def save_metadata_to_model_file(metadata, model_filename):
b = flatbuffers.Builder(0)
b.Finish(metadata.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
metadata_buffer = b.Output()
populator = _metadata.MetadataPopulator.with_model_file(model_filename)
populator.load_metadata_buffer(metadata_buffer)
# populator.load_associated_files(["your_path_to_label_file"]) # No associated files for this (e.g. No labels files)
populator.populate()
_metadata_fb.FeaturePropertiesT())
group = _metadata_fb.TensorGroupT()
group.name = "detection result"
group.tensorNames = [
output_location_meta.name, output_class_meta.name, output_score_meta.name
]
subgraph = _metadata_fb.SubGraphMetadataT()
subgraph.inputTensorMetadata = [input_meta]
subgraph.outputTensorMetadata = [
output_location_meta, output_class_meta, output_score_meta,
output_number_meta
]
subgraph.outputTensorGroups = [group]
model_meta.subgraphMetadata = [subgraph]
b = flatbuffers.Builder(0)
b.Finish(model_meta.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
metadta_buf = b.Output()
#본인 경로
populator = _metadata.MetadataPopulator.with_model_file("detect.tflite")
populator.load_metadata_buffer(metadta_buf)
#본인 경로
populator.load_associated_files(["label.txt"])
populator.populate()
# #메타 데이터 시각화
# import os
# displayer = _metadata.MetadataDisplayer.with_model_file("C:/Users/lee01/Jupyter_Projects/Safety-Helmet-Wearing-Dataset/ProcessingTfliteDetectionModel/final_model.tflite")
# export_json_file = os.path.join(FLAGS.export_directory,
def build_metadata(name: str,
version: str,
labels: List[LabelDescription],
output_types: List[OutputTensorType],
output_interpretation: str = "",
task: Optional[str] = None,
author: str = "",
task_params: str = "") -> bytearray:
"""
Args:
name:
version:
labels:
output_types:
output_interpretation:
task:
author:
task_params:
Returns:
"""
if task is None or task == "":
# Auto infer detection or localization from targets
types = set(output_types)
if types == {
OutputTensorType.BOX_SHAPE,
OutputTensorType.OBJECTNESS,
OutputTensorType.CLASSES,
}:
task = TaskType.OBJECT_DETECTION.value
elif types == {
OutputTensorType.BOX_SHAPE, OutputTensorType.OBJECTNESS
}:
task = TaskType.OBJECT_LOCALIZATION.value
else:
raise ValueError(
f"Cannot infer task_type from output_types: {output_types}")
metadata = ModelMetadata(
name=name,
version=version,
task=task,
output_interpretation=output_interpretation,
labels=labels,
task_params=task_params,
)
model_meta = _metadata_fb.ModelMetadataT()
model_meta.name = name
model_meta.description = json.dumps(metadata.asdict)
model_meta.version = version
model_meta.author = author
input_meta = _metadata_fb.TensorMetadataT()
input_meta.name = "image"
output_metas = []
for t in output_types:
output_meta = _metadata_fb.TensorMetadataT()
output_meta.name = t.value
output_metas.append(output_meta)
subgraph = _metadata_fb.SubGraphMetadataT()
subgraph.inputTensorMetadata = [input_meta]
subgraph.outputTensorMetadata = output_metas
model_meta.subgraphMetadata = [subgraph]
b = flatbuffers.Builder(0)
b.Finish(model_meta.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
metadata_buf = b.Output()
return metadata_buf
def build_metadata(
model_file_path: str,
#model_name: str,
#model_description: str,
#model_author: str,
#model_license: str,
vocab_file_path: str=None,
sentencepiece_model_path: str=None):
model_meta = _metadata_fb.ModelMetadataT()
model_meta.name = 'distilbert news'
model_meta.description = 'some model description'
model_meta.version = 'v1'
model_meta.author = 'Unknown'
model_meta.license = 'Apache License. Version 2.0'
model_meta.minParserVersion = '1.1.0'
# Creates input info.
ids = _metadata_fb.TensorMetadataT()
segment_ids = _metadata_fb.TensorMetadataT()
mask = _metadata_fb.TensorMetadataT()
ids.name = "ids"
ids.description = "Tokenized ids of input text."
ids.content = _metadata_fb.ContentT()
ids.content.contentProperties = _metadata_fb.FeaturePropertiesT()
ids.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
segment_ids.name = "segment_ids"
segment_ids.description = "0 for the first sequence, 1 for the second sequence if exits."
segment_ids.content = _metadata_fb.ContentT()
segment_ids.content.contentProperties = _metadata_fb.FeaturePropertiesT()
segment_ids.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
mask.name = "mask"
mask.description = "Mask with 1 for real tokens and 0 for padding tokens."
