Merge pull request #3738 from kinaryml:image-classification-python-impl

PiperOrigin-RevId: 483818404
This commit is contained in:
Copybara-Service 2022-10-25 17:26:32 -07:00
commit ae5b09e2b2
11 changed files with 1297 additions and 2 deletions

View File

@ -87,6 +87,7 @@ cc_library(
cc_library(
name = "builtin_task_graphs",
deps = [
"//mediapipe/tasks/cc/vision/image_classifier:image_classifier_graph",
"//mediapipe/tasks/cc/vision/object_detector:object_detector_graph",
],
)

View File

@ -14,7 +14,7 @@
"""The public facing packet getter APIs."""
from typing import List, Type
from typing import List
from google.protobuf import message
from google.protobuf import symbol_database
@ -39,7 +39,7 @@ get_image_frame = _packet_getter.get_image_frame
get_matrix = _packet_getter.get_matrix
def get_proto(packet: mp_packet.Packet) -> Type[message.Message]:
def get_proto(packet: mp_packet.Packet) -> message.Message:
"""Get the content of a MediaPipe proto Packet as a proto message.
Args:

View File

@ -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 +56,13 @@ py_library(
"//mediapipe/tasks/python/core:optional_dependencies",
],
)
py_library(
name = "classifications",
srcs = ["classifications.py"],
deps = [
":category",
"//mediapipe/tasks/cc/components/containers/proto:classifications_py_pb2",
"//mediapipe/tasks/python/core:optional_dependencies",
],
)

View File

@ -0,0 +1,168 @@
# Copyright 2022 The TensorFlow 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.
"""Classifications data class."""
import dataclasses
from typing import Any, List, Optional
from mediapipe.tasks.cc.components.containers.proto import classifications_pb2
from mediapipe.tasks.python.components.containers import category as category_module
from mediapipe.tasks.python.core.optional_dependencies import doc_controls
_ClassificationEntryProto = classifications_pb2.ClassificationEntry
_ClassificationsProto = classifications_pb2.Classifications
_ClassificationResultProto = classifications_pb2.ClassificationResult
@dataclasses.dataclass
class ClassificationEntry:
"""List of predicted classes (aka labels) for a given classifier head.
Attributes:
categories: The array of predicted categories, usually sorted by descending
scores (e.g. from high to low probability).
timestamp_ms: The optional timestamp (in milliseconds) associated to the
classification entry. This is useful for time series use cases, e.g.,
audio classification.
"""
categories: List[category_module.Category]
timestamp_ms: Optional[int] = None
@doc_controls.do_not_generate_docs
def to_pb2(self) -> _ClassificationEntryProto:
"""Generates a ClassificationEntry protobuf object."""
return _ClassificationEntryProto(
categories=[category.to_pb2() for category in self.categories],
timestamp_ms=self.timestamp_ms)
@classmethod
@doc_controls.do_not_generate_docs
def create_from_pb2(
cls, pb2_obj: _ClassificationEntryProto) -> 'ClassificationEntry':
"""Creates a `ClassificationEntry` object from the given protobuf object."""
return ClassificationEntry(
categories=[
category_module.Category.create_from_pb2(category)
for category in pb2_obj.categories
],
timestamp_ms=pb2_obj.timestamp_ms)
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, ClassificationEntry):
return False
return self.to_pb2().__eq__(other.to_pb2())
@dataclasses.dataclass
class Classifications:
"""Represents the classifications for a given classifier head.
Attributes:
entries: A list of `ClassificationEntry` objects.
head_index: The index of the classifier head these categories refer to. This
is useful for multi-head models.
head_name: The name of the classifier head, which is the corresponding
tensor metadata name.
"""
entries: List[ClassificationEntry]
head_index: int
head_name: str
@doc_controls.do_not_generate_docs
def to_pb2(self) -> _ClassificationsProto:
"""Generates a Classifications protobuf object."""
return _ClassificationsProto(
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: _ClassificationsProto) -> 'Classifications':
"""Creates a `Classifications` object from the given protobuf object."""
return Classifications(
entries=[
ClassificationEntry.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, Classifications):
return False
return self.to_pb2().__eq__(other.to_pb2())
@dataclasses.dataclass
class ClassificationResult:
"""Contains one set of results per classifier head.
Attributes:
classifications: A list of `Classifications` objects.
"""
classifications: List[Classifications]
@doc_controls.do_not_generate_docs
def to_pb2(self) -> _ClassificationResultProto:
"""Generates a ClassificationResult protobuf object."""
return _ClassificationResultProto(classifications=[
classification.to_pb2() for classification in self.classifications
])
@classmethod
@doc_controls.do_not_generate_docs
def create_from_pb2(
cls, pb2_obj: _ClassificationResultProto) -> 'ClassificationResult':
"""Creates a `ClassificationResult` object from the given protobuf object.
"""
return ClassificationResult(classifications=[
Classifications.create_from_pb2(classification)
for classification in pb2_obj.classifications
])
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, ClassificationResult):
return False
return self.to_pb2().__eq__(other.to_pb2())

