-
Notifications
You must be signed in to change notification settings - Fork 5.2k
/
metadata_schema.fbs
732 lines (648 loc) · 27.4 KB
/
metadata_schema.fbs
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
// Copyright 2022 The MediaPipe Authors.
//
// 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.
namespace tflite;
// TFLite metadata contains both human readable and machine readable information
// about what the model does and how to use the model. It can be used as a
// README file, which elaborates the details of the model, each input/ouput
// tensor, and each associated file.
//
// An important use case of TFLite metadata is the TFLite codegen tool, which
// automatically generates the model interface based on the properties of the
// model and the tensors. The model interface provides high-level APIs to
// interact with the model, such as preprocessing the input data and running
// inferences.
//
// Entries marked with "<Codegen usage>" are used in TFLite codegen tool to
// generate the model interface. It is recommended to fill in at least those
// enties to boost the codegen performance.
// The Metadata schema is versioned by the Semantic versioning number, such as
// MAJOR.MINOR.PATCH. It tracks the schema changes according to the rules below:
// * Bump up the MAJOR number when making potentially backwards incompatible
// changes. It must be incremented if the new changes break the backwards
// compatibility. It may also include minor and patch level changes as
// needed. The true backwards compatibility is indicated by the file
// identifier.
// * Bump up the MINOR number when making backwards compatible updates for
// major features, such as supporting new content types or adding new
// processing units.
// * Bump up the PATCH number when making small backwards compatible changes,
// such as adding a new fields or deprecating certain fields (not deleting
// them).
//
// ModelMetadata.min_parser_version indicates the minimum necessary metadata
// parser version to fully understand all fields in a given metadata flatbuffer.
//
// New fields and types will have associated comments with the schema version
// for which they were added.
//
// Schema Semantic version: 1.5.0
// This indicates the flatbuffer compatibility. The number will bump up when a
// break change is applied to the schema, such as removing fields or adding new
// fields to the middle of a table.
file_identifier "M001";
// History:
// 1.0.1 - Added VOCABULARY type to AssociatedFileType.
// 1.1.0 - Added BertTokenizerOptions to ProcessUnitOptions.
// Added SentencePieceTokenizerOptions to ProcessUnitOptions.
// Added input_process_units to SubGraphMetadata.
// Added output_process_units to SubGraphMetadata.
// 1.2.0 - Added input_tensor_group to SubGraphMetadata.
// Added output_tensor_group to SubGraphMetadata.
// 1.2.1 - Added RegexTokenizerOptions to ProcessUnitOptions.
// 1.3.0 - Added AudioProperties to ContentProperties.
// 1.4.0 - Added SCANN_INDEX_FILE type to AssociatedFileType.
// 1.4.1 - Added version to AssociatedFile.
// 1.5.0 - Added CustomMetadata in SubGraphMetadata.
// File extension of any written files.
file_extension "tflitemeta";
enum AssociatedFileType : byte {
UNKNOWN = 0,
// Files such as readme.txt.
DESCRIPTIONS = 1,
// Contains a list of labels (characters separated by "\n" or in lines) that
// annotate certain axis of the tensor. For example,
// the label file in image classification. Those labels annotate the
// the output tensor, such that each value in the output tensor is the
// probability of that corresponding category specified by the label. See the
// example label file used in image classification [1].
//
// <Codegen usage>:
// If an output tensor has an associated file as TENSOR_AXIS_LABELS, return
// the output as a mapping between the labels and probability in the model
// interface.
// If multiple files of the same type are present, the first one is used by
// default; additional ones are to be distinguished from one another by their
// specified locale.
//
// TODO: Add github example link.
TENSOR_AXIS_LABELS = 2,
// Contains a list of labels (characters separated by "\n" or in lines) that
// tensor values correspond to. For example, in
// the object detection model, one of the output tensors is the detected
// classes. And each value in the tensor refers to the index of label in the
// category label file. See the example label file used in object detection
// [1].
//
// <Codegen usage>:
// If an output tensor has an associated file as TENSOR_VALUE_LABELS, convert
// the tensor values into labels, and return a list of string as the output.
