From 71d5b695442ef36c1a574b2c7b6a956b5d6f2667 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Thu, 20 Oct 2022 02:29:14 -0700 Subject: [PATCH 01/10] Added files for the image embedder implementation and a simple test --- mediapipe/python/BUILD | 1 + .../tasks/python/components/containers/BUILD | 18 ++ .../components/containers/embeddings.py | 246 +++++++++++++++ .../python/components/containers/rect.py | 141 +++++++++ mediapipe/tasks/python/components/proto/BUILD | 28 ++ .../tasks/python/components/proto/__init__.py | 13 + .../components/proto/embedder_options.py | 72 +++++ mediapipe/tasks/python/test/vision/BUILD | 19 ++ .../python/test/vision/image_embedder_test.py | 98 ++++++ mediapipe/tasks/python/vision/BUILD | 19 ++ .../tasks/python/vision/image_embedder.py | 288 ++++++++++++++++++ 11 files changed, 943 insertions(+) create mode 100644 mediapipe/tasks/python/components/containers/embeddings.py create mode 100644 mediapipe/tasks/python/components/containers/rect.py create mode 100644 mediapipe/tasks/python/components/proto/BUILD create mode 100644 mediapipe/tasks/python/components/proto/__init__.py create mode 100644 mediapipe/tasks/python/components/proto/embedder_options.py create mode 100644 mediapipe/tasks/python/test/vision/image_embedder_test.py create mode 100644 mediapipe/tasks/python/vision/image_embedder.py diff --git a/mediapipe/python/BUILD b/mediapipe/python/BUILD index 2911e2fd6..3df0e2798 100644 --- a/mediapipe/python/BUILD +++ b/mediapipe/python/BUILD @@ -88,6 +88,7 @@ cc_library( name = "builtin_task_graphs", deps = [ "//mediapipe/tasks/cc/vision/object_detector:object_detector_graph", + "//mediapipe/tasks/cc/vision/image_embedder:image_embedder_graph", ], ) diff --git a/mediapipe/tasks/python/components/containers/BUILD b/mediapipe/tasks/python/components/containers/BUILD index 8dd9fcd60..cb123562f 100644 --- a/mediapipe/tasks/python/components/containers/BUILD +++ b/mediapipe/tasks/python/components/containers/BUILD @@ -27,6 +27,15 @@ py_library( ], ) +py_library( + name = "rect", + srcs = ["rect.py"], + deps = [ + "//mediapipe/framework/formats:rect_py_pb2", + "//mediapipe/tasks/python/core:optional_dependencies", + ], +) + py_library( name = "category", srcs = ["category.py"], @@ -47,3 +56,12 @@ py_library( "//mediapipe/tasks/python/core:optional_dependencies", ], ) + +py_library( + name = "embeddings", + srcs = ["embeddings.py"], + deps = [ + "//mediapipe/tasks/cc/components/containers/proto:embeddings_py_pb2", + "//mediapipe/tasks/python/core:optional_dependencies", + ], +) diff --git a/mediapipe/tasks/python/components/containers/embeddings.py b/mediapipe/tasks/python/components/containers/embeddings.py new file mode 100644 index 000000000..21f53670c --- /dev/null +++ b/mediapipe/tasks/python/components/containers/embeddings.py @@ -0,0 +1,246 @@ +# 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 Any, 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 +_EmbeddingEntryProto = embeddings_pb2.EmbeddingEntry +_EmbeddingsProto = embeddings_pb2.Embeddings +_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 + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _FloatEmbeddingProto: + """Generates a FloatEmbedding protobuf object.""" + return _FloatEmbeddingProto(values=self.values) + + @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.value_float, dtype=float)) + + 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, FloatEmbedding): + return False + + return self.to_pb2().__eq__(other.to_pb2()) + + +@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 + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _QuantizedEmbeddingProto: + """Generates a QuantizedEmbedding protobuf object.""" + return _QuantizedEmbeddingProto(values=self.values) + + @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.value_string), dtype=np.uint8)) + + 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, QuantizedEmbedding): + return False + + return self.to_pb2().__eq__(other.to_pb2()) + + +@dataclasses.dataclass +class EmbeddingEntry: + """Floating-point or scalar-quantized embedding with an optional timestamp. + + Attributes: + embedding: The actual embedding, either floating-point or scalar-quantized. + timestamp_ms: The optional timestamp (in milliseconds) associated to the + embedding entry. This is useful for time series use cases, e.g. audio + embedding. + """ + + embedding: np.ndarray + timestamp_ms: Optional[int] = None + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _EmbeddingEntryProto: + """Generates a EmbeddingEntry protobuf object.""" + + if self.embedding.dtype == float: + return _EmbeddingEntryProto(float_embedding=self.embedding) + + elif self.embedding.dtype == np.uint8: + return _EmbeddingEntryProto(quantized_embedding=bytes(self.embedding)) + + else: + raise ValueError("Invalid dtype. Only float and np.uint8 are supported.") + + @classmethod + @doc_controls.do_not_generate_docs + def create_from_pb2( + cls, pb2_obj: _EmbeddingEntryProto) -> 'EmbeddingEntry': + """Creates a `EmbeddingEntry` object from the given protobuf object.""" + + if pb2_obj.float_embedding: + return EmbeddingEntry( + embedding=np.array(pb2_obj.float_embedding.values, dtype=float)) + + elif pb2_obj.quantized_embedding: + return EmbeddingEntry( + embedding=np.array(bytearray(pb2_obj.quantized_embedding.values), + dtype=np.uint8)) + + else: + raise ValueError("Either float_embedding or quantized_embedding must " + "exist.") + + 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, EmbeddingEntry): + return False + + return self.to_pb2().__eq__(other.to_pb2()) + + +@dataclasses.dataclass +class Embeddings: + """Embeddings for a given embedder head. + Attributes: + entries: A list of `ClassificationEntry` objects. + 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. + """ + + entries: List[EmbeddingEntry] + head_index: int + head_name: str + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _EmbeddingsProto: + """Generates a Embeddings protobuf object.""" + return _EmbeddingsProto( + entries=[entry.to_pb2() for entry in self.entries], + head_index=self.head_index, + head_name=self.head_name) + + @classmethod + @doc_controls.do_not_generate_docs + def create_from_pb2(cls, pb2_obj: _EmbeddingsProto) -> 'Embeddings': + """Creates a `Embeddings` object from the given protobuf object.""" + return Embeddings( + entries=[ + EmbeddingEntry.create_from_pb2(entry) + for entry in pb2_obj.entries + ], + head_index=pb2_obj.head_index, + head_name=pb2_obj.head_name) + + 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, Embeddings): + return False + + return self.to_pb2().__eq__(other.to_pb2()) + + +@dataclasses.dataclass +class EmbeddingResult: + """Contains one set of results per embedder head. + Attributes: + embeddings: A list of `Embeddings` objects. + """ + + embeddings: List[Embeddings] + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _EmbeddingResultProto: + """Generates a EmbeddingResult protobuf object.""" + return _EmbeddingResultProto( + embeddings=[ + embedding.to_pb2() for embedding in self.embeddings + ]) + + @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=[ + Embeddings.create_from_pb2(embedding) + for embedding in pb2_obj.embeddings + ]) + + 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, EmbeddingResult): + return False + + return self.to_pb2().__eq__(other.to_pb2()) diff --git a/mediapipe/tasks/python/components/containers/rect.py b/mediapipe/tasks/python/components/containers/rect.py new file mode 100644 index 000000000..aadb404db --- /dev/null +++ b/mediapipe/tasks/python/components/containers/rect.py @@ -0,0 +1,141 @@ +# 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. +"""Rect data class.""" + +import dataclasses +from typing import Any, Optional + +from mediapipe.framework.formats import rect_pb2 +from mediapipe.tasks.python.core.optional_dependencies import doc_controls + +_RectProto = rect_pb2.Rect +_NormalizedRectProto = rect_pb2.NormalizedRect + + +@dataclasses.dataclass +class Rect: + """A rectangle with rotation in image coordinates. + + Attributes: + x_center : The X coordinate of the top-left corner, in pixels. + y_center : The Y coordinate of the top-left corner, in pixels. + width: The width of the rectangle, in pixels. + height: The height of the rectangle, in pixels. + rotation: Rotation angle is clockwise in radians. + rect_id: Optional unique id to help associate different rectangles to each + other. + """ + + x_center: int + y_center: int + width: int + height: int + rotation: Optional[float] = 0.0 + rect_id: Optional[int] = None + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _RectProto: + """Generates a Rect protobuf object.""" + return _RectProto( + x_center=self.x_center, + y_center=self.y_center, + width=self.width, + height=self.height, + ) + + @classmethod + @doc_controls.do_not_generate_docs + def create_from_pb2(cls, pb2_obj: _RectProto) -> 'Rect': + """Creates a `Rect` object from the given protobuf object.""" + return Rect( + x_center=pb2_obj.x_center, + y_center=pb2_obj.y_center, + width=pb2_obj.width, + height=pb2_obj.height) + + 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, Rect): + return False + + return self.to_pb2().__eq__(other.to_pb2()) + + +@dataclasses.dataclass +class NormalizedRect: + """A rectangle with rotation in normalized coordinates. The values of box + center location and size are within [0, 1]. + + Attributes: + x_center : The X normalized coordinate of the top-left corner. + y_center : The Y normalized coordinate of the top-left corner. + width: The width of the rectangle. + height: The height of the rectangle. + rotation: Rotation angle is clockwise in radians. + rect_id: Optional unique id to help associate different rectangles to each + other. + """ + + x_center: float + y_center: float + width: float + height: float + rotation: Optional[float] = 0.0 + rect_id: Optional[int] = None + + @doc_controls.do_not_generate_docs + def to_pb2(self) -> _NormalizedRectProto: + """Generates a NormalizedRect protobuf object.""" + return _NormalizedRectProto( + x_center=self.x_center, + y_center=self.y_center, + width=self.width, + height=self.height, + rotation=self.rotation, + rect_id=self.rect_id + ) + + @classmethod + @doc_controls.do_not_generate_docs + def create_from_pb2(cls, pb2_obj: _NormalizedRectProto) -> 'NormalizedRect': + """Creates a `NormalizedRect` object from the given protobuf object.""" + return NormalizedRect( + x_center=pb2_obj.x_center, + y_center=pb2_obj.y_center, + width=pb2_obj.width, + height=pb2_obj.height, + rotation=pb2_obj.rotation, + rect_id=pb2_obj.rect_id + ) + + 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, NormalizedRect): + return False + + return self.to_pb2().