = ["image_embedder_test.py"],
+    data = [
+        "//mediapipe/tasks/testdata/vision:test_images",
+        "//mediapipe/tasks/testdata/vision:test_models",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/tasks/python/components/containers:embedding_result",
+        "//mediapipe/tasks/python/components/containers:rect",
+        "//mediapipe/tasks/python/components/processors:embedder_options",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/test:test_utils",
+        "//mediapipe/tasks/python/vision:image_embedder",
+        "//mediapipe/tasks/python/vision/core:image_processing_options",
+        "//mediapipe/tasks/python/vision/core:vision_task_running_mode",
+    ],
+)
+
 py_test(
     name = "image_segmenter_test",
     srcs = ["image_segmenter_test.py"],

mediapipe/tasks/python/test/vision/image_embedder_test.py

@@ -0,0 +1,402 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for image embedder."""
+
+import enum
+import os
+from unittest import mock
+
+from absl.testing import absltest
+from absl.testing import parameterized
+import numpy as np
+
+from mediapipe.python._framework_bindings import image as image_module
+from mediapipe.tasks.python.components.containers import embedding_result as embedding_result_module
+from mediapipe.tasks.python.components.containers import rect
+from mediapipe.tasks.python.components.processors import embedder_options as embedder_options_module
+from mediapipe.tasks.python.core import base_options as base_options_module
+from mediapipe.tasks.python.test import test_utils
+from mediapipe.tasks.python.vision import image_embedder
+from mediapipe.tasks.python.vision.core import image_processing_options as image_processing_options_module
+from mediapipe.tasks.python.vision.core import vision_task_running_mode as running_mode_module
+
+ImageEmbedderResult = embedding_result_module.EmbeddingResult
+_Rect = rect.Rect
+_BaseOptions = base_options_module.BaseOptions
+_EmbedderOptions = embedder_options_module.EmbedderOptions
+_FloatEmbedding = embedding_result_module.FloatEmbedding
+_QuantizedEmbedding = embedding_result_module.QuantizedEmbedding
+_Embedding = embedding_result_module.Embedding
+_Image = image_module.Image
+_ImageEmbedder = image_embedder.ImageEmbedder
+_ImageEmbedderOptions = image_embedder.ImageEmbedderOptions
+_RUNNING_MODE = running_mode_module.VisionTaskRunningMode
+_ImageProcessingOptions = image_processing_options_module.ImageProcessingOptions
+
+_MODEL_FILE = 'mobilenet_v3_small_100_224_embedder.tflite'
+_BURGER_IMAGE_FILE = 'burger.jpg'
+_BURGER_CROPPED_IMAGE_FILE = 'burger_crop.jpg'
+_TEST_DATA_DIR = 'mediapipe/tasks/testdata/vision'
+# Tolerance for embedding vector coordinate values.
+_EPSILON = 1e-4
+# Tolerance for cosine similarity evaluation.
+_SIMILARITY_TOLERANCE = 1e-6
+
+
+class ModelFileType(enum.Enum):
+  FILE_CONTENT = 1
+  FILE_NAME = 2
+
+
+class ImageEmbedderTest(parameterized.TestCase):
+
+  def setUp(self):
+    super().setUp()
+    self.test_image = _Image.create_from_file(
+        test_utils.get_test_data_path(
+            os.path.join(_TEST_DATA_DIR, _BURGER_IMAGE_FILE)))
+    self.test_cropped_image = _Image.create_from_file(
+        test_utils.get_test_data_path(
+            os.path.join(_TEST_DATA_DIR, _BURGER_CROPPED_IMAGE_FILE)))
+    self.model_path = test_utils.get_test_data_path(
+        os.path.join(_TEST_DATA_DIR, _MODEL_FILE))
+
+  def test_create_from_file_succeeds_with_valid_model_path(self):
+    # Creates with default option and valid model file successfully.
+    with _ImageEmbedder.create_from_model_path(self.model_path) as embedder:
+      self.assertIsInstance(embedder, _ImageEmbedder)
+
+  def test_create_from_options_succeeds_with_valid_model_path(self):
+    # Creates with options containing model file successfully.