View File

@ -0,0 +1,140 @@
# 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())

View File

@ -0,0 +1,30 @@
# Copyright 2022 The MediaPipe Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Placeholder for internal Python strict library compatibility macro.
# Placeholder for internal Python strict library and test compatibility macro.
package(default_visibility = ["//mediapipe/tasks:internal"])
licenses(["notice"])
py_library(
name = "classifier_options",
srcs = ["classifier_options.py"],
deps = [
"//mediapipe/tasks/cc/components/processors/proto:classifier_options_py_pb2",
"//mediapipe/tasks/python/core:optional_dependencies",
],
)

View File

@ -0,0 +1,86 @@
# Copyright 2022 The TensorFlow 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.
"""Classifier options data class."""
import dataclasses
from typing import Any, List, Optional
from mediapipe.tasks.cc.components.processors.proto import classifier_options_pb2
from mediapipe.tasks.python.core.optional_dependencies import doc_controls
_ClassifierOptionsProto = classifier_options_pb2.ClassifierOptions
@dataclasses.dataclass
class ClassifierOptions:
"""Options for classification processor.
Attributes:
display_names_locale: The locale to use for display names specified through
the TFLite Model Metadata.
max_results: The maximum number of top-scored classification results to
return.
score_threshold: Overrides the ones provided in the model metadata. Results
below this value are rejected.
category_allowlist: Allowlist of category names. If non-empty, detection
results whose category name is not in this set will be filtered out.
Duplicate or unknown category names are ignored. Mutually exclusive with
`category_denylist`.
category_denylist: Denylist of category names. If non-empty, detection
results whose category name is in this set will be filtered out. Duplicate
or unknown category names are ignored. Mutually exclusive with
`category_allowlist`.
"""
display_names_locale: Optional[str] = None
max_results: Optional[int] = None
score_threshold: Optional[float] = None
category_allowlist: Optional[List[str]] = None
category_denylist: Optional[List[str]] = None
@doc_controls.do_not_generate_docs
def to_pb2(self) -> _ClassifierOptionsProto:
"""Generates a ClassifierOptions protobuf object."""
return _ClassifierOptionsProto(
score_threshold=self.score_threshold,
category_allowlist=self.category_allowlist,
category_denylist=self.category_denylist,
display_names_locale=self.display_names_locale,
max_results=self.max_results)
@classmethod
@doc_controls.do_not_generate_docs
def create_from_pb2(cls,
pb2_obj: _ClassifierOptionsProto) -> 'ClassifierOptions':
"""Creates a `ClassifierOptions` object from the given protobuf object."""
return ClassifierOptions(
score_threshold=pb2_obj.score_threshold,
category_allowlist=[str(name) for name in pb2_obj.category_allowlist],
category_denylist=[str(name) for name in pb2_obj.category_denylist],
display_names_locale=pb2_obj.display_names_locale,
max_results=pb2_obj.max_results)
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, ClassifierOptions):
return False
return self.to_pb2().__eq__(other.to_pb2())

View File

@ -36,3 +36,23 @@ py_test(
"//mediapipe/tasks/python/vision/core:vision_task_running_mode",
],
)
py_test(
name = "image_classifier_test",
srcs = ["image_classifier_test.py"],
data = [
"//mediapipe/tasks/testdata/vision:test_images",
"//mediapipe/tasks/testdata/vision:test_models",
],
deps = [
"//mediapipe/python:_framework_bindings",
"//mediapipe/tasks/python/components/containers:category",
"//mediapipe/tasks/python/components/containers:classifications",
"//mediapipe/tasks/python/components/containers:rect",
"//mediapipe/tasks/python/components/processors:classifier_options",
"//mediapipe/tasks/python/core:base_options",
"//mediapipe/tasks/python/test:test_utils",
"//mediapipe/tasks/python/vision:image_classifier",
"//mediapipe/tasks/python/vision/core:vision_task_running_mode",
],
)