__eq__(other.to_pb2()) diff --git a/mediapipe/tasks/python/components/proto/BUILD b/mediapipe/tasks/python/components/proto/BUILD new file mode 100644 index 000000000..973f150ca --- /dev/null +++ b/mediapipe/tasks/python/components/proto/BUILD @@ -0,0 +1,28 @@ +# 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. + +package(default_visibility = ["//mediapipe/tasks:internal"]) + +licenses(["notice"]) + +py_library( + name = "embedder_options", + srcs = ["embedder_options.py"], + deps = [ + "//mediapipe/tasks/cc/components/proto:embedder_options_py_pb2", + "//mediapipe/tasks/python/core:optional_dependencies", + ], +) diff --git a/mediapipe/tasks/python/components/proto/__init__.py b/mediapipe/tasks/python/components/proto/__init__.py new file mode 100644 index 000000000..65c1214af --- /dev/null +++ b/mediapipe/tasks/python/components/proto/__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. diff --git a/mediapipe/tasks/python/components/proto/embedder_options.py b/mediapipe/tasks/python/components/proto/embedder_options.py new file mode 100644 index 000000000..3c257b976 --- /dev/null +++ b/mediapipe/tasks/python/components/proto/embedder_options.py @@ -0,0 +1,72 @@ +# 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.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()) diff --git a/mediapipe/tasks/python/test/vision/BUILD b/mediapipe/tasks/python/test/vision/BUILD index 290b665e7..62595d377 100644 --- a/mediapipe/tasks/python/test/vision/BUILD +++ b/mediapipe/tasks/python/test/vision/BUILD @@ -36,3 +36,22 @@ py_test( "//mediapipe/tasks/python/vision/core:vision_task_running_mode", ], ) + +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/proto:embedder_options", + "//mediapipe/tasks/python/components/containers:embeddings", + "//mediapipe/tasks/python/components/containers:rect", + "//mediapipe/tasks/python/core:base_options", + "//mediapipe/tasks/python/test:test_utils", + "//mediapipe/tasks/python/vision:image_embedder", + "//mediapipe/tasks/python/vision/core:vision_task_running_mode", + ], +) diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py new file mode 100644 index 000000000..8ddf3c992 --- /dev/null +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -0,0 +1,98 @@ +# 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 +from unittest import mock + +import numpy as np +from absl.testing import absltest +from absl.testing import parameterized + +from mediapipe.python._framework_bindings import image as image_module +from mediapipe.tasks.python.components.proto import embedder_options as embedder_options_module +from mediapipe.tasks.python.components.containers import embeddings as embeddings_module +from mediapipe.tasks.python.components.containers import rect as rect_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 vision_task_running_mode as running_mode_module + +_NormalizedRect = rect_module.NormalizedRect +_BaseOptions = base_options_module.BaseOptions +_EmbedderOptions = embedder_options_module.EmbedderOptions +_FloatEmbedding = embeddings_module.FloatEmbedding +_QuantizedEmbedding = embeddings_module.QuantizedEmbedding +_ClassificationEntry = embeddings_module.EmbeddingEntry +_Classifications = embeddings_module.Embeddings +_ClassificationResult = embeddings_module.EmbeddingResult +_Image = image_module.Image +_ImageEmbedder = image_embedder.ImageEmbedder +_ImageEmbedderOptions = image_embedder.ImageEmbedderOptions +_RUNNING_MODE = running_mode_module.VisionTaskRunningMode + +_MODEL_FILE = 'mobilenet_v3_small_100_224_embedder.tflite' +_IMAGE_FILE = 'burger.jpg' +_ALLOW_LIST = ['cheeseburger', 'guacamole'] +_DENY_LIST = ['cheeseburger'] +_SCORE_THRESHOLD = 0.5 +_MAX_RESULTS = 3 + + +class ModelFileType(enum.Enum): + FILE_CONTENT = 1 + FILE_NAME = 2 + + +class ImageClassifierTest(parameterized.TestCase): + + def setUp(self): + super().setUp() + self.test_image = _Image.create_from_file( + test_utils.get_test_data_path(_IMAGE_FILE)) + self.model_path = test_utils.get_test_data_path(_MODEL_FILE) + + @parameterized.parameters( + (ModelFileType.FILE_NAME, False, False), + (ModelFileType.FILE_CONTENT, False, False)) + def test_embed(self, model_file_type, l2_normalize, quantize): + # 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) + + # Performs image embedding extraction on the input. + image_result = embedder.embed(self.test_image) + + # TODO: Verify results. + + # Closes the embedder explicitly when the classifier is not used in + # a context. + embedder.close() + + +if __name__ == '__main__': + absltest.main() diff --git a/mediapipe/tasks/python/vision/BUILD b/mediapipe/tasks/python/vision/BUILD index 7ff818610..08c2709fc 100644 --- a/mediapipe/tasks/python/vision/BUILD +++ b/mediapipe/tasks/python/vision/BUILD @@ -36,3 +36,22 @@ py_library( "//mediapipe/tasks/python/vision/core:vision_task_running_mode", ], ) + +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/vision/image_embedder/proto:image_embedder_graph_options_py_pb2", + "//mediapipe/tasks/python/components/containers:embeddings", + "//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:vision_task_running_mode", + ], +) diff --git a/mediapipe/tasks/python/vision/image_embedder.py b/mediapipe/tasks/python/vision/image_embedder.py new file mode 100644 index 000000000..23ef492e5 --- /dev/null +++ b/mediapipe/tasks/python/vision/image_embedder.py @@ -0,0 +1,288 @@ +# 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.python._framework_bindings import task_runner as task_runner_module +from mediapipe.tasks.cc.vision.image_embedder.proto import image_embedder_graph_options_pb2 +from mediapipe.tasks.python.components.proto import embedder_options +from mediapipe.tasks.python.components.containers import embeddings as embeddings_module +from mediapipe.tasks.python.components.containers import rect as rect_module +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 vision_task_running_mode as running_mode_module + +_NormalizedRect = rect_module.NormalizedRect +_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 +_TaskRunner = task_runner_module.TaskRunner + +_EMBEDDING_RESULT_OUT_STREAM_NAME = 'embedding_result_out' +_EMBEDDING_RESULT_TAG = 'EMBEDDING_RESULT' +_IMAGE_IN_STREAM_NAME = 'image_in' +_IMAGE_OUT_STREAM_NAME = 'image_out' +_IMAGE_TAG = 'IMAGE' +_NORM_RECT_NAME = 'norm_rect_in' +_NORM_RECT_TAG = 'NORM_RECT' +_TASK_GRAPH_NAME = 'mediapipe.tasks.vision.image_embedder.ImageEmbedderGraph' +_MICRO_SECONDS_PER_MILLISECOND = 1000 + + +def _build_full_image_norm_rect() -> _NormalizedRect: + # Builds a NormalizedRect covering the entire image. + return _NormalizedRect(x_center=0.5, y_center=0.5, width=1, height=1) + + +@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[[embeddings_module.EmbeddingResult, 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 `ImageClassifier` 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 = packet_getter.get_proto( + output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) + + embedding_result = embeddings_module.EmbeddingResult([ + embeddings_module.Embeddings.create_from_pb2(embedding) + for embedding in embedding_result_proto.embeddings + ]) + image = packet_getter.get_image(output_packets[_IMAGE_OUT_STREAM_NAME]) + timestamp = output_packets[_IMAGE_OUT_STREAM_NAME].timestamp + options.result_callback(embedding_result, 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_NAME]), + ], + output_streams=[ + ':'.join([_EMBEDDING_RESULT_TAG, + _EMBEDDING_RESULT_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, + roi: Optional[_NormalizedRect] = None + ) -> embeddings_module.EmbeddingResult: + """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. + roi: The region of interest. + + Returns: + A 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. + """ + norm_rect = roi if roi is not None else _build_full_image_norm_rect() + output_packets = self._process_image_data({ + _IMAGE_IN_STREAM_NAME: packet_creator.create_image(image), + _NORM_RECT_NAME: packet_creator.create_proto(norm_rect.to_pb2())}) + embedding_result_proto = packet_getter.get_proto( + output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) + + return embeddings_module.EmbeddingResult([ + embeddings_module.Embeddings.create_from_pb2(embedding) + for embedding in embedding_result_proto.embeddings + ]) + + def embed_for_video( + self, image: image_module.Image, + timestamp_ms: int, + roi: Optional[_NormalizedRect] = None + ) -> embeddings_module.EmbeddingResult: + """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. + roi: The region of interest. + + Returns: + A 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. + """ + norm_rect = roi if roi is not None else _build_full_image_norm_rect() + output_packets = self._process_video_data({ + _IMAGE_IN_STREAM_NAME: packet_creator.create_image(image).at( + timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND), + _NORM_RECT_NAME: packet_creator.create_proto(norm_rect.to_pb2()).at( + timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND) + }) + embedding_result_proto = packet_getter.get_proto( + output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) + + return embeddings_module.EmbeddingResult([ + embeddings_module.Embeddings.create_from_pb2(embedding) + for embedding in embedding_result_proto.embeddings + ]) + + def embed_async( + self, + image: image_module.Image, + timestamp_ms: int, + roi: Optional[_NormalizedRect] = None + ) -> None: + """ Sends live image data to embedder, and 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: + - A 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. + roi: The region of interest. + + Raises: + ValueError: If the current input timestamp is smaller than what the image + embedder has already processed. + """ + norm_rect = roi if roi is not None else _build_full_image_norm_rect() + self._send_live_stream_data({ + _IMAGE_IN_STREAM_NAME: packet_creator.create_image(image).at( + timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND), + _NORM_RECT_NAME: packet_creator.create_proto(norm_rect.to_pb2()).at( + timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND) + }) From 492607152afb8ac697e7d84d17db0425e87a9469 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Thu, 3 Nov 2022 10:56:38 -0700 Subject: [PATCH 02/10] Reverted changes to rect --- .../python/components/containers/rect.py | 84 ++++++------------- 1 file changed, 26 insertions(+), 58 deletions(-) diff --git a/mediapipe/tasks/python/components/containers/rect.py b/mediapipe/tasks/python/components/containers/rect.py index aadb404db..4fdccbafb 100644 --- a/mediapipe/tasks/python/components/containers/rect.py +++ b/mediapipe/tasks/python/components/containers/rect.py @@ -19,78 +19,48 @@ from typing import Any, Optional from mediapipe.framework.formats import rect_pb2 from mediapipe.tasks.python.core.optional_dependencies import doc_controls -_RectProto = rect_pb2.Rect _NormalizedRectProto = rect_pb2.NormalizedRect @dataclasses.dataclass class Rect: - """A rectangle with rotation in image coordinates. + """A rectangle, used as part of detection results or as input region-of-interest. + + The coordinates are normalized wrt the image dimensions, i.e. generally in + [0,1] but they may exceed these bounds if describing a region overlapping the + image. The origin is on the top-left corner of the image. Attributes: - x_center : The X coordinate of the top-left corner, in pixels. - y_center : The Y coordinate of the top-left corner, in pixels. - width: The width of the rectangle, in pixels. - height: The height of the rectangle, in pixels. - rotation: Rotation angle is clockwise in radians. - rect_id: Optional unique id to help associate different rectangles to each - other. + left: The X coordinate of the left side of the rectangle. + top: The Y coordinate of the top of the rectangle. + right: The X coordinate of the right side of the rectangle. + bottom: The Y coordinate of the bottom of the rectangle. """ - x_center: int - y_center: int - width: int - height: int - rotation: Optional[float] = 0.0 - rect_id: Optional[int] = None - - @doc_controls.do_not_generate_docs - def to_pb2(self) -> _RectProto: - """Generates a Rect protobuf object.""" - return _RectProto( - x_center=self.x_center, - y_center=self.y_center, - width=self.width, - height=self.height, - ) - - @classmethod - @doc_controls.do_not_generate_docs - def create_from_pb2(cls, pb2_obj: _RectProto) -> 'Rect': - """Creates a `Rect` object from the given protobuf object.""" - return Rect( - x_center=pb2_obj.x_center, - y_center=pb2_obj.y_center, - width=pb2_obj.width, - height=pb2_obj.height) - - 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, Rect): - return False - - return self.to_pb2().__eq__(other.to_pb2()) + left: float + top: float + right: float + bottom: float @dataclasses.dataclass class NormalizedRect: - """A rectangle with rotation in normalized coordinates. The values of box - center location and size are within [0, 1]. + """A rectangle with rotation in normalized coordinates. + + Location of the center of the rectangle in image coordinates. The (0.0, 0.0) + point is at the (top, left) corner. + + The values of box center location and size are within [0, 1]. Attributes: - x_center : The X normalized coordinate of the top-left corner. - y_center : The Y normalized coordinate of the top-left corner. + x_center: The normalized X coordinate of the rectangle, in image + coordinates. + y_center: The normalized Y coordinate of the rectangle, in image + coordinates. width: The width of the rectangle. height: The height of the rectangle. rotation: Rotation angle is clockwise in radians. - rect_id: Optional unique id to help associate different rectangles to each + rect_id: Optional unique id to help associate different rectangles to each other. """ @@ -110,8 +80,7 @@ class NormalizedRect: width=self.width, height=self.height, rotation=self.rotation, - rect_id=self.rect_id - ) + rect_id=self.rect_id) @classmethod @doc_controls.do_not_generate_docs @@ -123,8 +92,7 @@ class NormalizedRect: width=pb2_obj.width, height=pb2_obj.height, rotation=pb2_obj.rotation, - rect_id=pb2_obj.rect_id - ) + rect_id=pb2_obj.rect_id) def __eq__(self, other: Any) -> bool: """Checks if this object is equal to the given object. From e2d50745ac0c0d84b56f1759c3e26165e5979159 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Thu, 3 Nov 2022 14:30:21 -0700 Subject: [PATCH 03/10] Revised image embedder implementation --- .../components/containers/embeddings.py | 16 +- .../components/proto/embedder_options.py | 6 +- mediapipe/tasks/python/components/utils/BUILD | 28 ++ .../tasks/python/components/utils/__init__.py | 13 + .../components/utils/cosine_similarity.py | 61 ++++ mediapipe/tasks/python/test/vision/BUILD | 2 + .../python/test/vision/image_embedder_test.py | 316 ++++++++++++++++-- mediapipe/tasks/python/vision/BUILD | 1 + .../tasks/python/vision/image_embedder.py | 69 ++-- 9 files changed, 456 insertions(+), 56 deletions(-) create mode 100644 mediapipe/tasks/python/components/utils/BUILD create mode 100644 mediapipe/tasks/python/components/utils/__init__.py create mode 100644 mediapipe/tasks/python/components/utils/cosine_similarity.py diff --git a/mediapipe/tasks/python/components/containers/embeddings.py b/mediapipe/tasks/python/components/containers/embeddings.py index 21f53670c..c1185c84f 100644 --- a/mediapipe/tasks/python/components/containers/embeddings.py +++ b/mediapipe/tasks/python/components/containers/embeddings.py @@ -131,18 +131,14 @@ class EmbeddingEntry: cls, pb2_obj: _EmbeddingEntryProto) -> 'EmbeddingEntry': """Creates a `EmbeddingEntry` object from the given protobuf object.""" - if pb2_obj.float_embedding: - return EmbeddingEntry( - embedding=np.array(pb2_obj.float_embedding.values, dtype=float)) - - elif pb2_obj.quantized_embedding: - return EmbeddingEntry( - embedding=np.array(bytearray(pb2_obj.quantized_embedding.values), - dtype=np.uint8)) + quantized_embedding = np.array( + bytearray(pb2_obj.quantized_embedding.values)) + float_embedding = np.array(pb2_obj.float_embedding.values, dtype=float) + if len(quantized_embedding) == 0: + return EmbeddingEntry(embedding=float_embedding) else: - raise ValueError("Either float_embedding or quantized_embedding must " - "exist.") + return EmbeddingEntry(embedding=quantized_embedding) def __eq__(self, other: Any) -> bool: """Checks if this object is equal to the given object. diff --git a/mediapipe/tasks/python/components/proto/embedder_options.py b/mediapipe/tasks/python/components/proto/embedder_options.py index 3c257b976..49bcfb985 100644 --- a/mediapipe/tasks/python/components/proto/embedder_options.py +++ b/mediapipe/tasks/python/components/proto/embedder_options.py @@ -45,8 +45,7 @@ class EmbedderOptions: """Generates a EmbedderOptions protobuf object.""" return _EmbedderOptionsProto( l2_normalize=self.l2_normalize, - quantize=self.quantize - ) + quantize=self.quantize) @classmethod @doc_controls.do_not_generate_docs @@ -54,8 +53,7 @@ class EmbedderOptions: """Creates a `EmbedderOptions` object from the given protobuf object.""" return EmbedderOptions( l2_normalize=pb2_obj.l2_normalize, - quantize=pb2_obj.quantize - ) + quantize=pb2_obj.quantize) def __eq__(self, other: Any) -> bool: """Checks if this object is equal to the given object. diff --git a/mediapipe/tasks/python/components/utils/BUILD b/mediapipe/tasks/python/components/utils/BUILD new file mode 100644 index 000000000..7ec01a034 --- /dev/null +++ b/mediapipe/tasks/python/components/utils/BUILD @@ -0,0 +1,28 @@ +# 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. + +package(default_visibility = ["//mediapipe/tasks:internal"]) + +licenses(["notice"]) + +py_library( + name = "cosine_similarity", + srcs = ["cosine_similarity.py"], + deps = [ + "//mediapipe/tasks/python/components/containers:embeddings", + "//mediapipe/tasks/python/components/proto:embedder_options", + ], +) diff --git a/mediapipe/tasks/python/components/utils/__init__.py b/mediapipe/tasks/python/components/utils/__init__.py new file mode 100644 index 000000000..65c1214af --- /dev/null +++ b/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. diff --git a/mediapipe/tasks/python/components/utils/cosine_similarity.py b/mediapipe/tasks/python/components/utils/cosine_similarity.py new file mode 100644 index 000000000..8723a55eb --- /dev/null +++ b/mediapipe/tasks/python/components/utils/cosine_similarity.py @@ -0,0 +1,61 @@ +# 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 embeddings +from mediapipe.tasks.python.components.proto import embedder_options + +_EmbeddingEntry = embeddings.EmbeddingEntry +_EmbedderOptions = embedder_options.EmbedderOptions + + +def _compute_cosine_similarity(u, v): + 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: _EmbeddingEntry, v: _EmbeddingEntry) -> 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. + """ + 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.") diff --git a/mediapipe/tasks/python/test/vision/BUILD b/mediapipe/tasks/python/test/vision/BUILD index 6adab6807..9fee8a023 100644 --- a/mediapipe/tasks/python/test/vision/BUILD +++ b/mediapipe/tasks/python/test/vision/BUILD @@ -84,11 +84,13 @@ py_test( deps = [ "//mediapipe/python:_framework_bindings", "//mediapipe/tasks/python/components/proto:embedder_options", + "//mediapipe/tasks/python/components/utils:cosine_similarity", "//mediapipe/tasks/python/components/containers:embeddings", "//mediapipe/tasks/python/components/containers:rect", "//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", ], ) diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index 8ddf3c992..4f109ea29 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -14,6 +14,7 @@ """Tests for image embedder.""" import enum +import os from unittest import mock import numpy as np @@ -23,31 +24,32 @@ from absl.testing import parameterized from mediapipe.python._framework_bindings import image as image_module from mediapipe.tasks.python.components.proto import embedder_options as embedder_options_module from mediapipe.tasks.python.components.containers import embeddings as embeddings_module -from mediapipe.tasks.python.components.containers