+    base_options = _BaseOptions(model_asset_path=self.model_path)
+    options = _ImageEmbedderOptions(base_options=base_options)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      self.assertIsInstance(embedder, _ImageEmbedder)
+
+  def test_create_from_options_fails_with_invalid_model_path(self):
+    with self.assertRaisesRegex(
+        RuntimeError, 'Unable to open file at /path/to/invalid/model.tflite'):
+      base_options = _BaseOptions(
+          model_asset_path='/path/to/invalid/model.tflite')
+      options = _ImageEmbedderOptions(base_options=base_options)
+      _ImageEmbedder.create_from_options(options)
+
+  def test_create_from_options_succeeds_with_valid_model_content(self):
+    # Creates with options containing model content successfully.
+    with open(self.model_path, 'rb') as f:
+      base_options = _BaseOptions(model_asset_buffer=f.read())
+      options = _ImageEmbedderOptions(base_options=base_options)
+      embedder = _ImageEmbedder.create_from_options(options)
+      self.assertIsInstance(embedder, _ImageEmbedder)
+
+  def _check_embedding_value(self, result, expected_first_value):
+    # Check embedding first value.
+    self.assertAlmostEqual(
+        result.embeddings[0].embedding[0], expected_first_value, delta=_EPSILON)
+
+  def _check_embedding_size(self, result, quantize, expected_embedding_size):
+    # Check embedding size.
+    self.assertLen(result.embeddings, 1)
+    embedding_result = result.embeddings[0]
+    self.assertLen(embedding_result.embedding, expected_embedding_size)
+    if quantize:
+      self.assertEqual(embedding_result.embedding.dtype, np.uint8)
+    else:
+      self.assertEqual(embedding_result.embedding.dtype, float)
+
+  def _check_cosine_similarity(self, result0, result1, expected_similarity):
+    # Checks cosine similarity.
+    similarity = _ImageEmbedder.cosine_similarity(result0.embeddings[0],
+                                                  result1.embeddings[0])
+    self.assertAlmostEqual(
+        similarity, expected_similarity, delta=_SIMILARITY_TOLERANCE)
+
+  @parameterized.parameters(
+      (False, False, False, ModelFileType.FILE_NAME, 0.925519, 1024,
+       (-0.2101883, -0.193027)),
+      (True, False, False, ModelFileType.FILE_NAME, 0.925519, 1024,
+       (-0.0142344, -0.0131606)),
+      # (False, True, False, ModelFileType.FILE_NAME,
+      #  0.926791, 1024, (229, 231)),
+      (False, False, True, ModelFileType.FILE_CONTENT, 0.999931, 1024,
+       (-0.195062, -0.193027)))
+  def test_embed(self, l2_normalize, quantize, with_roi, model_file_type,
+                 expected_similarity, expected_size, expected_first_values):
+    # Creates embedder.
+    if model_file_type is ModelFileType.FILE_NAME:
+      base_options = _BaseOptions(model_asset_path=self.model_path)
+    elif model_file_type is ModelFileType.FILE_CONTENT:
+      with open(self.model_path, 'rb') as f:
+        model_content = f.read()
+      base_options = _BaseOptions(model_asset_buffer=model_content)
+    else:
+      # Should never happen
+      raise ValueError('model_file_type is invalid.')
+
+    embedder_options = _EmbedderOptions(
+        l2_normalize=l2_normalize, quantize=quantize)
+    options = _ImageEmbedderOptions(
+        base_options=base_options, embedder_options=embedder_options)
+    embedder = _ImageEmbedder.create_from_options(options)
+
+    image_processing_options = None
+    if with_roi:
+      # Region-of-interest in "burger.jpg" corresponding to "burger_crop.jpg".
+      roi = _Rect(left=0, top=0, right=0.833333, bottom=1)
+      image_processing_options = _ImageProcessingOptions(roi)
+
+    # Extracts both embeddings.
+    image_result = embedder.embed(self.test_image, image_processing_options)
+    crop_result = embedder.embed(self.test_cropped_image)
+
+    # Checks embeddings and cosine similarity.
+    expected_result0_value, expected_result1_value = expected_first_values
+    self._check_embedding_size(image_result, quantize, expected_size)
+    self._check_embedding_size(crop_result, quantize, expected_size)
+    self._check_embedding_value(image_result, expected_result0_value)
+    self._check_embedding_value(crop_result, expected_result1_value)
+    self._check_cosine_similarity(image_result, crop_result,
+                                  expected_similarity)
+    # Closes the embedder explicitly when the embedder is not used in
+    # a context.