// If multiple files of the same type are present, the first one is used by
// default; additional ones are to be distinguished from one another by their
// specified locale.
//
// TODO: Add github example link.
TENSOR_VALUE_LABELS = 3,
// Contains sigmoid-based score calibration parameters, formatted as CSV.
// Lines contain for each index of an output tensor the scale, slope, offset
// and (optional) min_score parameters to be used for sigmoid fitting (in this
// order and in `strtof`-compatible [1] format). Scale should be a
// non-negative value.
// A line may be left empty to default calibrated scores for this index to
// default_score.
// In summary, each line should thus contain 0, 3 or 4 comma-separated values.
//
// See the example score calibration file used in image classification [2].
//
// See documentation for ScoreCalibrationOptions for details.
//
// [1]: https://en.cppreference.com/w/c/string/byte/strtof
// TODO: Add github example link.
TENSOR_AXIS_SCORE_CALIBRATION = 4,
// Contains a list of unique words (characters separated by "\n" or in lines)
// that help to convert natural language words to embedding vectors.
//
// See the example vocab file used in text classification [1].
//
// TODO: Add github example link.
// Added in: 1.0.1
VOCABULARY = 5,
// TODO: introduce the ScaNN index file with links once the code
// is released.
// Contains on-device ScaNN index file with LevelDB format.
// Added in: 1.4.0
SCANN_INDEX_FILE = 6,
}
table AssociatedFile {
// Name of this file. Need to be exact the same as the name of the actual file
// packed into the TFLite model as a zip file.
//
// <Codegen usage>:
// Locates to the actual file in the TFLite model.
name:string;
// A description of what the file is.
description:string;
// Type of the associated file. There may be special pre/post processing for
// some types. For example in image classification, a label file of the output
// will be used to convert object index into string.
//
// <Codegen usage>:
// Determines how to process the corresponding tensor.
type:AssociatedFileType;
// An optional locale for this associated file (if applicable). It is
// recommended to use an ISO 639-1 letter code (e.g. "en" for English),
// optionally completed by a two letter region code (e.g. "en-US" for US
// English and "en-CA" for Canadian English).
// Leverage this in order to specify e.g multiple label files translated in
// different languages.
locale:string;
// Version of the file specified by model creators.
// Added in: 1.4.1
version:string;
}
// The basic content type for all tensors.
//
// <Codegen usage>:
// Input feature tensors:
// 1. Generates the method to load data from a TensorBuffer.
// 2. Creates the preprocessing logic. The default processing pipeline is:
// [NormalizeOp, QuantizeOp].
// Output feature tensors:
// 1. Generates the method to return the output data to a TensorBuffer.
// 2. Creates the post-processing logic. The default processing pipeline is:
// [DeQuantizeOp].
table FeatureProperties {
}
// The type of color space of an image.
enum ColorSpaceType : byte {
UNKNOWN = 0,
RGB = 1,
GRAYSCALE = 2,
}
table ImageSize {
width:uint;
height:uint;
}
// The properties for image tensors.
//
// <Codegen usage>:
// Input image tensors:
// 1. Generates the method to load an image from a TensorImage.
// 2. Creates the preprocessing logic. The default processing pipeline is:
// [ResizeOp, NormalizeOp, QuantizeOp].
// Output image tensors:
// 1. Generates the method to return the output data to a TensorImage.
// 2. Creates the post-processing logic. The default processing pipeline is:
// [DeQuantizeOp].
table ImageProperties {
// The color space of the image.
//
// <Codegen usage>:
// Determines how to convert the color space of a given image from users.
color_space:ColorSpaceType;
// Indicates the default value of image width and height if the tensor shape
// is dynamic. For fixed-size tensor, this size will be consistent with the
// expected size.
default_size:ImageSize;
}
// The properties for tensors representing bounding boxes.
//
// <Codegen usage>:
// Input image tensors: NA.
// Output image tensors: parses the values into a data structure that represents
// bounding boxes. For example, in the generated wrapper for Android, it returns
// the output as android.graphics.Rect objects.
enum BoundingBoxType : byte {
UNKNOWN = 0,
// Represents the bounding box by using the combination of boundaries,
// {left, top, right, bottom}.