import rect as rect_module +from mediapipe.tasks.python.components.containers import rect 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 -_NormalizedRect = rect_module.NormalizedRect +_Rect = rect.Rect _BaseOptions = base_options_module.BaseOptions _EmbedderOptions = embedder_options_module.EmbedderOptions _FloatEmbedding = embeddings_module.FloatEmbedding _QuantizedEmbedding = embeddings_module.QuantizedEmbedding -_ClassificationEntry = embeddings_module.EmbeddingEntry -_Classifications = embeddings_module.Embeddings -_ClassificationResult = embeddings_module.EmbeddingResult +_EmbeddingEntry = embeddings_module.EmbeddingEntry +_Embeddings = embeddings_module.Embeddings +_EmbeddingResult = embeddings_module.EmbeddingResult _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' -_IMAGE_FILE = 'burger.jpg' -_ALLOW_LIST = ['cheeseburger', 'guacamole'] -_DENY_LIST = ['cheeseburger'] -_SCORE_THRESHOLD = 0.5 -_MAX_RESULTS = 3 +_BURGER_IMAGE_FILE = 'burger.jpg' +_BURGER_CROPPED_IMAGE_FILE = 'burger_crop.jpg' +_TEST_DATA_DIR = 'mediapipe/tasks/testdata/vision' +_SIMILARITY_TOLERANCE = 1e-6 class ModelFileType(enum.Enum): @@ -55,18 +57,55 @@ class ModelFileType(enum.Enum): FILE_NAME = 2 -class ImageClassifierTest(parameterized.TestCase): +class ImageEmbedderTest(parameterized.TestCase): def setUp(self): super().setUp() self.test_image = _Image.create_from_file( - test_utils.get_test_data_path(_IMAGE_FILE)) - self.model_path = test_utils.get_test_data_path(_MODEL_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 _check_cosine_similarity(self, result0, result1, quantize, + expected_similarity): + # Checks head_index and head_name. + self.assertEqual(result0.embeddings[0].head_index, 0) + self.assertEqual(result1.embeddings[0].head_index, 0) + self.assertEqual(result0.embeddings[0].head_name, 'feature') + self.assertEqual(result1.embeddings[0].head_name, 'feature') + + # Check embedding sizes. + def _check_embedding_size(result): + self.assertLen(result.embeddings, 1) + embedding_entry = result.embeddings[0].entries[0] + self.assertLen(embedding_entry.embedding, 1024) + if quantize: + self.assertEqual(embedding_entry.embedding.dtype, np.uint8) + else: + self.assertEqual(embedding_entry.embedding.dtype, float) + + # Checks results sizes. + _check_embedding_size(result0) + _check_embedding_size(result1) + + # Checks cosine similarity. + similarity = _ImageEmbedder.cosine_similarity( + result0.embeddings[0].entries[0], result1.embeddings[0].entries[0]) + self.assertAlmostEqual(similarity, expected_similarity, + delta=_SIMILARITY_TOLERANCE) @parameterized.parameters( - (ModelFileType.FILE_NAME, False, False), - (ModelFileType.FILE_CONTENT, False, False)) - def test_embed(self, model_file_type, l2_normalize, quantize): + (False, False, False, ModelFileType.FILE_NAME, 0.925519, -0.2101883), + (True, False, False, ModelFileType.FILE_NAME, 0.925519, -0.0142344), + # (False, True, False, ModelFileType.FILE_NAME, 0.926791, 229), + (False, False, True, ModelFileType.FILE_CONTENT, 0.999931, -0.195062) + ) + def test_embed(self, l2_normalize, quantize, with_roi, model_file_type, + expected_similarity, expected_first_value): # Creates embedder. if model_file_type is ModelFileType.FILE_NAME: base_options = _BaseOptions(model_asset_path=self.model_path) @@ -84,15 +123,254 @@ class ImageClassifierTest(parameterized.TestCase): base_options=base_options, embedder_options=embedder_options) embedder = _ImageEmbedder.create_from_options(options) - # Performs image embedding extraction on the input. - image_result = embedder.embed(self.test_image) + 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) - # TODO: Verify results. + # Extracts both embeddings. + image_result = embedder.embed(self.test_image, image_processing_options) + crop_result = embedder.embed(self.test_cropped_image) + # Check embedding value. + self.assertAlmostEqual(image_result.embeddings[0].entries[0].embedding[0], + expected_first_value) + + # Checks cosine similarity. + self._check_cosine_similarity(image_result, crop_result, quantize, + expected_similarity) # Closes the embedder explicitly when the classifier 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, quantize, + 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, quantize=False, + 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, quantize=False, + 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_classify_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_classify_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: _EmbeddingResult, output_image: _Image, + timestamp_ms: int): + # Checks cosine similarity. + self._check_cosine_similarity(result, crop_result, quantize=False, + 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_classify_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: _EmbeddingResult, output_image: _Image, + timestamp_ms: int): + # Checks cosine similarity. + self._check_cosine_similarity(result, crop_result, quantize=False, + 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() diff --git a/mediapipe/tasks/python/vision/BUILD b/mediapipe/tasks/python/vision/BUILD index 96cf6288d..9d406be28 100644 --- a/mediapipe/tasks/python/vision/BUILD +++ b/mediapipe/tasks/python/vision/BUILD @@ -118,6 +118,7 @@ py_library( "//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", ], ) diff --git a/mediapipe/tasks/python/vision/image_embedder.py b/mediapipe/tasks/python/vision/image_embedder.py index 23ef492e5..e287593f5 100644 --- a/mediapipe/tasks/python/vision/image_embedder.py +++ b/mediapipe/tasks/python/vision/image_embedder.py @@ -23,20 +23,21 @@ from mediapipe.python._framework_bindings import packet as packet_module from mediapipe.python._framework_bindings import task_runner as task_runner_module from mediapipe.tasks.cc.vision.image_embedder.proto import image_embedder_graph_options_pb2 from mediapipe.tasks.python.components.proto import embedder_options +from mediapipe.tasks.python.components.utils import cosine_similarity from mediapipe.tasks.python.components.containers import embeddings as embeddings_module -from mediapipe.tasks.python.components.containers import rect as rect_module 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 -_NormalizedRect = rect_module.NormalizedRect _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 _TaskRunner = task_runner_module.TaskRunner _EMBEDDING_RESULT_OUT_STREAM_NAME = 'embedding_result_out' @@ -44,17 +45,12 @@ _EMBEDDING_RESULT_TAG = 'EMBEDDING_RESULT' _IMAGE_IN_STREAM_NAME = 'image_in' _IMAGE_OUT_STREAM_NAME = 'image_out' _IMAGE_TAG = 'IMAGE' -_NORM_RECT_NAME = 'norm_rect_in' +_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 -def _build_full_image_norm_rect() -> _NormalizedRect: - # Builds a NormalizedRect covering the entire image. - return _NormalizedRect(x_center=0.5, y_center=0.5, width=1, height=1) - - @dataclasses.dataclass class ImageEmbedderOptions: """Options for the image embedder task. @@ -75,6 +71,8 @@ class ImageEmbedderOptions: """ base_options: _BaseOptions running_mode: _RunningMode = _RunningMode.IMAGE + l2_normalize: Optional[bool] = None + quantize: Optional[bool] = None embedder_options: _EmbedderOptions = _EmbedderOptions() result_callback: Optional[ Callable[[embeddings_module.EmbeddingResult, image_module.Image, @@ -157,7 +155,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): task_graph=_TASK_GRAPH_NAME, input_streams=[ ':'.join([_IMAGE_TAG, _IMAGE_IN_STREAM_NAME]), - ':'.join([_NORM_RECT_TAG, _NORM_RECT_NAME]), + ':'.join([_NORM_RECT_TAG, _NORM_RECT_STREAM_NAME]), ], output_streams=[ ':'.join([_EMBEDDING_RESULT_TAG, @@ -174,7 +172,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): def embed( self, image: image_module.Image, - roi: Optional[_NormalizedRect] = None + image_processing_options: Optional[_ImageProcessingOptions] = None ) -> embeddings_module.EmbeddingResult: """Performs image embedding extraction on the provided MediaPipe Image. Extraction is performed on the region of interest specified by the `roi` @@ -182,7 +180,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): Args: image: MediaPipe Image. - roi: The region of interest. + image_processing_options: Options for image processing. Returns: A embedding result object that contains a list of embeddings. @@ -191,10 +189,11 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): ValueError: If any of the input arguments is invalid. RuntimeError: If image embedder failed to run. """ - norm_rect = roi if roi is not None else _build_full_image_norm_rect() + 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_NAME: packet_creator.create_proto(norm_rect.to_pb2())}) + _NORM_RECT_STREAM_NAME: packet_creator.create_proto( + normalized_rect.to_pb2())}) embedding_result_proto = packet_getter.get_proto( output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) @@ -206,7 +205,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): def embed_for_video( self, image: image_module.Image, timestamp_ms: int, - roi: Optional[_NormalizedRect] = None + image_processing_options: Optional[_ImageProcessingOptions] = None ) -> embeddings_module.EmbeddingResult: """Performs image embedding extraction on the provided video frames. Extraction is performed on the region of interested specified by the `roi` @@ -220,7 +219,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): Args: image: MediaPipe Image. timestamp_ms: The timestamp of the input video frame in milliseconds. - roi: The region of interest. + image_processing_options: Options for image processing. Returns: A embedding result object that contains a list of embeddings. @@ -229,12 +228,13 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): ValueError: If any of the input arguments is invalid. RuntimeError: If image embedder failed to run. """ - norm_rect = roi if roi is not None else _build_full_image_norm_rect() + 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_NAME: packet_creator.create_proto(norm_rect.to_pb2()).