+    embedder.close()
+
+  @parameterized.parameters(
+      (False, False, ModelFileType.FILE_NAME, 0.925519),
+      (False, False, ModelFileType.FILE_CONTENT, 0.925519))
+  def test_embed_in_context(self, l2_normalize, quantize, model_file_type,
+                            expected_similarity):
+    # Creates embedder.
+    if model_file_type is ModelFileType.FILE_NAME:
+      base_options = _BaseOptions(model_asset_path=self.model_path)
+    elif model_file_type is ModelFileType.FILE_CONTENT:
+      with open(self.model_path, 'rb') as f:
+        model_content = f.read()
+      base_options = _BaseOptions(model_asset_buffer=model_content)
+    else:
+      # Should never happen
+      raise ValueError('model_file_type is invalid.')
+
+    embedder_options = _EmbedderOptions(
+        l2_normalize=l2_normalize, quantize=quantize)
+    options = _ImageEmbedderOptions(
+        base_options=base_options, embedder_options=embedder_options)
+
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      # Extracts both embeddings.
+      image_result = embedder.embed(self.test_image)
+      crop_result = embedder.embed(self.test_cropped_image)
+
+      # Checks cosine similarity.
+      self._check_cosine_similarity(image_result, crop_result,
+                                    expected_similarity)
+
+  def test_missing_result_callback(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.LIVE_STREAM)
+    with self.assertRaisesRegex(ValueError,
+                                r'result callback must be provided'):
+      with _ImageEmbedder.create_from_options(options) as unused_embedder:
+        pass
+
+  @parameterized.parameters((_RUNNING_MODE.IMAGE), (_RUNNING_MODE.VIDEO))
+  def test_illegal_result_callback(self, running_mode):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=running_mode,
+        result_callback=mock.MagicMock())
+    with self.assertRaisesRegex(ValueError,
+                                r'result callback should not be provided'):
+      with _ImageEmbedder.create_from_options(options) as unused_embedder:
+        pass
+
+  def test_calling_embed_for_video_in_image_mode(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.IMAGE)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      with self.assertRaisesRegex(ValueError,
+                                  r'not initialized with the video mode'):
+        embedder.embed_for_video(self.test_image, 0)
+
+  def test_calling_embed_async_in_image_mode(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.IMAGE)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      with self.assertRaisesRegex(ValueError,
+                                  r'not initialized with the live stream mode'):
+        embedder.embed_async(self.test_image, 0)
+
+  def test_calling_embed_in_video_mode(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.VIDEO)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      with self.assertRaisesRegex(ValueError,
+                                  r'not initialized with the image mode'):
+        embedder.embed(self.test_image)
+
+  def test_calling_embed_async_in_video_mode(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.VIDEO)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      with self.assertRaisesRegex(ValueError,
+                                  r'not initialized with the live stream mode'):
+        embedder.embed_async(self.test_image, 0)
+
+  def test_embed_for_video_with_out_of_order_timestamp(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.VIDEO)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      unused_result = embedder.embed_for_video(self.test_image, 1)
+      with self.assertRaisesRegex(
+          ValueError, r'Input timestamp must be monotonically increasing'):
+        embedder.embed_for_video(self.test_image, 0)
+
+  def test_embed_for_video(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.VIDEO)
+    with _ImageEmbedder.create_from_options(options) as embedder0, \
+         _ImageEmbedder.create_from_options(options) as embedder1:
+      for timestamp in range(0, 300, 30):
+        # Extracts both embeddings.
+        image_result = embedder0.embed_for_video(self.test_image, timestamp)
+        crop_result = embedder1.embed_for_video(self.test_cropped_image,
+                                                timestamp)
+        # Checks cosine similarity.
+        self._check_cosine_similarity(
+            image_result, crop_result, expected_similarity=0.925519)
+
+  def test_embed_for_video_succeeds_with_region_of_interest(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.VIDEO)
+    with _ImageEmbedder.create_from_options(options) as embedder0, \
+         _ImageEmbedder.create_from_options(options) as embedder1:
+      # Region-of-interest in "burger.jpg" corresponding to "burger_crop.jpg".
+      roi = _Rect(left=0, top=0, right=0.833333, bottom=1)
+      image_processing_options = _ImageProcessingOptions(roi)
+
+      for timestamp in range(0, 300, 30):
+        # Extracts both embeddings.
+        image_result = embedder0.embed_for_video(self.test_image, timestamp,
+                                                 image_processing_options)
+        crop_result = embedder1.embed_for_video(self.test_cropped_image,
+                                                timestamp)
+
+        # Checks cosine similarity.