mask.content = _metadata_fb.ContentT()
mask.content.contentProperties = _metadata_fb.FeaturePropertiesT()
mask.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
# Creates output info.
output_meta = _metadata_fb.TensorMetadataT()
output_meta.name = "output"
output_meta.description = "the id of the output class"
output_meta.content = _metadata_fb.ContentT()
output_meta.content.contentProperties = _metadata_fb.FeaturePropertiesT()
output_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
"""
label_file = _metadata_fb.AssociatedFileT()
label_file.name = os.path.basename(label_file_path)
label_file.description = "Labels for the categories to be predicted."
label_file.type = _metadata_fb.AssociatedFileType.TENSOR_AXIS_LABELS
output_meta.associatedFiles = [label_file]
"""
input_process_units = []
if sentencepiece_model_path is not None:
#Add sentencepiece specific process unit
sentencepiece_process_unit = _metadata_fb.ProcessUnitT()
sentencepiece_process_unit.optionsType = (
_metadata_fb.ProcessUnitOptions.SentencePieceTokenizerOptions)
sentencepiece_process_unit.options = _metadata_fb.SentencePieceTokenizerOptionsT()
sentencepiece_model = AssociatedFileT()
sentencepiece_model.name="30k-clean-model",
sentencepiece_model.description="The sentence piece model file."
sentencepiece_process_unit.options.sentencePieceModel = [sentencepiece_model]
input_process_units.append(sentencepiece_process_unit)
if vocab_file_path is not None:
model_process_unit = _metadata_fb.ProcessUnitT()
model_process_unit.optionsType = (
_metadata_fb.ProcessUnitOptions.BertTokenizerOptions)
model_process_unit.options = _metadata_fb.BertTokenizerOptionsT()
vocab_file = AssociatedFileT()
vocab_file.name="jp word piece vocab",
vocab_file.description="Japanese Vocabulary file for the BERT tokenizer.",
vocab_file.type=_metadata_fb.AssociatedFileType.VOCABULARY
model_process_unit.options.vocabFile = [vocab_file]
input_process_units.append(model_process_unit)
#Put metadata together
subgraph = _metadata_fb.SubGraphMetadataT()
subgraph.inputTensorMetadata = [ids, mask, segment_ids]
subgraph.outputTensorMetadata = [output_meta]
subgraph.inputProcessUnits = input_process_units
model_meta.subgraphMetadata = [subgraph]
#create flat buffers for metadata
b = flatbuffers.Builder(0)
b.Finish(
model_meta.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
metadata_buf = b.Output()
#associate files to metadata
populator = _metadata.MetadataPopulator.with_model_file(model_file_path)
populator.load_metadata_buffer(metadata_buf)
if vocab_file_path is not None:
files = [vocab_file_path]
else:
files = [sentencepiece_model_path]
populator.load_associated_files(files)
populator.populate()
def create_metadata(model_file_name, label_map_file_name, num_labels):
from tflite_support import flatbuffers
from tflite_support import metadata as _metadata
from tflite_support import metadata_schema_py_generated as _metadata_fb
""" ... """
"""Creates the metadata for an image classifier."""
# Creates model info.
model_meta = _metadata_fb.ModelMetadataT()
model_meta.name = 'MobileNetV1 image classifier'
model_meta.description = ('Identify the most prominent object in the '
'image from a set of 1,001 categories such as '
'trees, animals, food, vehicles, person etc.')
model_meta.version = 'v1'
model_meta.author = 'TensorFlow'
model_meta.license = ('Apache License. Version 2.0 '
'http://www.apache.org/licenses/LICENSE-2.0.')
# Creates input info.
input_meta = _metadata_fb.TensorMetadataT()
input_meta.name = 'image'
input_meta.description = (
'Input image to be classified. The expected image is {0} x {1}, with '
'three channels (red, blue, and green) per pixel. Each value in the '
'tensor is a single byte between 0 and 255.'.format(416, 416))
input_meta.content = _metadata_fb.ContentT()
input_meta.content.contentProperties = _metadata_fb.ImagePropertiesT()
input_meta.content.contentProperties.colorSpace = (
_metadata_fb.ColorSpaceType.RGB)
input_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.ImageProperties)
input_normalization = _metadata_fb.ProcessUnitT()
input_normalization.optionsType = (
_metadata_fb.ProcessUnitOptions.NormalizationOptions)
input_normalization.options = _metadata_fb.NormalizationOptionsT()
input_normalization.options.mean = [127.5]
input_normalization.options.std = [127.5]
input_meta.processUnits = [input_normalization]
input_stats = _metadata_fb.StatsT()
input_stats.max = [255]
input_stats.min = [0]
input_meta.stats = input_stats
# Creates output info.
bbox_meta = _metadata_fb.TensorMetadataT()
bbox_meta.name = 'bbox'
bbox_meta.description = '.'