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 = packet_getter.get_proto( output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) @@ -248,7 +248,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): self, image: image_module.Image, timestamp_ms: int, - roi: Optional[_NormalizedRect] = None + image_processing_options: Optional[_ImageProcessingOptions] = None ) -> None: """ Sends live image data to embedder, and the results will be available via the "result_callback" provided in the ImageEmbedderOptions. Embedding @@ -273,16 +273,39 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): Args: image: MediaPipe Image. timestamp_ms: The timestamp of the input image in milliseconds. - roi: The region of interest. + image_processing_options: Options for image processing. Raises: ValueError: If the current input timestamp is smaller than what the image embedder has already processed. """ - norm_rect = roi if roi is not None else _build_full_image_norm_rect() + 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_NAME: packet_creator.create_proto(norm_rect.to_pb2()).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) }) + + @staticmethod + def cosine_similarity(u: embeddings_module.EmbeddingEntry, + v: embeddings_module.EmbeddingEntry) -> float: + """Utility function to compute cosine similarity [1] 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 a + 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) From 664d9c49e7cf120b753fb572c936be569e93b217 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Mon, 7 Nov 2022 13:59:07 -0800 Subject: [PATCH 04/10] Revised image embedder implementation --- mediapipe/python/BUILD | 1 - .../components/containers/embeddings.py | 99 ++++++------------- .../tasks/python/components/processors/BUILD | 9 ++ .../{proto => processors}/embedder_options.py | 2 +- mediapipe/tasks/python/components/proto/BUILD | 28 ------ .../tasks/python/components/proto/__init__.py | 13 --- mediapipe/tasks/python/components/utils/BUILD | 2 +- .../components/utils/cosine_similarity.py | 16 +-- mediapipe/tasks/python/test/vision/BUILD | 2 +- .../python/test/vision/image_embedder_test.py | 17 ++-- mediapipe/tasks/python/vision/BUILD | 1 + .../tasks/python/vision/image_embedder.py | 39 ++++---- 12 files changed, 79 insertions(+), 150 deletions(-) rename mediapipe/tasks/python/components/{proto => processors}/embedder_options.py (96%) delete mode 100644 mediapipe/tasks/python/components/proto/BUILD delete mode 100644 mediapipe/tasks/python/components/proto/__init__.py diff --git a/mediapipe/python/BUILD b/mediapipe/python/BUILD index 2423370e6..0f049f305 100644 --- a/mediapipe/python/BUILD +++ b/mediapipe/python/BUILD @@ -92,7 +92,6 @@ cc_library( "//mediapipe/tasks/cc/vision/image_segmenter:image_segmenter_graph", "//mediapipe/tasks/cc/vision/object_detector:object_detector_graph", "//mediapipe/tasks/cc/vision/image_embedder:image_embedder_graph", - ], ] + select({ # TODO: Build text_classifier_graph on Windows. "//mediapipe:windows": [], diff --git a/mediapipe/tasks/python/components/containers/embeddings.py b/mediapipe/tasks/python/components/containers/embeddings.py index c1185c84f..3a024079a 100644 --- a/mediapipe/tasks/python/components/containers/embeddings.py +++ b/mediapipe/tasks/python/components/containers/embeddings.py @@ -22,8 +22,7 @@ from mediapipe.tasks.python.core.optional_dependencies import doc_controls _FloatEmbeddingProto = embeddings_pb2.FloatEmbedding _QuantizedEmbeddingProto = embeddings_pb2.QuantizedEmbedding -_EmbeddingEntryProto = embeddings_pb2.EmbeddingEntry -_EmbeddingsProto = embeddings_pb2.Embeddings +_EmbeddingProto = embeddings_pb2.Embedding _EmbeddingResultProto = embeddings_pb2.EmbeddingResult @@ -99,28 +98,34 @@ class QuantizedEmbedding: @dataclasses.dataclass -class EmbeddingEntry: - """Floating-point or scalar-quantized embedding with an optional timestamp. +class Embedding: + """Embedding result for a given embedder head. Attributes: embedding: The actual embedding, either floating-point or scalar-quantized. - timestamp_ms: The optional timestamp (in milliseconds) associated to the - embedding entry. This is useful for time series use cases, e.g. audio - embedding. + 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 - timestamp_ms: Optional[int] = None + head_index: int + head_name: str @doc_controls.do_not_generate_docs - def to_pb2(self) -> _EmbeddingEntryProto: - """Generates a EmbeddingEntry protobuf object.""" + def to_pb2(self) -> _EmbeddingProto: + """Generates a Embedding protobuf object.""" if self.embedding.dtype == float: - return _EmbeddingEntryProto(float_embedding=self.embedding) + return _EmbeddingProto(float_embedding=self.embedding, + head_index=self.head_index, + head_name=self.head_name) elif self.embedding.dtype == np.uint8: - return _EmbeddingEntryProto(quantized_embedding=bytes(self.embedding)) + return _EmbeddingProto(quantized_embedding=bytes(self.embedding), + head_index=self.head_index, + head_name=self.head_name) else: raise ValueError("Invalid dtype. Only float and np.uint8 are supported.") @@ -128,17 +133,21 @@ class EmbeddingEntry: @classmethod @doc_controls.do_not_generate_docs def create_from_pb2( - cls, pb2_obj: _EmbeddingEntryProto) -> 'EmbeddingEntry': - """Creates a `EmbeddingEntry` object from the given protobuf object.""" + 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 len(quantized_embedding) == 0: - return EmbeddingEntry(embedding=float_embedding) + return Embedding(embedding=float_embedding, + head_index=pb2_obj.head_index, + head_name=pb2_obj.head_name) else: - return EmbeddingEntry(embedding=quantized_embedding) + return Embedding(embedding=quantized_embedding, + head_index=pb2_obj.head_index, + head_name=pb2_obj.head_name) def __eq__(self, other: Any) -> bool: """Checks if this object is equal to the given object. @@ -147,55 +156,7 @@ class EmbeddingEntry: Returns: True if the objects are equal. """ - if not isinstance(other, EmbeddingEntry): - return False - - return self.to_pb2().__eq__(other.to_pb2()) - - -@dataclasses.dataclass -class Embeddings: - """Embeddings for a given embedder head. - Attributes: - entries: A list of `ClassificationEntry` objects. - 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. - """ - - entries: List[EmbeddingEntry] - head_index: int - head_name: str - - @doc_controls.do_not_generate_docs - def to_pb2(self) -> _EmbeddingsProto: - """Generates a Embeddings protobuf object.""" - return _EmbeddingsProto( - entries=[entry.to_pb2() for entry in self.entries], - head_index=self.head_index, - head_name=self.head_name) - - @classmethod - @doc_controls.do_not_generate_docs - def create_from_pb2(cls, pb2_obj: _EmbeddingsProto) -> 'Embeddings': - """Creates a `Embeddings` object from the given protobuf object.""" - return Embeddings( - entries=[ - EmbeddingEntry.create_from_pb2(entry) - for entry in pb2_obj.entries - ], - head_index=pb2_obj.head_index, - head_name=pb2_obj.head_name) - - 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, Embeddings): + if not isinstance(other, Embedding): return False return self.to_pb2().__eq__(other.to_pb2()) @@ -203,12 +164,12 @@ class Embeddings: @dataclasses.dataclass class EmbeddingResult: - """Contains one set of results per embedder head. + """Embedding results for a given embedder model. Attributes: - embeddings: A list of `Embeddings` objects. + embeddings: A list of `Embedding` objects. """ - embeddings: List[Embeddings] + embeddings: List[Embedding] @doc_controls.do_not_generate_docs def to_pb2(self) -> _EmbeddingResultProto: @@ -225,7 +186,7 @@ class EmbeddingResult: """Creates a `EmbeddingResult` object from the given protobuf object.""" return EmbeddingResult( embeddings=[ - Embeddings.create_from_pb2(embedding) + Embedding.create_from_pb2(embedding) for embedding in pb2_obj.embeddings ]) diff --git a/mediapipe/tasks/python/components/processors/BUILD b/mediapipe/tasks/python/components/processors/BUILD index f87a579b0..eef368db0 100644 --- a/mediapipe/tasks/python/components/processors/BUILD +++ b/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", + ], +) diff --git a/mediapipe/tasks/python/components/proto/embedder_options.py b/mediapipe/tasks/python/components/processors/embedder_options.py similarity index 96% rename from mediapipe/tasks/python/components/proto/embedder_options.py rename to mediapipe/tasks/python/components/processors/embedder_options.py index 49bcfb985..dcd316dcd 100644 --- a/mediapipe/tasks/python/components/proto/embedder_options.py +++ b/mediapipe/tasks/python/components/processors/embedder_options.py @@ -16,7 +16,7 @@ import dataclasses from typing import Any, Optional -from mediapipe.tasks.cc.components.proto import embedder_options_pb2 +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 diff --git a/mediapipe/tasks/python/components/proto/BUILD b/mediapipe/tasks/python/components/proto/BUILD deleted file mode 100644 index 973f150ca..000000000 --- a/mediapipe/tasks/python/components/proto/BUILD +++ /dev/null @@ -1,28 +0,0 @@ -# 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. - -package(default_visibility = ["//mediapipe/tasks:internal"]) - -licenses(["notice"]) - -py_library( - name = "embedder_options", - srcs = ["embedder_options.py"], - deps = [ - "//mediapipe/tasks/cc/components/proto:embedder_options_py_pb2", - "//mediapipe/tasks/python/core:optional_dependencies", - ], -) diff --git a/mediapipe/tasks/python/components/proto/__init__.py b/mediapipe/tasks/python/components/proto/__init__.py deleted file mode 100644 index 65c1214af..000000000 --- a/mediapipe/tasks/python/components/proto/__init__.py +++ /dev/null @@ -1,13 +0,0 @@ -# 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. diff --git a/mediapipe/tasks/python/components/utils/BUILD b/mediapipe/tasks/python/components/utils/BUILD index 7ec01a034..6d00bb31a 100644 --- a/mediapipe/tasks/python/components/utils/BUILD +++ b/mediapipe/tasks/python/components/utils/BUILD @@ -23,6 +23,6 @@ py_library( srcs = ["cosine_similarity.py"], deps = [ "//mediapipe/tasks/python/components/containers:embeddings", - "//mediapipe/tasks/python/components/proto:embedder_options", + "//mediapipe/tasks/python/components/processors:embedder_options", ], ) diff --git a/mediapipe/tasks/python/components/utils/cosine_similarity.py b/mediapipe/tasks/python/components/utils/cosine_similarity.py index 8723a55eb..616f2651f 100644 --- a/mediapipe/tasks/python/components/utils/cosine_similarity.py +++ b/mediapipe/tasks/python/components/utils/cosine_similarity.py @@ -16,9 +16,9 @@ import numpy as np from mediapipe.tasks.python.components.containers