+        self._check_cosine_similarity(
+            image_result, crop_result, expected_similarity=0.999931)
+
+  def test_calling_embed_in_live_stream_mode(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.LIVE_STREAM,
+        result_callback=mock.MagicMock())
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      with self.assertRaisesRegex(ValueError,
+                                  r'not initialized with the image mode'):
+        embedder.embed(self.test_image)
+
+  def test_calling_embed_for_video_in_live_stream_mode(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.LIVE_STREAM,
+        result_callback=mock.MagicMock())
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      with self.assertRaisesRegex(ValueError,
+                                  r'not initialized with the video mode'):
+        embedder.embed_for_video(self.test_image, 0)
+
+  def test_embed_async_calls_with_illegal_timestamp(self):
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.LIVE_STREAM,
+        result_callback=mock.MagicMock())
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      embedder.embed_async(self.test_image, 100)
+      with self.assertRaisesRegex(
+          ValueError, r'Input timestamp must be monotonically increasing'):
+        embedder.embed_async(self.test_image, 0)
+
+  def test_embed_async_calls(self):
+    # Get the embedding result for the cropped image.
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.IMAGE)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      crop_result = embedder.embed(self.test_cropped_image)
+
+    observed_timestamp_ms = -1
+
+    def check_result(result: ImageEmbedderResult, output_image: _Image,
+                     timestamp_ms: int):
+      # Checks cosine similarity.
+      self._check_cosine_similarity(
+          result, crop_result, expected_similarity=0.925519)
+      self.assertTrue(
+          np.array_equal(output_image.numpy_view(),
+                         self.test_image.numpy_view()))
+      self.assertLess(observed_timestamp_ms, timestamp_ms)
+      self.observed_timestamp_ms = timestamp_ms
+
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.LIVE_STREAM,
+        result_callback=check_result)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      for timestamp in range(0, 300, 30):
+        embedder.embed_async(self.test_image, timestamp)
+
+  def test_embed_async_succeeds_with_region_of_interest(self):
+    # Get the embedding result for the cropped image.
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.IMAGE)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      crop_result = embedder.embed(self.test_cropped_image)
+
+    # Region-of-interest in "burger.jpg" corresponding to "burger_crop.jpg".
+    roi = _Rect(left=0, top=0, right=0.833333, bottom=1)
+    image_processing_options = _ImageProcessingOptions(roi)
+    observed_timestamp_ms = -1
+
+    def check_result(result: ImageEmbedderResult, output_image: _Image,
+                     timestamp_ms: int):
+      # Checks cosine similarity.
+      self._check_cosine_similarity(
+          result, crop_result, expected_similarity=0.999931)
+      self.assertTrue(
+          np.array_equal(output_image.numpy_view(),
+                         self.test_image.numpy_view()))
+      self.assertLess(observed_timestamp_ms, timestamp_ms)
+      self.observed_timestamp_ms = timestamp_ms
+
+    options = _ImageEmbedderOptions(
+        base_options=_BaseOptions(model_asset_path=self.model_path),
+        running_mode=_RUNNING_MODE.LIVE_STREAM,
+        result_callback=check_result)
+    with _ImageEmbedder.create_from_options(options) as embedder:
+      for timestamp in range(0, 300, 30):
+        embedder.embed_async(self.test_image, timestamp,
+                             image_processing_options)
+
+
+if __name__ == '__main__':
+  absltest.main()

mediapipe/tasks/python/vision/BUILD

@@ -79,6 +79,29 @@ py_library(
     ],
 )
 
+py_library(
+    name = "image_embedder",
+    srcs = [
+        "image_embedder.py",
+    ],
+    deps = [
+        "//mediapipe/python:_framework_bindings",
+        "//mediapipe/python:packet_creator",
+        "//mediapipe/python:packet_getter",
+        "//mediapipe/tasks/cc/components/containers/proto:embeddings_py_pb2",
+        "//mediapipe/tasks/cc/vision/image_embedder/proto:image_embedder_graph_options_py_pb2",
+        "//mediapipe/tasks/python/components/containers:embedding_result",
+        "//mediapipe/tasks/python/components/processors:embedder_options",
+        "//mediapipe/tasks/python/components/utils:cosine_similarity",
+        "//mediapipe/tasks/python/core:base_options",
+        "//mediapipe/tasks/python/core:optional_dependencies",
+        "//mediapipe/tasks/python/core:task_info",
+        "//mediapipe/tasks/python/vision/core:base_vision_task_api",
+        "//mediapipe/tasks/python/vision/core:image_processing_options",
+        "//mediapipe/tasks/python/vision/core:vision_task_running_mode",
+    ],
+)
+
 py_library(
     name = "gesture_recognizer",
     srcs = [

mediapipe/tasks/python/vision/image_embedder.py

@@ -0,0 +1,309 @@
+# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""MediaPipe image embedder task."""