bbox_meta.content = _metadata_fb.ContentT()
bbox_meta.content.content_properties = _metadata_fb.FeaturePropertiesT()
bbox_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
bbox_stats = _metadata_fb.StatsT()
bbox_stats.max = [416.0]
bbox_stats.min = [0.0]
bbox_meta.stats = bbox_stats
classes_meta = _metadata_fb.TensorMetadataT()
classes_meta.name = 'classes'
classes_meta.description = '.'
classes_meta.content = _metadata_fb.ContentT()
classes_meta.content.content_properties = _metadata_fb.FeaturePropertiesT()
classes_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
classes_stats = _metadata_fb.StatsT()
classes_stats.max = [num_labels]
classes_stats.min = [0]
classes_meta.stats = classes_stats
label_file = _metadata_fb.AssociatedFileT()
label_file.name = os.path.basename(label_map_file_name)
label_file.description = 'Labels for objects that the model can recognize.'
label_file.type = _metadata_fb.AssociatedFileType.TENSOR_AXIS_LABELS
classes_meta.associatedFiles = [label_file]
confidence_meta = _metadata_fb.TensorMetadataT()
confidence_meta.name = 'confidence'
confidence_meta.description = '.'
confidence_meta.content = _metadata_fb.ContentT()
confidence_meta.content.content_properties = _metadata_fb.FeaturePropertiesT(
)
confidence_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
confidence_stats = _metadata_fb.StatsT()
confidence_stats.max = [1.0]
confidence_stats.min = [0.0]
confidence_meta.stats = confidence_stats
num_dets_meta = _metadata_fb.TensorMetadataT()
num_dets_meta.name = 'num_dets'
num_dets_meta.description = '.'
num_dets_meta.content = _metadata_fb.ContentT()
num_dets_meta.content.content_properties = _metadata_fb.FeaturePropertiesT(
)
num_dets_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
num_dets_stats = _metadata_fb.StatsT()
num_dets_stats.max = [200]
num_dets_stats.min = [0]
num_dets_meta.stats = num_dets_stats
raw_output_meta = _metadata_fb.TensorMetadataT()
raw_output_meta.name = 'raw_output'
raw_output_meta.description = '.'
raw_output_meta.content = _metadata_fb.ContentT()
raw_output_meta.content.content_properties = _metadata_fb.FeaturePropertiesT(
)
raw_output_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
subgraph = _metadata_fb.SubGraphMetadataT()
subgraph.inputTensorMetadata = [input_meta]
subgraph.outputTensorMetadata = [
bbox_meta, classes_meta, confidence_meta, num_dets_stats
] # raw_output_meta
model_meta.subgraphMetadata = [subgraph]
b = flatbuffers.Builder(0)
b.Finish(model_meta.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
metadata_buf = b.Output()
populator = _metadata.MetadataPopulator.with_model_file(model_file_name)
populator.load_metadata_buffer(metadata_buf)
populator.load_associated_files([label_map_file_name])
populator.populate()
def _create_metadata(self):
"""Creates the metadata for the selfie2anime model."""
# Creates model info.
model_meta = _metadata_fb.ModelMetadataT()
model_meta.name = "Selfie2Anime"
model_meta.description = ("Convert selfie to anime.")
model_meta.version = "v1"
model_meta.author = "TensorFlow"
model_meta.license = ("Apache License. Version 2.0 "
"http://www.apache.org/licenses/LICENSE-2.0.")
# Creates info for the input, selfie image.
input_image_meta = _metadata_fb.TensorMetadataT()
input_image_meta.name = "selfie_image"
input_image_meta.description = (
"The expected image is 256 x 256, with three channels "
"(red, blue, and green) per pixel. Each value in the tensor is between"
" 0 and 1.")
input_image_meta.content = _metadata_fb.ContentT()
input_image_meta.content.contentProperties = (
_metadata_fb.ImagePropertiesT())
input_image_meta.content.contentProperties.colorSpace = (
_metadata_fb.ColorSpaceType.RGB)
input_image_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.ImageProperties)
input_image_normalization = _metadata_fb.ProcessUnitT()
input_image_normalization.optionsType = (
_metadata_fb.ProcessUnitOptions.NormalizationOptions)
input_image_normalization.options = _metadata_fb.NormalizationOptionsT(
)
input_image_normalization.options.mean = [0.0]
input_image_normalization.options.std = [255]
input_image_meta.processUnits = [input_image_normalization]
input_image_stats = _metadata_fb.StatsT()
input_image_stats.max = [1.0]
input_image_stats.min = [0.0]
input_image_meta.stats = input_image_stats
# Creates output info, anime image
output_image_meta = _metadata_fb.TensorMetadataT()
output_image_meta.name = "anime_image"
output_image_meta.description = "Image styled."