import embeddings -from mediapipe.tasks.python.components.proto import embedder_options +from mediapipe.tasks.python.components.processors import embedder_options -_EmbeddingEntry = embeddings.EmbeddingEntry +_Embedding = embeddings.Embedding _EmbedderOptions = embedder_options.EmbedderOptions @@ -36,15 +36,15 @@ def _compute_cosine_similarity(u, v): return np.dot(u.embedding, v.embedding.T) / (norm_u * norm_v) -def cosine_similarity(u: _EmbeddingEntry, v: _EmbeddingEntry) -> 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 +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 entry. - v: An embedding entry. + u: An embedding. + v: An embedding. """ if len(u.embedding) != len(v.embedding): raise ValueError(f"Cannot compute cosine similarity between embeddings " diff --git a/mediapipe/tasks/python/test/vision/BUILD b/mediapipe/tasks/python/test/vision/BUILD index 9fee8a023..5fd396161 100644 --- a/mediapipe/tasks/python/test/vision/BUILD +++ b/mediapipe/tasks/python/test/vision/BUILD @@ -83,7 +83,7 @@ py_test( ], deps = [ "//mediapipe/python:_framework_bindings", - "//mediapipe/tasks/python/components/proto:embedder_options", + "//mediapipe/tasks/python/components/processors:embedder_options", "//mediapipe/tasks/python/components/utils:cosine_similarity", "//mediapipe/tasks/python/components/containers:embeddings", "//mediapipe/tasks/python/components/containers:rect", diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index 4f109ea29..7bcf48d71 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -22,7 +22,7 @@ from absl.testing import absltest from absl.testing import parameterized from mediapipe.python._framework_bindings import image as image_module -from mediapipe.tasks.python.components.proto import embedder_options as embedder_options_module +from mediapipe.tasks.python.components.processors import embedder_options as embedder_options_module from mediapipe.tasks.python.components.containers import embeddings as embeddings_module from mediapipe.tasks.python.components.containers import rect from mediapipe.tasks.python.core import base_options as base_options_module @@ -36,8 +36,7 @@ _BaseOptions = base_options_module.BaseOptions _EmbedderOptions = embedder_options_module.EmbedderOptions _FloatEmbedding = embeddings_module.FloatEmbedding _QuantizedEmbedding = embeddings_module.QuantizedEmbedding -_EmbeddingEntry = embeddings_module.EmbeddingEntry -_Embeddings = embeddings_module.Embeddings +_Embedding = embeddings_module.Embedding _EmbeddingResult = embeddings_module.EmbeddingResult _Image = image_module.Image _ImageEmbedder = image_embedder.ImageEmbedder @@ -81,12 +80,12 @@ class ImageEmbedderTest(parameterized.TestCase): # Check embedding sizes. def _check_embedding_size(result): self.assertLen(result.embeddings, 1) - embedding_entry = result.embeddings[0].entries[0] - self.assertLen(embedding_entry.embedding, 1024) + embedding_result = result.embeddings[0] + self.assertLen(embedding_result.embedding, 1024) if quantize: - self.assertEqual(embedding_entry.embedding.dtype, np.uint8) + self.assertEqual(embedding_result.embedding.dtype, np.uint8) else: - self.assertEqual(embedding_entry.embedding.dtype, float) + self.assertEqual(embedding_result.embedding.dtype, float) # Checks results sizes. _check_embedding_size(result0) @@ -94,7 +93,7 @@ class ImageEmbedderTest(parameterized.TestCase): # Checks cosine similarity. similarity = _ImageEmbedder.cosine_similarity( - result0.embeddings[0].entries[0], result1.embeddings[0].entries[0]) + result0.embeddings[0], result1.embeddings[0]) self.assertAlmostEqual(similarity, expected_similarity, delta=_SIMILARITY_TOLERANCE) @@ -134,7 +133,7 @@ class ImageEmbedderTest(parameterized.TestCase): crop_result = embedder.embed(self.test_cropped_image) # Check embedding value. - self.assertAlmostEqual(image_result.embeddings[0].entries[0].embedding[0], + self.assertAlmostEqual(image_result.embeddings[0].embedding[0], expected_first_value) # Checks cosine similarity. diff --git a/mediapipe/tasks/python/vision/BUILD b/mediapipe/tasks/python/vision/BUILD index 9d406be28..1d60fa5b5 100644 --- a/mediapipe/tasks/python/vision/BUILD +++ b/mediapipe/tasks/python/vision/BUILD @@ -114,6 +114,7 @@ py_library( "//mediapipe/python:packet_getter", "//mediapipe/tasks/cc/vision/image_embedder/proto:image_embedder_graph_options_py_pb2", "//mediapipe/tasks/python/components/containers:embeddings", + "//mediapipe/tasks/python/components/processors:classifier_options", "//mediapipe/tasks/python/core:base_options", "//mediapipe/tasks/python/core:optional_dependencies", "//mediapipe/tasks/python/core:task_info", diff --git a/mediapipe/tasks/python/vision/image_embedder.py b/mediapipe/tasks/python/vision/image_embedder.py index e287593f5..2851970d8 100644 --- a/mediapipe/tasks/python/vision/image_embedder.py +++ b/mediapipe/tasks/python/vision/image_embedder.py @@ -22,7 +22,7 @@ from mediapipe.python._framework_bindings import image as image_module from mediapipe.python._framework_bindings import packet as packet_module from mediapipe.python._framework_bindings import task_runner as task_runner_module from mediapipe.tasks.cc.vision.image_embedder.proto import image_embedder_graph_options_pb2 -from mediapipe.tasks.python.components.proto import embedder_options +from mediapipe.tasks.python.components.processors import embedder_options from mediapipe.tasks.python.components.utils import cosine_similarity from mediapipe.tasks.python.components.containers import embeddings as embeddings_module from mediapipe.tasks.python.core import base_options as base_options_module @@ -32,6 +32,7 @@ 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 = embeddings_module.EmbeddingResult _BaseOptions = base_options_module.BaseOptions _ImageEmbedderGraphOptionsProto = image_embedder_graph_options_pb2.ImageEmbedderGraphOptions _EmbedderOptions = embedder_options.EmbedderOptions @@ -40,8 +41,8 @@ _TaskInfo = task_info_module.TaskInfo _ImageProcessingOptions = image_processing_options_module.ImageProcessingOptions _TaskRunner = task_runner_module.TaskRunner -_EMBEDDING_RESULT_OUT_STREAM_NAME = 'embedding_result_out' -_EMBEDDING_RESULT_TAG = 'EMBEDDING_RESULT' +_EMBEDDINGS_OUT_STREAM_NAME = 'embeddings_out' +_EMBEDDINGS_TAG = 'EMBEDDINGS' _IMAGE_IN_STREAM_NAME = 'image_in' _IMAGE_OUT_STREAM_NAME = 'image_out' _IMAGE_TAG = 'IMAGE' @@ -140,15 +141,15 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): if output_packets[_IMAGE_OUT_STREAM_NAME].is_empty(): return embedding_result_proto = packet_getter.get_proto( - output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) + output_packets[_EMBEDDINGS_OUT_STREAM_NAME]) - embedding_result = embeddings_module.EmbeddingResult([ - embeddings_module.Embeddings.create_from_pb2(embedding) + embeddings = embeddings_module.EmbeddingResult([ + embeddings_module.Embedding.create_from_pb2(embedding) for embedding in embedding_result_proto.embeddings ]) image = packet_getter.get_image(output_packets[_IMAGE_OUT_STREAM_NAME]) timestamp = output_packets[_IMAGE_OUT_STREAM_NAME].timestamp - options.result_callback(embedding_result, image, + options.result_callback(embeddings, image, timestamp.value // _MICRO_SECONDS_PER_MILLISECOND) task_info = _TaskInfo( @@ -158,8 +159,8 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): ':'.join([_NORM_RECT_TAG, _NORM_RECT_STREAM_NAME]), ], output_streams=[ - ':'.join([_EMBEDDING_RESULT_TAG, - _EMBEDDING_RESULT_OUT_STREAM_NAME]), + ':'.join([_EMBEDDINGS_TAG, + _EMBEDDINGS_OUT_STREAM_NAME]), ':'.join([_IMAGE_TAG, _IMAGE_OUT_STREAM_NAME]) ], task_options=options) @@ -173,7 +174,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): self, image: image_module.Image, image_processing_options: Optional[_ImageProcessingOptions] = None - ) -> embeddings_module.EmbeddingResult: + ) -> _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. @@ -195,18 +196,18 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): _NORM_RECT_STREAM_NAME: packet_creator.create_proto( normalized_rect.to_pb2())}) embedding_result_proto = packet_getter.get_proto( - output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) + output_packets[_EMBEDDINGS_OUT_STREAM_NAME]) return embeddings_module.EmbeddingResult([ - embeddings_module.Embeddings.create_from_pb2(embedding) - for embedding in embedding_result_proto.embeddings + embeddings_module.Embedding.create_from_pb2(embedding) + for embedding in embedding_result_proto.embeddings ]) def embed_for_video( self, image: image_module.Image, timestamp_ms: int, image_processing_options: Optional[_ImageProcessingOptions] = None - ) -> embeddings_module.EmbeddingResult: + ) -> _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. @@ -237,11 +238,11 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND) }) embedding_result_proto = packet_getter.get_proto( - output_packets[_EMBEDDING_RESULT_OUT_STREAM_NAME]) + output_packets[_EMBEDDINGS_OUT_STREAM_NAME]) return embeddings_module.EmbeddingResult([ - embeddings_module.Embeddings.create_from_pb2(embedding) - for embedding in embedding_result_proto.embeddings + embeddings_module.Embedding.create_from_pb2(embedding) + for embedding in embedding_result_proto.embeddings ]) def embed_async( @@ -289,8 +290,8 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): }) @staticmethod - def cosine_similarity(u: embeddings_module.EmbeddingEntry, - v: embeddings_module.EmbeddingEntry) -> float: + def cosine_similarity(u: embeddings_module.Embedding, + v: embeddings_module.Embedding) -> float: """Utility function to compute cosine similarity [1] 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 a From dd6fdedd5f4f280a9cd99c43f22e66aaa2869df5 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Tue, 8 Nov 2022 23:15:58 -0800 Subject: [PATCH 05/10] Added some sanity tests --- .../python/test/vision/image_embedder_test.py | 28 +++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index 7bcf48d71..624a9d99a 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -69,6 +69,34 @@ class ImageEmbedderTest(parameterized.TestCase): 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_cosine_similarity(self, result0, result1, quantize, expected_similarity): # Checks head_index and head_name. From d327b3649df98a55049bd2d5e216584bf040a392 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Tue, 8 Nov 2022 23:47:58 -0800 Subject: [PATCH 06/10] Fixed some typos in methods and comments --- mediapipe/tasks/python/test/vision/image_embedder_test.