+
+import dataclasses
+from typing import Callable, Mapping, Optional
+
+from mediapipe.python import packet_creator
+from mediapipe.python import packet_getter
+from mediapipe.python._framework_bindings import image as image_module
+from mediapipe.python._framework_bindings import packet as packet_module
+from mediapipe.tasks.cc.components.containers.proto import embeddings_pb2
+from mediapipe.tasks.cc.vision.image_embedder.proto import image_embedder_graph_options_pb2
+from mediapipe.tasks.python.components.containers import embedding_result as embedding_result_module
+from mediapipe.tasks.python.components.processors import embedder_options
+from mediapipe.tasks.python.components.utils import cosine_similarity
+from mediapipe.tasks.python.core import base_options as base_options_module
+from mediapipe.tasks.python.core import task_info as task_info_module
+from mediapipe.tasks.python.core.optional_dependencies import doc_controls
+from mediapipe.tasks.python.vision.core import base_vision_task_api
+from mediapipe.tasks.python.vision.core import image_processing_options as image_processing_options_module
+from mediapipe.tasks.python.vision.core import vision_task_running_mode as running_mode_module
+
+ImageEmbedderResult = embedding_result_module.EmbeddingResult
+_BaseOptions = base_options_module.BaseOptions
+_ImageEmbedderGraphOptionsProto = image_embedder_graph_options_pb2.ImageEmbedderGraphOptions
+_EmbedderOptions = embedder_options.EmbedderOptions
+_RunningMode = running_mode_module.VisionTaskRunningMode
+_TaskInfo = task_info_module.TaskInfo
+_ImageProcessingOptions = image_processing_options_module.ImageProcessingOptions
+
+_EMBEDDINGS_OUT_STREAM_NAME = 'embeddings_out'
+_EMBEDDINGS_TAG = 'EMBEDDINGS'
+_IMAGE_IN_STREAM_NAME = 'image_in'
+_IMAGE_OUT_STREAM_NAME = 'image_out'
+_IMAGE_TAG = 'IMAGE'
+_NORM_RECT_STREAM_NAME = 'norm_rect_in'
+_NORM_RECT_TAG = 'NORM_RECT'
+_TASK_GRAPH_NAME = 'mediapipe.tasks.vision.image_embedder.ImageEmbedderGraph'
+_MICRO_SECONDS_PER_MILLISECOND = 1000
+
+
+@dataclasses.dataclass
+class ImageEmbedderOptions:
+  """Options for the image embedder task.
+
+  Attributes:
+    base_options: Base options for the image embedder task.
+    running_mode: The running mode of the task. Default to the image mode. Image
+      embedder task has three running modes: 1) The image mode for embedding
+      image on single image inputs. 2) The video mode for embedding image on the
+      decoded frames of a video. 3) The live stream mode for embedding image on
+      a live stream of input data, such as from camera.
+    embedder_options: Options for the image embedder task.
+    result_callback: The user-defined result callback for processing live stream
+      data. The result callback should only be specified when the running mode
+      is set to the live stream mode.
+  """
+  base_options: _BaseOptions
+  running_mode: _RunningMode = _RunningMode.IMAGE
+  embedder_options: _EmbedderOptions = _EmbedderOptions()
+  result_callback: Optional[Callable[
+      [ImageEmbedderResult, image_module.Image, int], None]] = None
+
+  @doc_controls.do_not_generate_docs
+  def to_pb2(self) -> _ImageEmbedderGraphOptionsProto:
+    """Generates an ImageEmbedderOptions protobuf object."""
+    base_options_proto = self.base_options.to_pb2()
+    base_options_proto.use_stream_mode = False if self.running_mode == _RunningMode.IMAGE else True
+    embedder_options_proto = self.embedder_options.to_pb2()
+
+    return _ImageEmbedderGraphOptionsProto(
+        base_options=base_options_proto,
+        embedder_options=embedder_options_proto)
+
+
+class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi):
+  """Class that performs embedding extraction on images."""