View File

@ -0,0 +1,515 @@
# 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 classifier."""
import enum
from unittest import mock
from absl.testing import absltest
from absl.testing import parameterized
import numpy as np
from mediapipe.python._framework_bindings import image
from mediapipe.tasks.python.components.containers import category
from mediapipe.tasks.python.components.containers import classifications as classifications_module
from mediapipe.tasks.python.components.containers import rect
from mediapipe.tasks.python.components.processors import classifier_options
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_classifier
from mediapipe.tasks.python.vision.core import vision_task_running_mode
_NormalizedRect = rect.NormalizedRect
_BaseOptions = base_options_module.BaseOptions
_ClassifierOptions = classifier_options.ClassifierOptions
_Category = category.Category
_ClassificationEntry = classifications_module.ClassificationEntry
_Classifications = classifications_module.Classifications
_ClassificationResult = classifications_module.ClassificationResult
_Image = image.Image
_ImageClassifier = image_classifier.ImageClassifier
_ImageClassifierOptions = image_classifier.ImageClassifierOptions
_RUNNING_MODE = vision_task_running_mode.VisionTaskRunningMode
_MODEL_FILE = 'mobilenet_v2_1.0_224.tflite'
_IMAGE_FILE = 'burger.jpg'
_ALLOW_LIST = ['cheeseburger', 'guacamole']
_DENY_LIST = ['cheeseburger']
_SCORE_THRESHOLD = 0.5
_MAX_RESULTS = 3
# TODO: Port assertProtoEquals
def _assert_proto_equals(expected, actual): # pylint: disable=unused-argument
pass
def _generate_empty_results(timestamp_ms: int) -> _ClassificationResult:
return _ClassificationResult(classifications=[
_Classifications(
entries=[
_ClassificationEntry(categories=[], timestamp_ms=timestamp_ms)
],
head_index=0,
head_name='probability')
])
def _generate_burger_results(timestamp_ms: int) -> _ClassificationResult:
return _ClassificationResult(classifications=[
_Classifications(
entries=[
_ClassificationEntry(
categories=[
_Category(
index=934,
score=0.7939587831497192,
display_name='',
category_name='cheeseburger'),
_Category(
index=932,
score=0.02739289402961731,
display_name='',
category_name='bagel'),
_Category(
index=925,
score=0.01934075355529785,
display_name='',
category_name='guacamole'),
_Category(
index=963,
score=0.006327860057353973,
display_name='',
category_name='meat loaf')
],
timestamp_ms=timestamp_ms)
],
head_index=0,
head_name='probability')
])
def _generate_soccer_ball_results(timestamp_ms: int) -> _ClassificationResult:
return _ClassificationResult(classifications=[
_Classifications(
entries=[
_ClassificationEntry(
categories=[
_Category(
index=806,
score=0.9965274930000305,
display_name='',
category_name='soccer ball')
],
timestamp_ms=timestamp_ms)
],
head_index=0,
head_name='probability')
])
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)
def test_create_from_file_succeeds_with_valid_model_path(self):
# Creates with default option and valid model file successfully.
with _ImageClassifier.create_from_model_path(self.model_path) as classifier:
self.assertIsInstance(classifier, _ImageClassifier)
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 = _ImageClassifierOptions(base_options=base_options)
with _ImageClassifier.create_from_options(options) as classifier:
self.assertIsInstance(classifier, _ImageClassifier)
def test_create_from_options_fails_with_invalid_model_path(self):
# Invalid empty model path.
with self.assertRaisesRegex(
ValueError,
r"ExternalFile must specify at least one of 'file_content', "
r"'file_name', 'file_pointer_meta' or 'file_descriptor_meta'."):
base_options = _BaseOptions(model_asset_path='')
options = _ImageClassifierOptions(base_options=base_options)
_ImageClassifier.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 = _ImageClassifierOptions(base_options=base_options)
classifier = _ImageClassifier.create_from_options(options)
self.assertIsInstance(classifier, _ImageClassifier)
@parameterized.parameters(
(ModelFileType.FILE_NAME, 4, _generate_burger_results(0)),
(ModelFileType.FILE_CONTENT, 4, _generate_burger_results(0)))
def test_classify(self, model_file_type, max_results,
expected_classification_result):
# Creates classifier.
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.')
custom_classifier_options = _ClassifierOptions(max_results=max_results)
options = _ImageClassifierOptions(
base_options=base_options, classifier_options=custom_classifier_options)
classifier = _ImageClassifier.create_from_options(options)
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
# Comparing results.
_assert_proto_equals(image_result.to_pb2(),
expected_classification_result.to_pb2())
# Closes the classifier explicitly when the classifier is not used in
# a context.
classifier.close()
@parameterized.parameters(
(ModelFileType.FILE_NAME, 4, _generate_burger_results(0)),
(ModelFileType.FILE_CONTENT, 4, _generate_burger_results(0)))