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index 624a9d99a..0dfce91c2 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -167,7 +167,7 @@ class ImageEmbedderTest(parameterized.TestCase): # Checks cosine similarity. self._check_cosine_similarity(image_result, crop_result, quantize, expected_similarity) - # Closes the embedder explicitly when the classifier is not used in + # Closes the embedder explicitly when the embedder is not used in # a context. embedder.close() @@ -315,7 +315,7 @@ class ImageEmbedderTest(parameterized.TestCase): r'not initialized with the image mode'): embedder.embed(self.test_image) - def test_calling_classify_for_video_in_live_stream_mode(self): + 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, @@ -325,7 +325,7 @@ class ImageEmbedderTest(parameterized.TestCase): r'not initialized with the video mode'): embedder.embed_for_video(self.test_image, 0) - def test_classify_async_calls_with_illegal_timestamp(self): + 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, @@ -365,7 +365,7 @@ class ImageEmbedderTest(parameterized.TestCase): for timestamp in range(0, 300, 30): embedder.embed_async(self.test_image, timestamp) - def test_classify_async_succeeds_with_region_of_interest(self): + 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), From dc30cf973273bd8693ce94465839205bf717ec1e Mon Sep 17 00:00:00 2001 From: kinaryml Date: Tue, 8 Nov 2022 23:51:55 -0800 Subject: [PATCH 07/10] Updated embeddings container to use a timestamp and made parameters - head_index,head_name optional --- .../tasks/python/components/containers/embeddings.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/mediapipe/tasks/python/components/containers/embeddings.py b/mediapipe/tasks/python/components/containers/embeddings.py index 3a024079a..35882b9f3 100644 --- a/mediapipe/tasks/python/components/containers/embeddings.py +++ b/mediapipe/tasks/python/components/containers/embeddings.py @@ -110,8 +110,8 @@ class Embedding: """ embedding: np.ndarray - head_index: int - head_name: str + head_index: Optional[int] = None + head_name: Optional[str] = None @doc_controls.do_not_generate_docs def to_pb2(self) -> _EmbeddingProto: @@ -167,9 +167,16 @@ 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 @doc_controls.do_not_generate_docs def to_pb2(self) -> _EmbeddingResultProto: From 17bb174444ed374578aaa38387f59ab8389b8822 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Wed, 9 Nov 2022 11:42:32 -0800 Subject: [PATCH 08/10] Refactored embeddings to embedding_result --- .../tasks/python/components/containers/BUILD | 4 +- .../{embeddings.py => embedding_result.py} | 0 mediapipe/tasks/python/components/utils/BUILD | 2 +- .../components/utils/cosine_similarity.py | 4 +- mediapipe/tasks/python/test/vision/BUILD | 2 +- .../python/test/vision/image_embedder_test.py | 14 ++--- mediapipe/tasks/python/vision/BUILD | 3 +- .../tasks/python/vision/image_embedder.py | 52 +++++++++---------- 8 files changed, 40 insertions(+), 41 deletions(-) rename mediapipe/tasks/python/components/containers/{embeddings.py => embedding_result.py} (100%) diff --git a/mediapipe/tasks/python/components/containers/BUILD b/mediapipe/tasks/python/components/containers/BUILD index 1f01e2955..7091c1f41 100644 --- a/mediapipe/tasks/python/components/containers/BUILD +++ b/mediapipe/tasks/python/components/containers/BUILD @@ -106,8 +106,8 @@ py_library( ) py_library( - name = "embeddings", - srcs = ["embeddings.py"], + name = "embedding_result", + srcs = ["embedding_result.py"], deps = [ "//mediapipe/tasks/cc/components/containers/proto:embeddings_py_pb2", "//mediapipe/tasks/python/core:optional_dependencies", diff --git a/mediapipe/tasks/python/components/containers/embeddings.py b/mediapipe/tasks/python/components/containers/embedding_result.py similarity index 100% rename from mediapipe/tasks/python/components/containers/embeddings.py rename to mediapipe/tasks/python/components/containers/embedding_result.py diff --git a/mediapipe/tasks/python/components/utils/BUILD b/mediapipe/tasks/python/components/utils/BUILD index 6d00bb31a..50d4094c0 100644 --- a/mediapipe/tasks/python/components/utils/BUILD +++ b/mediapipe/tasks/python/components/utils/BUILD @@ -22,7 +22,7 @@ py_library( name = "cosine_similarity", srcs = ["cosine_similarity.py"], deps = [ - "//mediapipe/tasks/python/components/containers:embeddings", + "//mediapipe/tasks/python/components/containers:embedding_result", "//mediapipe/tasks/python/components/processors:embedder_options", ], ) diff --git a/mediapipe/tasks/python/components/utils/cosine_similarity.py b/mediapipe/tasks/python/components/utils/cosine_similarity.py index 616f2651f..d6102f0b5 100644 --- a/mediapipe/tasks/python/components/utils/cosine_similarity.py +++ b/mediapipe/tasks/python/components/utils/cosine_similarity.py @@ -15,10 +15,10 @@ import numpy as np -from mediapipe.tasks.python.components.containers import embeddings +from mediapipe.tasks.python.components.containers import embedding_result from mediapipe.tasks.python.components.processors import embedder_options -_Embedding = embeddings.Embedding +_Embedding = embedding_result.Embedding _EmbedderOptions = embedder_options.EmbedderOptions diff --git a/mediapipe/tasks/python/test/vision/BUILD b/mediapipe/tasks/python/test/vision/BUILD index 5fd396161..553e1f5a6 100644 --- a/mediapipe/tasks/python/test/vision/BUILD +++ b/mediapipe/tasks/python/test/vision/BUILD @@ -85,7 +85,7 @@ py_test( "//mediapipe/python:_framework_bindings", "//mediapipe/tasks/python/components/processors:embedder_options", "//mediapipe/tasks/python/components/utils:cosine_similarity", - "//mediapipe/tasks/python/components/containers:embeddings", + "//mediapipe/tasks/python/components/containers:embedding_result", "//mediapipe/tasks/python/components/containers:rect", "//mediapipe/tasks/python/core:base_options", "//mediapipe/tasks/python/test:test_utils", diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index 0dfce91c2..e9ff50ed4 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -23,7 +23,7 @@ from absl.testing import parameterized from mediapipe.python._framework_bindings import image as image_module from mediapipe.tasks.python.components.processors import embedder_options as embedder_options_module -from mediapipe.tasks.python.components.containers import embeddings as embeddings_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.core import base_options as base_options_module from mediapipe.tasks.python.test import test_utils @@ -31,13 +31,13 @@ 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 = embeddings_module.FloatEmbedding -_QuantizedEmbedding = embeddings_module.QuantizedEmbedding -_Embedding = embeddings_module.Embedding -_EmbeddingResult = embeddings_module.EmbeddingResult +_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 @@ -346,7 +346,7 @@ class ImageEmbedderTest(parameterized.TestCase): observed_timestamp_ms = -1 - def check_result(result: _EmbeddingResult, output_image: _Image, + def check_result(result: ImageEmbedderResult, output_image: _Image, timestamp_ms: int): # Checks cosine similarity. self._check_cosine_similarity(result, crop_result, quantize=False, @@ -378,7 +378,7 @@ class ImageEmbedderTest(parameterized.TestCase): image_processing_options = _ImageProcessingOptions(roi) observed_timestamp_ms = -1 - def check_result(result: _EmbeddingResult, output_image: _Image, + def check_result(result: ImageEmbedderResult, output_image: _Image, timestamp_ms: int): # Checks cosine similarity. self._check_cosine_similarity(result, crop_result, quantize=False, diff --git a/mediapipe/tasks/python/vision/BUILD b/mediapipe/tasks/python/vision/BUILD index 3c040fb4d..0af8f07e1 100644 --- a/mediapipe/tasks/python/vision/BUILD +++ b/mediapipe/tasks/python/vision/BUILD @@ -113,7 +113,8 @@ py_library( "//mediapipe/python:packet_creator", "//mediapipe/python:packet_getter", "//mediapipe/tasks/cc/vision/image_embedder/proto:image_embedder_graph_options_py_pb2", - "//mediapipe/tasks/python/components/containers:embeddings", + "//mediapipe/tasks/cc/components/containers/proto:embeddings_py_pb2", + "//mediapipe/tasks/python/components/containers:embedding_result", "//mediapipe/tasks/python/components/processors:classifier_options", "//mediapipe/tasks/python/core:base_options", "//mediapipe/tasks/python/core:optional_dependencies", diff --git a/mediapipe/tasks/python/vision/image_embedder.py b/mediapipe/tasks/python/vision/image_embedder.py index 2851970d8..bec9682d0 100644 --- a/mediapipe/tasks/python/vision/image_embedder.py +++ b/mediapipe/tasks/python/vision/image_embedder.py @@ -22,9 +22,10 @@ from mediapipe.python._framework_bindings import image as image_module from mediapipe.python._framework_bindings import packet as packet_module from mediapipe.python._framework_bindings import task_runner as task_runner_module from mediapipe.tasks.cc.vision.image_embedder.proto import image_embedder_graph_options_pb2 +from mediapipe.tasks.cc.components.containers.proto import embeddings_pb2 from mediapipe.tasks.python.components.processors import embedder_options from mediapipe.tasks.python.components.utils import cosine_similarity -from mediapipe.tasks.python.components.containers import embeddings as embeddings_module +from mediapipe.tasks.python.components.containers import embedding_result as embedding_result_module 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 @@ -32,7 +33,7 @@ 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 = embeddings_module.EmbeddingResult +ImageEmbedderResult = embedding_result_module.EmbeddingResult _BaseOptions = base_options_module.BaseOptions _ImageEmbedderGraphOptionsProto = image_embedder_graph_options_pb2.ImageEmbedderGraphOptions _EmbedderOptions = embedder_options.EmbedderOptions @@ -76,7 +77,7 @@ class ImageEmbedderOptions: quantize: Optional[bool] = None embedder_options: _EmbedderOptions = _EmbedderOptions() result_callback: Optional[ - Callable[[embeddings_module.EmbeddingResult, image_module.Image, + Callable[[ImageEmbedderResult, image_module.Image, int], None]] = None @doc_controls.do_not_generate_docs @@ -140,17 +141,17 