+
+  @classmethod
+  def create_from_model_path(cls, model_path: str) -> 'ImageEmbedder':
+    """Creates an `ImageEmbedder` object from a TensorFlow Lite model and the default `ImageEmbedderOptions`.
+
+    Note that the created `ImageEmbedder` instance is in image mode, for
+    embedding image on single image inputs.
+
+    Args:
+      model_path: Path to the model.
+
+    Returns:
+      `ImageEmbedder` object that's created from the model file and the default
+      `ImageEmbedderOptions`.
+
+    Raises:
+      ValueError: If failed to create `ImageEmbedder` object from the provided
+        file such as invalid file path.
+      RuntimeError: If other types of error occurred.
+    """
+    base_options = _BaseOptions(model_asset_path=model_path)
+    options = ImageEmbedderOptions(
+        base_options=base_options, running_mode=_RunningMode.IMAGE)
+    return cls.create_from_options(options)
+
+  @classmethod
+  def create_from_options(cls,
+                          options: ImageEmbedderOptions) -> 'ImageEmbedder':
+    """Creates the `ImageEmbedder` object from image embedder options.
+
+    Args:
+      options: Options for the image embedder task.
+
+    Returns:
+      `ImageEmbedder` object that's created from `options`.
+
+    Raises:
+      ValueError: If failed to create `ImageEmbedder` object from
+        `ImageEmbedderOptions` such as missing the model.
+      RuntimeError: If other types of error occurred.
+    """
+
+    def packets_callback(output_packets: Mapping[str, packet_module.Packet]):
+      if output_packets[_IMAGE_OUT_STREAM_NAME].is_empty():
+        return
+
+      embedding_result_proto = embeddings_pb2.EmbeddingResult()
+      embedding_result_proto.CopyFrom(
+          packet_getter.get_proto(output_packets[_EMBEDDINGS_OUT_STREAM_NAME]))
+
+      image = packet_getter.get_image(output_packets[_IMAGE_OUT_STREAM_NAME])
+      timestamp = output_packets[_IMAGE_OUT_STREAM_NAME].timestamp
+      options.result_callback(
+          ImageEmbedderResult.create_from_pb2(embedding_result_proto), image,
+          timestamp.value // _MICRO_SECONDS_PER_MILLISECOND)
+
+    task_info = _TaskInfo(
+        task_graph=_TASK_GRAPH_NAME,
+        input_streams=[
+            ':'.join([_IMAGE_TAG, _IMAGE_IN_STREAM_NAME]),
+            ':'.join([_NORM_RECT_TAG, _NORM_RECT_STREAM_NAME]),
+        ],
+        output_streams=[
+            ':'.join([_EMBEDDINGS_TAG, _EMBEDDINGS_OUT_STREAM_NAME]),
+            ':'.join([_IMAGE_TAG, _IMAGE_OUT_STREAM_NAME])
+        ],
+        task_options=options)
+    return cls(
+        task_info.generate_graph_config(
+            enable_flow_limiting=options.running_mode ==
+            _RunningMode.LIVE_STREAM), options.running_mode,
+        packets_callback if options.result_callback else None)
+
+  def embed(
+      self,
+      image: image_module.Image,
+      image_processing_options: Optional[_ImageProcessingOptions] = None
+  ) -> ImageEmbedderResult:
+    """Performs image embedding extraction on the provided MediaPipe Image.
+
+     Extraction is performed on the region of interest specified by the `roi`
+     argument if provided, or on the entire image otherwise.
+
+    Args:
+      image: MediaPipe Image.
+      image_processing_options: Options for image processing.
+
+    Returns:
+      An embedding result object that contains a list of embeddings.
+
+    Raises:
+      ValueError: If any of the input arguments is invalid.
+      RuntimeError: If image embedder failed to run.
+    """
+    normalized_rect = self.convert_to_normalized_rect(image_processing_options)
+    output_packets = self._process_image_data({
+        _IMAGE_IN_STREAM_NAME:
+            packet_creator.create_image(image),
+        _NORM_RECT_STREAM_NAME:
+            packet_creator.create_proto(normalized_rect.to_pb2())
+    })
+
+    embedding_result_proto = embeddings_pb2.EmbeddingResult()
+    embedding_result_proto.CopyFrom(
+        packet_getter.get_proto(output_packets[_EMBEDDINGS_OUT_STREAM_NAME]))
+
+    return ImageEmbedderResult.create_from_pb2(embedding_result_proto)
+
+  def embed_for_video(
+      self,
+      image: image_module.Image,
+      timestamp_ms: int,
+      image_processing_options: Optional[_ImageProcessingOptions] = None
+  ) -> ImageEmbedderResult:
+    """Performs image embedding extraction on the provided video frames.