def test_classify_in_context(self, model_file_type, max_results,
expected_classification_result):
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.')
custom_classifier_options = _ClassifierOptions(max_results=max_results)
options = _ImageClassifierOptions(
base_options=base_options, classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
# Comparing results.
_assert_proto_equals(image_result.to_pb2(),
expected_classification_result.to_pb2())
def test_classify_succeeds_with_region_of_interest(self):
base_options = _BaseOptions(model_asset_path=self.model_path)
custom_classifier_options = _ClassifierOptions(max_results=1)
options = _ImageClassifierOptions(
base_options=base_options, classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Load the test image.
test_image = _Image.create_from_file(
test_utils.get_test_data_path('multi_objects.jpg'))
# NormalizedRect around the soccer ball.
roi = _NormalizedRect(
x_center=0.532, y_center=0.521, width=0.164, height=0.427)
# Performs image classification on the input.
image_result = classifier.classify(test_image, roi)
# Comparing results.
_assert_proto_equals(image_result.to_pb2(),
_generate_soccer_ball_results(0).to_pb2())
def test_score_threshold_option(self):
custom_classifier_options = _ClassifierOptions(
score_threshold=_SCORE_THRESHOLD)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
classifications = image_result.classifications
for classification in classifications:
for entry in classification.entries:
score = entry.categories[0].score
self.assertGreaterEqual(
score, _SCORE_THRESHOLD,
f'Classification with score lower than threshold found. '
f'{classification}')
def test_max_results_option(self):
custom_classifier_options = _ClassifierOptions(
score_threshold=_SCORE_THRESHOLD)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
categories = image_result.classifications[0].entries[0].categories
self.assertLessEqual(
len(categories), _MAX_RESULTS, 'Too many results returned.')
def test_allow_list_option(self):
custom_classifier_options = _ClassifierOptions(
category_allowlist=_ALLOW_LIST)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
classifications = image_result.classifications
for classification in classifications:
for entry in classification.entries:
label = entry.categories[0].category_name
self.assertIn(label, _ALLOW_LIST,
f'Label {label} found but not in label allow list')
def test_deny_list_option(self):
custom_classifier_options = _ClassifierOptions(category_denylist=_DENY_LIST)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
classifications = image_result.classifications
for classification in classifications:
for entry in classification.entries:
label = entry.categories[0].category_name
self.assertNotIn(label, _DENY_LIST,
f'Label {label} found but in deny list.')
def test_combined_allowlist_and_denylist(self):
# Fails with combined allowlist and denylist
with self.assertRaisesRegex(
ValueError,
r'`category_allowlist` and `category_denylist` are mutually '
r'exclusive options.'):
custom_classifier_options = _ClassifierOptions(
category_allowlist=['foo'], category_denylist=['bar'])
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as unused_classifier:
pass
def test_empty_classification_outputs(self):
custom_classifier_options = _ClassifierOptions(score_threshold=1)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Performs image classification on the input.
image_result = classifier.classify(self.test_image)
self.assertEmpty(image_result.classifications[0].entries[0].categories)
def test_missing_result_callback(self):
options = _ImageClassifierOptions(
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 _ImageClassifier.create_from_options(options) as unused_classifier:
pass
@parameterized.parameters((_RUNNING_MODE.IMAGE), (_RUNNING_MODE.VIDEO))
def test_illegal_result_callback(self, running_mode):
options = _ImageClassifierOptions(
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 _ImageClassifier.create_from_options(options) as unused_classifier:
pass
def test_calling_classify_for_video_in_image_mode(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.IMAGE)
with _ImageClassifier.create_from_options(options) as classifier:
with self.assertRaisesRegex(ValueError,
r'not initialized with the video mode'):
classifier.classify_for_video(self.test_image, 0)
def test_calling_classify_async_in_image_mode(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.IMAGE)
with _ImageClassifier.create_from_options(options) as classifier:
with self.assertRaisesRegex(ValueError,
r'not initialized with the live stream mode'):
classifier.classify_async(self.test_image, 0)
def test_calling_classify_in_video_mode(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO)
with _ImageClassifier.create_from_options(options) as classifier:
with self.assertRaisesRegex(ValueError,
r'not initialized with the image mode'):
classifier.classify(self.test_image)