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): def packets_callback(output_packets: Mapping[str, packet_module.Packet]): if output_packets[_IMAGE_OUT_STREAM_NAME].is_empty(): return - embedding_result_proto = packet_getter.get_proto( - output_packets[_EMBEDDINGS_OUT_STREAM_NAME]) - embeddings = embeddings_module.EmbeddingResult([ - embeddings_module.Embedding.create_from_pb2(embedding) - for embedding in embedding_result_proto.embeddings - ]) + 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(embeddings, image, - timestamp.value // _MICRO_SECONDS_PER_MILLISECOND) + 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, @@ -174,7 +175,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): self, image: image_module.Image, image_processing_options: Optional[_ImageProcessingOptions] = None - ) -> _ImageEmbedderResult: + ) -> 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. @@ -195,19 +196,18 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): _IMAGE_IN_STREAM_NAME: packet_creator.create_image(image), _NORM_RECT_STREAM_NAME: packet_creator.create_proto( normalized_rect.to_pb2())}) - embedding_result_proto = packet_getter.get_proto( - output_packets[_EMBEDDINGS_OUT_STREAM_NAME]) - return embeddings_module.EmbeddingResult([ - embeddings_module.Embedding.create_from_pb2(embedding) - for embedding in embedding_result_proto.embeddings - ]) + 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: + ) -> 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. @@ -237,13 +237,11 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): normalized_rect.to_pb2()).at( timestamp_ms * _MICRO_SECONDS_PER_MILLISECOND) }) - embedding_result_proto = packet_getter.get_proto( - output_packets[_EMBEDDINGS_OUT_STREAM_NAME]) + embedding_result_proto = embeddings_pb2.EmbeddingResult() + embedding_result_proto.CopyFrom( + packet_getter.get_proto(output_packets[_EMBEDDINGS_OUT_STREAM_NAME])) - return embeddings_module.EmbeddingResult([ - embeddings_module.Embedding.create_from_pb2(embedding) - for embedding in embedding_result_proto.embeddings - ]) + return ImageEmbedderResult.create_from_pb2(embedding_result_proto) def embed_async( self, @@ -290,8 +288,8 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): }) @staticmethod - def cosine_similarity(u: embeddings_module.Embedding, - v: embeddings_module.Embedding) -> float: + def cosine_similarity(u: embedding_result_module.Embedding, + v: embedding_result_module.Embedding) -> float: """Utility function to compute cosine similarity [1] 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 a From 7ec0d8cf3b236c4ee5bccd02197984e3ccaae4de Mon Sep 17 00:00:00 2001 From: kinaryml Date: Wed, 9 Nov 2022 11:53:30 -0800 Subject: [PATCH 09/10] Added tolerance for vector coordinate values in embeddings --- mediapipe/tasks/python/test/vision/image_embedder_test.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index e9ff50ed4..18fd644a2 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -48,6 +48,9 @@ _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 @@ -162,7 +165,7 @@ class ImageEmbedderTest(parameterized.TestCase): # Check embedding value. self.assertAlmostEqual(image_result.embeddings[0].embedding[0], - expected_first_value) + expected_first_value, delta=_EPSILON) # Checks cosine similarity. self._check_cosine_similarity(image_result, crop_result, quantize, From 0e9b9257262abc625089664e13571af7168ff576 Mon Sep 17 00:00:00 2001 From: kinaryml Date: Thu, 10 Nov 2022 02:30:17 -0800 Subject: [PATCH 10/10] Fixed some typos and revised image embedder tests --- .../python/test/vision/image_embedder_test.py | 75 +++++++++---------- .../tasks/python/vision/image_embedder.py | 10 +-- 2 files changed, 40 insertions(+), 45 deletions(-) diff --git a/mediapipe/tasks/python/test/vision/image_embedder_test.py b/mediapipe/tasks/python/test/vision/image_embedder_test.py index 18fd644a2..3df1ed1c5 100644 --- a/mediapipe/tasks/python/test/vision/image_embedder_test.py +++ b/mediapipe/tasks/python/test/vision/image_embedder_test.py @@ -100,28 +100,22 @@ class ImageEmbedderTest(parameterized.TestCase): embedder = _ImageEmbedder.create_from_options(options) self.assertIsInstance(embedder, _ImageEmbedder) - def _check_cosine_similarity(self, result0, result1, quantize, - expected_similarity): - # Checks head_index and head_name. - self.assertEqual(result0.embeddings[0].head_index, 0) - self.assertEqual(result1.embeddings[0].head_index, 0) - self.assertEqual(result0.embeddings[0].head_name, 'feature') - self.assertEqual(result1.embeddings[0].head_name, 'feature') + 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) - # Check embedding sizes. - def _check_embedding_size(result): - self.assertLen(result.embeddings, 1) - embedding_result = result.embeddings[0] - self.assertLen(embedding_result.embedding, 1024) - if quantize: - self.assertEqual(embedding_result.embedding.dtype, np.uint8) - else: - self.assertEqual(embedding_result.embedding.dtype, float) - - # Checks results sizes. - _check_embedding_size(result0) - _check_embedding_size(result1) + 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]) @@ -129,13 +123,17 @@ class ImageEmbedderTest(parameterized.TestCase): delta=_SIMILARITY_TOLERANCE) @parameterized.parameters( - (False, False, False, ModelFileType.FILE_NAME, 0.925519, -0.2101883), - (True, False, False, ModelFileType.FILE_NAME, 0.925519, -0.0142344), - # (False, True, False, ModelFileType.FILE_NAME, 0.926791, 229), - (False, False, True, ModelFileType.FILE_CONTENT, 0.999931, -0.195062) + (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_first_value): + 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) @@ -163,12 +161,13 @@ class ImageEmbedderTest(parameterized.TestCase): image_result = embedder.embed(self.test_image, image_processing_options) crop_result = embedder.embed(self.test_cropped_image) - # Check embedding value. - self.assertAlmostEqual(image_result.embeddings[0].embedding[0], - expected_first_value, delta=_EPSILON) - - # Checks cosine similarity. - self._check_cosine_similarity(image_result, crop_result, quantize, + # 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. @@ -201,7 +200,7 @@ class ImageEmbedderTest(parameterized.TestCase): crop_result = embedder.embed(self.test_cropped_image) # Checks cosine similarity. - self._check_cosine_similarity(image_result, crop_result, quantize, + self._check_cosine_similarity(image_result, crop_result, expected_similarity) def test_missing_result_callback(self): @@ -283,8 +282,7 @@ class ImageEmbedderTest(parameterized.TestCase): timestamp) # Checks cosine similarity. self._check_cosine_similarity( - image_result, crop_result, quantize=False, - expected_similarity=0.925519) + image_result, crop_result, expected_similarity=0.925519) def test_embed_for_video_succeeds_with_region_of_interest(self): options = _ImageEmbedderOptions( @@ -305,8 +303,7 @@ class ImageEmbedderTest(parameterized.TestCase): # Checks cosine similarity. self._check_cosine_similarity( - image_result, crop_result, quantize=False, - expected_similarity=0.999931) + image_result, crop_result, expected_similarity=0.999931) def test_calling_embed_in_live_stream_mode(self): options = _ImageEmbedderOptions( @@ -352,8 +349,8 @@ class ImageEmbedderTest(parameterized.TestCase): def check_result(result: ImageEmbedderResult, output_image: _Image, timestamp_ms: int): # Checks cosine similarity. - self._check_cosine_similarity(result, crop_result, quantize=False, - expected_similarity=0.925519) + 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())) @@ -384,7 +381,7 @@ class ImageEmbedderTest(parameterized.TestCase): def check_result(result: ImageEmbedderResult, output_image: _Image, timestamp_ms: int): # Checks cosine similarity. - self._check_cosine_similarity(result, crop_result, quantize=False, + self._check_cosine_similarity(result, crop_result, expected_similarity=0.999931) self.assertTrue( np.array_equal(output_image.numpy_view(), diff --git a/mediapipe/tasks/python/vision/image_embedder.py b/mediapipe/tasks/python/vision/image_embedder.py index bec9682d0..e696ebdc8 100644 --- a/mediapipe/tasks/python/vision/image_embedder.py +++ b/mediapipe/tasks/python/vision/image_embedder.py @@ -20,7 +20,6 @@ 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.python._framework_bindings import task_runner as task_runner_module from mediapipe.tasks.cc.vision.image_embedder.proto import image_embedder_graph_options_pb2 from mediapipe.tasks.cc.components.containers.proto import embeddings_pb2 from mediapipe.tasks.python.components.processors import embedder_options @@ -40,7 +39,6 @@ _EmbedderOptions = embedder_options.EmbedderOptions _RunningMode = running_mode_module.VisionTaskRunningMode _TaskInfo = task_info_module.TaskInfo _ImageProcessingOptions = image_processing_options_module.ImageProcessingOptions -_TaskRunner = task_runner_module.TaskRunner _EMBEDDINGS_OUT_STREAM_NAME = 'embeddings_out' _EMBEDDINGS_TAG = 'EMBEDDINGS' @@ -112,7 +110,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): `ImageEmbedderOptions`. Raises: - ValueError: If failed to create `ImageClassifier` object from the provided + ValueError: If failed to create `ImageEmbedder` object from the provided file such as invalid file path. RuntimeError: If other types of error occurred. """ @@ -185,7 +183,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): image_processing_options: Options for image processing. Returns: - A embedding result object that contains a list of embeddings. + An embedding result object that contains a list of embeddings. Raises: ValueError: If any of the input arguments is invalid. @@ -223,7 +221,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): image_processing_options: Options for image processing. Returns: - A embedding result object that contains a list of embeddings. + An embedding result object that contains a list of embeddings. Raises: ValueError: If any of the input arguments is invalid. @@ -265,7 +263,7 @@ class ImageEmbedder(base_vision_task_api.BaseVisionTaskApi): per input image. The `result_callback` provides: - - A embedding result object that contains a list of embeddings. + - An embedding result object that contains a list of embeddings. - The input image that the image embedder runs on. - The input timestamp in milliseconds.