+
+    Extraction is performed on the region of interested specified by the `roi`
+    argument if provided, or on the entire image otherwise.
+
+    Only use this method when the ImageEmbedder is created with the video
+    running mode. It's required to provide the video frame's timestamp (in
+    milliseconds) along with the video frame. The input timestamps should be
+    monotonically increasing for adjacent calls of this method.
+
+    Args:
+      image: MediaPipe Image.
+      timestamp_ms: The timestamp of the input video frame in milliseconds.
+      image_processing_options: Options for image processing.
+
+    Returns:
+      An embedding result object that contains a list of embeddings.
+
+    Raises:
+      ValueError: If any of the input arguments is invalid.
+      RuntimeError: If image embedder failed to run.
+    """
+    normalized_rect = self.convert_to_normalized_rect(image_processing_options)
+    output_packets = self._process_video_data({
+        _IMAGE_IN_STREAM_NAME:
+            packet_creator.create_image(image).at(
+                timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND),
+        _NORM_RECT_STREAM_NAME:
+            packet_creator.create_proto(normalized_rect.to_pb2()).at(
+                timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND)
+    })
+    embedding_result_proto = embeddings_pb2.EmbeddingResult()
+    embedding_result_proto.CopyFrom(
+        packet_getter.get_proto(output_packets[_EMBEDDINGS_OUT_STREAM_NAME]))
+
+    return ImageEmbedderResult.create_from_pb2(embedding_result_proto)
+
+  def embed_async(
+      self,
+      image: image_module.Image,
+      timestamp_ms: int,
+      image_processing_options: Optional[_ImageProcessingOptions] = None
+  ) -> None:
+    """Sends live image data to embedder.
+
+    The results will be available via the "result_callback" provided in the
+    ImageEmbedderOptions. Embedding extraction is performed on the region of
+    interested specified by the `roi` argument if provided, or on the entire
+    image otherwise.
+
+    Only use this method when the ImageEmbedder is created with the live
+    stream running mode. The input timestamps should be monotonically increasing
+    for adjacent calls of this method. This method will return immediately after
+    the input image is accepted. The results will be available via the
+    `result_callback` provided in the `ImageEmbedderOptions`. The
+    `embed_async` method is designed to process live stream data such as
+    camera input. To lower the overall latency, image embedder may drop the
+    input images if needed. In other words, it's not guaranteed to have output
+    per input image.
+
+    The `result_callback` provides:
+      - An embedding result object that contains a list of embeddings.
+      - The input image that the image embedder runs on.
+      - The input timestamp in milliseconds.
+
+    Args:
+      image: MediaPipe Image.
+      timestamp_ms: The timestamp of the input image in milliseconds.
+      image_processing_options: Options for image processing.
+
+    Raises:
+      ValueError: If the current input timestamp is smaller than what the image
+        embedder has already processed.
+    """
+    normalized_rect = self.convert_to_normalized_rect(image_processing_options)
+    self._send_live_stream_data({
+        _IMAGE_IN_STREAM_NAME:
+            packet_creator.create_image(image).at(
+                timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND),
+        _NORM_RECT_STREAM_NAME:
+            packet_creator.create_proto(normalized_rect.to_pb2()).at(
+                timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND)
+    })
+
+  @classmethod
+  def cosine_similarity(cls, u: embedding_result_module.Embedding,
+                        v: embedding_result_module.Embedding) -> float:
+    """Utility function to compute cosine similarity between two embedding entries.
+
+    May return an InvalidArgumentError if e.g. the feature vectors are of
+    different types (quantized vs. float), have different sizes, or have an
+    L2-norm of 0.
+
+    Args:
+      u: An embedding entry.
+      v: An embedding entry.
+
+    Returns:
+      The cosine similarity for the two embeddings.
+
+    Raises:
+      ValueError: May return an error if e.g. the feature vectors are of
+        different types (quantized vs. float), have different sizes, or have
+        an L2-norm of 0.
+    """
+    return cosine_similarity.cosine_similarity(u, v)