def test_calling_classify_async_in_video_mode(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO)
with _ImageClassifier.create_from_options(options) as classifier:
with self.assertRaisesRegex(ValueError,
r'not initialized with the live stream mode'):
classifier.classify_async(self.test_image, 0)
def test_classify_for_video_with_out_of_order_timestamp(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO)
with _ImageClassifier.create_from_options(options) as classifier:
unused_result = classifier.classify_for_video(self.test_image, 1)
with self.assertRaisesRegex(
ValueError, r'Input timestamp must be monotonically increasing'):
classifier.classify_for_video(self.test_image, 0)
def test_classify_for_video(self):
custom_classifier_options = _ClassifierOptions(max_results=4)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO,
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
for timestamp in range(0, 300, 30):
classification_result = classifier.classify_for_video(
self.test_image, timestamp)
_assert_proto_equals(classification_result.to_pb2(),
_generate_burger_results(timestamp).to_pb2())
def test_classify_for_video_succeeds_with_region_of_interest(self):
custom_classifier_options = _ClassifierOptions(max_results=1)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.VIDEO,
classifier_options=custom_classifier_options)
with _ImageClassifier.create_from_options(options) as classifier:
# Load the test image.
test_image = _Image.create_from_file(
test_utils.get_test_data_path('multi_objects.jpg'))
# NormalizedRect around the soccer ball.
roi = _NormalizedRect(
x_center=0.532, y_center=0.521, width=0.164, height=0.427)
for timestamp in range(0, 300, 30):
classification_result = classifier.classify_for_video(
test_image, timestamp, roi)
self.assertEqual(classification_result,
_generate_soccer_ball_results(timestamp))
def test_calling_classify_in_live_stream_mode(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
result_callback=mock.MagicMock())
with _ImageClassifier.create_from_options(options) as classifier:
with self.assertRaisesRegex(ValueError,
r'not initialized with the image mode'):
classifier.classify(self.test_image)
def test_calling_classify_for_video_in_live_stream_mode(self):
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
result_callback=mock.MagicMock())
with _ImageClassifier.create_from_options(options) as classifier:
with self.assertRaisesRegex(ValueError,
r'not initialized with the video mode'):
classifier.classify_for_video(self.test_image, 0)
def test_classify_async_calls_with_illegal_timestamp(self):
custom_classifier_options = _ClassifierOptions(max_results=4)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
classifier_options=custom_classifier_options,
result_callback=mock.MagicMock())
with _ImageClassifier.create_from_options(options) as classifier:
classifier.classify_async(self.test_image, 100)
with self.assertRaisesRegex(
ValueError, r'Input timestamp must be monotonically increasing'):
classifier.classify_async(self.test_image, 0)
@parameterized.parameters((0, _generate_burger_results),
(1, _generate_empty_results))
def test_classify_async_calls(self, threshold, expected_result_fn):
observed_timestamp_ms = -1
def check_result(result: _ClassificationResult, output_image: _Image,
timestamp_ms: int):
_assert_proto_equals(result.to_pb2(),
expected_result_fn(timestamp_ms).to_pb2())
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
custom_classifier_options = _ClassifierOptions(
max_results=4, score_threshold=threshold)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
classifier_options=custom_classifier_options,
result_callback=check_result)
with _ImageClassifier.create_from_options(options) as classifier:
for timestamp in range(0, 300, 30):
classifier.classify_async(self.test_image, timestamp)
def test_classify_async_succeeds_with_region_of_interest(self):
# Load the test image.
test_image = _Image.create_from_file(
test_utils.get_test_data_path('multi_objects.jpg'))
# NormalizedRect around the soccer ball.
roi = _NormalizedRect(
x_center=0.532, y_center=0.521, width=0.164, height=0.427)
observed_timestamp_ms = -1
def check_result(result: _ClassificationResult, output_image: _Image,
timestamp_ms: int):
_assert_proto_equals(result.to_pb2(),
_generate_soccer_ball_results(timestamp_ms).to_pb2())
self.assertEqual(output_image.width, test_image.width)
self.assertEqual(output_image.height, test_image.height)
self.assertLess(observed_timestamp_ms, timestamp_ms)
self.observed_timestamp_ms = timestamp_ms
custom_classifier_options = _ClassifierOptions(max_results=1)
options = _ImageClassifierOptions(
base_options=_BaseOptions(model_asset_path=self.model_path),
running_mode=_RUNNING_MODE.LIVE_STREAM,
classifier_options=custom_classifier_options,
result_callback=check_result)
with _ImageClassifier.create_from_options(options) as classifier:
for timestamp in range(0, 300, 30):
classifier.classify_async(test_image, timestamp, roi)
if __name__ == '__main__':
absltest.main()