output_image_meta.content = _metadata_fb.ContentT()
output_image_meta.content.contentProperties = _metadata_fb.ImagePropertiesT(
)
output_image_meta.content.contentProperties.colorSpace = (
_metadata_fb.ColorSpaceType.RGB)
output_image_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.ImageProperties)
output_image_normalization = _metadata_fb.ProcessUnitT()
output_image_normalization.optionsType = (
_metadata_fb.ProcessUnitOptions.NormalizationOptions)
output_image_normalization.options = _metadata_fb.NormalizationOptionsT(
)
output_image_normalization.options.mean = [0.0]
output_image_normalization.options.std = [0.003921568627] # 1/255
output_image_meta.processUnits = [output_image_normalization]
output_image_stats = _metadata_fb.StatsT()
output_image_stats.max = [1.0]
output_image_stats.min = [0.0]
output_image_meta.stats = output_image_stats
# Creates subgraph info.
subgraph = _metadata_fb.SubGraphMetadataT()
subgraph.inputTensorMetadata = [input_image_meta
] # Updated by Margaret
subgraph.outputTensorMetadata = [output_image_meta
] # Updated by Margaret
model_meta.subgraphMetadata = [subgraph]
b = flatbuffers.Builder(0)
b.Finish(model_meta.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
self.metadata_buf = b.Output()
def _create_metadata(self):
"""Creates the metadata for an image classifier."""
# Creates model info.
model_meta = _metadata_fb.ModelMetadataT()
model_meta.name = self.model_info.name
model_meta.description = ("Identify the most prominent object in the "
"image from a set of categories.")
model_meta.version = self.model_info.version
model_meta.author = "TFLite Model Maker"
model_meta.license = ("Apache License. Version 2.0 "
"http://www.apache.org/licenses/LICENSE-2.0.")
# Creates input info.
input_meta = _metadata_fb.TensorMetadataT()
input_meta.name = "image"
input_meta.description = (
"Input image to be classified. The expected image is {0} x {1}, with "
"three channels (red, blue, and green) per pixel. Each value in the "
"tensor is a single byte between {2} and {3}.".format(
self.model_info.image_width, self.model_info.image_height,
self.model_info.image_min, self.model_info.image_max))
input_meta.content = _metadata_fb.ContentT()
input_meta.content.contentProperties = _metadata_fb.ImagePropertiesT()
input_meta.content.contentProperties.colorSpace = (
_metadata_fb.ColorSpaceType.RGB)
input_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.ImageProperties)
input_normalization = _metadata_fb.ProcessUnitT()
input_normalization.optionsType = (
_metadata_fb.ProcessUnitOptions.NormalizationOptions)
input_normalization.options = _metadata_fb.NormalizationOptionsT()
input_normalization.options.mean = self.model_info.mean
input_normalization.options.std = self.model_info.std
input_meta.processUnits = [input_normalization]
input_stats = _metadata_fb.StatsT()
input_stats.max = [self.model_info.image_max]
input_stats.min = [self.model_info.image_min]
input_meta.stats = input_stats
# Creates output info.
output_meta = _metadata_fb.TensorMetadataT()
output_meta.name = "probability"
output_meta.description = "Probabilities of the labels respectively."
output_meta.content = _metadata_fb.ContentT()
output_meta.content.content_properties = _metadata_fb.FeaturePropertiesT(
)
output_meta.content.contentPropertiesType = (
_metadata_fb.ContentProperties.FeatureProperties)
output_stats = _metadata_fb.StatsT()
output_stats.max = [1.0]
output_stats.min = [0.0]
output_meta.stats = output_stats
label_file = _metadata_fb.AssociatedFileT()
label_file.name = os.path.basename(self.label_file_path)
label_file.description = "Labels that %s can recognize." % model_meta.name
label_file.type = _metadata_fb.AssociatedFileType.TENSOR_AXIS_LABELS
output_meta.associatedFiles = [label_file]
# Creates subgraph info.
subgraph = _metadata_fb.SubGraphMetadataT()
subgraph.inputTensorMetadata = [input_meta]
subgraph.outputTensorMetadata = [output_meta]
model_meta.subgraphMetadata = [subgraph]
b = flatbuffers.Builder(0)
b.Finish(model_meta.Pack(b),
_metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
self.metadata_buf = b.Output()