View File

@ -36,3 +36,25 @@ py_library(
"//mediapipe/tasks/python/vision/core:vision_task_running_mode",
],
)
py_library(
name = "image_classifier",
srcs = [
"image_classifier.py",
],
deps = [
"//mediapipe/python:_framework_bindings",
"//mediapipe/python:packet_creator",
"//mediapipe/python:packet_getter",
"//mediapipe/tasks/cc/components/containers/proto:classifications_py_pb2",
"//mediapipe/tasks/cc/vision/image_classifier/proto:image_classifier_graph_options_py_pb2",
"//mediapipe/tasks/python/components/containers:classifications",
"//mediapipe/tasks/python/components/containers:rect",
"//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",
"//mediapipe/tasks/python/vision/core:base_vision_task_api",
"//mediapipe/tasks/python/vision/core:vision_task_running_mode",
],
)

View File

@ -0,0 +1,294 @@
# 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 classifier task."""
import dataclasses
from typing import Callable, Mapping, Optional
from mediapipe.python import packet_creator
from mediapipe.python import packet_getter
# TODO: Import MPImage directly one we have an alias
from mediapipe.python._framework_bindings import image as image_module
from mediapipe.python._framework_bindings import packet
from mediapipe.python._framework_bindings import task_runner
from mediapipe.tasks.cc.components.containers.proto import classifications_pb2
from mediapipe.tasks.cc.vision.image_classifier.proto import image_classifier_graph_options_pb2
from mediapipe.tasks.python.components.containers import classifications
from mediapipe.tasks.python.components.containers import rect
from mediapipe.tasks.python.components.processors import classifier_options
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
_NormalizedRect = rect.NormalizedRect
_BaseOptions = base_options_module.BaseOptions
_ImageClassifierGraphOptionsProto = image_classifier_graph_options_pb2.ImageClassifierGraphOptions
_ClassifierOptions = classifier_options.ClassifierOptions
_RunningMode = vision_task_running_mode.VisionTaskRunningMode
_TaskInfo = task_info_module.TaskInfo
_TaskRunner = task_runner.TaskRunner
_CLASSIFICATION_RESULT_OUT_STREAM_NAME = 'classification_result_out'
_CLASSIFICATION_RESULT_TAG = 'CLASSIFICATION_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_classifier.ImageClassifierGraph'
_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 ImageClassifierOptions:
"""Options for the image classifier task.
Attributes:
base_options: Base options for the image classifier task.
running_mode: The running mode of the task. Default to the image mode. Image
classifier task has three running modes: 1) The image mode for classifying
objects on single image inputs. 2) The video mode for classifying objects
on the decoded frames of a video. 3) The live stream mode for classifying
objects on a live stream of input data, such as from camera.
classifier_options: Options for the image classification 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
classifier_options: _ClassifierOptions = _ClassifierOptions()
result_callback: Optional[
Callable[[classifications.ClassificationResult, image_module.Image, int],
None]] = None
@doc_controls.do_not_generate_docs
def to_pb2(self) -> _ImageClassifierGraphOptionsProto:
"""Generates an ImageClassifierOptions 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
classifier_options_proto = self.classifier_options.to_pb2()
return _ImageClassifierGraphOptionsProto(
base_options=base_options_proto,
classifier_options=classifier_options_proto)
class ImageClassifier(base_vision_task_api.BaseVisionTaskApi):
"""Class that performs image classification on images."""
@classmethod
def create_from_model_path(cls, model_path: str) -> 'ImageClassifier':
"""Creates an `ImageClassifier` object from a TensorFlow Lite model and the default `ImageClassifierOptions`.
Note that the created `ImageClassifier` instance is in image mode, for
classifying objects on single image inputs.
Args:
model_path: Path to the model.
Returns:
`ImageClassifier` object that's created from the model file and the
default `ImageClassifierOptions`.
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 = ImageClassifierOptions(
base_options=base_options, running_mode=_RunningMode.IMAGE)
return cls.create_from_options(options)
@classmethod
def create_from_options(cls,
options: ImageClassifierOptions) -> 'ImageClassifier':
"""Creates the `ImageClassifier` object from image classifier options.
Args:
options: Options for the image classifier task.
Returns:
`ImageClassifier` object that's created from `options`.
Raises:
ValueError: If failed to create `ImageClassifier` object from
`ImageClassifierOptions` such as missing the model.
RuntimeError: If other types of error occurred.
"""
def packets_callback(output_packets: Mapping[str, packet.Packet]):
if output_packets[_IMAGE_OUT_STREAM_NAME].is_empty():
return
classification_result_proto = classifications_pb2.ClassificationResult()
classification_result_proto.CopyFrom(
packet_getter.get_proto(
output_packets[_CLASSIFICATION_RESULT_OUT_STREAM_NAME]))
classification_result = classifications.ClassificationResult([
classifications.Classifications.create_from_pb2(classification)
for classification in classification_result_proto.classifications
])
image = packet_getter.get_image(output_packets[_IMAGE_OUT_STREAM_NAME])
timestamp = output_packets[_IMAGE_OUT_STREAM_NAME].timestamp
options.result_callback(classification_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([
_CLASSIFICATION_RESULT_TAG,
_CLASSIFICATION_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)
# TODO: Replace _NormalizedRect with ImageProcessingOption
def classify(
self,
image: image_module.Image,
roi: Optional[_NormalizedRect] = None
) -> classifications.ClassificationResult:
"""Performs image classification on the provided MediaPipe Image.
Args:
image: MediaPipe Image.
roi: The region of interest.
Returns:
A classification result object that contains a list of classifications.
Raises:
ValueError: If any of the input arguments is invalid.
RuntimeError: If image classification 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())
})
classification_result_proto = classifications_pb2.ClassificationResult()
classification_result_proto.CopyFrom(
packet_getter.get_proto(
output_packets[_CLASSIFICATION_RESULT_OUT_STREAM_NAME]))
return classifications.ClassificationResult([
classifications.Classifications.create_from_pb2(classification)
for classification in classification_result_proto.classifications
])
def classify_for_video(
self,
image: image_module.Image,
timestamp_ms: int,
roi: Optional[_NormalizedRect] = None
) -> classifications.ClassificationResult:
"""Performs image classification on the provided video frames.
Only use this method when the ImageClassifier 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 classification result object that contains a list of classifications.
Raises:
ValueError: If any of the input arguments is invalid.
RuntimeError: If image classification 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)
})
classification_result_proto = classifications_pb2.ClassificationResult()
classification_result_proto.CopyFrom(
packet_getter.get_proto(
output_packets[_CLASSIFICATION_RESULT_OUT_STREAM_NAME]))
return classifications.ClassificationResult([
classifications.Classifications.create_from_pb2(classification)
for classification in classification_result_proto.classifications
])
def classify_async(self,
image: image_module.Image,
timestamp_ms: int,
roi: Optional[_NormalizedRect] = None) -> None:
"""Sends live image data (an Image with a unique timestamp) to perform image classification.
Only use this method when the ImageClassifier 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 `ImageClassifierOptions`. The
`classify_async` method is designed to process live stream data such as
camera input. To lower the overall latency, image classifier 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 classification result object that contains a list of classifications.
- The input image that the image classifier 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
classifier 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)
})