// The default order is {left, top, right, bottom}. Other orders can be
// indicated by BoundingBoxProperties.index.
BOUNDARIES = 1,
// Represents the bounding box by using the upper_left corner, width and
// height.
// The default order is {upper_left_x, upper_left_y, width, height}. Other
// orders can be indicated by BoundingBoxProperties.index.
UPPER_LEFT = 2,
// Represents the bounding box by using the center of the box, width and
// height. The default order is {center_x, center_y, width, height}. Other
// orders can be indicated by BoundingBoxProperties.index.
CENTER = 3,
}
// The properties for audio tensors.
// Added in: 1.3.0
table AudioProperties {
// The sample rate in Hz when the audio was captured.
sample_rate:uint;
// The channel count of the audio.
channels:uint;
}
enum CoordinateType : byte {
// The coordinates are float values from 0 to 1.
RATIO = 0,
// The coordinates are integers.
PIXEL = 1,
}
table BoundingBoxProperties {
// Denotes the order of the elements defined in each bounding box type. An
// empty index array represent the default order of each bounding box type.
// For example, to denote the default order of BOUNDARIES, {left, top, right,
// bottom}, the index should be {0, 1, 2, 3}. To denote the order {left,
// right, top, bottom}, the order should be {0, 2, 1, 3}.
//
// The index array can be applied to all bounding box types to adjust the
// order of their corresponding underlying elements.
//
// <Codegen usage>:
// Indicates how to parse the bounding box values.
index:[uint];
// <Codegen usage>:
// Indicates how to parse the bounding box values.
type:BoundingBoxType;
// <Codegen usage>:
// Indicates how to convert the bounding box back to the original image in
// pixels.
coordinate_type:CoordinateType;
}
union ContentProperties {
FeatureProperties,
ImageProperties,
BoundingBoxProperties,
// Added in: 1.3.0
AudioProperties,
}
table ValueRange {
min:int;
max:int;
}
table Content {
// The properties that the content may have, indicating the type of the
// Content.
//
// <Codegen usage>:
// Indicates how to process the tensor.
content_properties:ContentProperties;
// The range of dimensions that the content corresponds to. A NULL
// "range" indicates that the content uses up all dimensions,
// except the batch axis if applied.
//
// Here are all the possible situations of how a tensor is composed.
// Case 1: The tensor is a single object, such as an image.
// For example, the input of an image classifier
// (https://www.tensorflow.org/lite/models/image_classification/overview),
// a tensor of shape [1, 224, 224, 3]. Dimensions 1 to 3 correspond to the
// image. Since dimension 0 is a batch axis, which can be ignored,
// "range" can be left as NULL.
//
// Case 2: The tensor contains multiple instances of the same object.
// For example, the output tensor of detected bounding boxes of an object
// detection model
// (https://www.tensorflow.org/lite/models/object_detection/overview).
// The tensor shape is [1, 10, 4]. Here is the what the three dimensions
// represent for:
// dimension 0: the batch axis.
// dimension 1: the 10 objects detected with the highest confidence.
// dimension 2: the bounding boxes of the 10 detected objects.
// The tensor is essentially 10 bounding boxes. In this case,
// "range" should be {min=2; max=2;}.
//
// The output tensor of scores of the above object detection model has shape
// [1, 10], where
// dimension 0: the batch axis;
// dimension 1: the scores of the 10 detected objects.
// Set "range" to the number of dimensions which is {min=2; max=2;} to denote
// that every element in the tensor is an individual content object, i.e. a
// score in this example.
//
// Another example is the pose estimation model
// (https://www.tensorflow.org/lite/models/pose_estimation/overview).
// The output tensor of heatmaps is in the shape of [1, 9, 9, 17].
// Here is the what the four dimensions represent for:
// dimension 0: the batch axis.
// dimension 1/2: the heatmap image.
// dimension 3: 17 body parts of a person.
// Even though the last axis is body part, the real content of this tensor is
// the heatmap. "range" should be [min=1; max=2].
//
// Case 3: The tensor contains multiple different objects. (Not supported by
// Content at this point).
// Sometimes a tensor may contain multiple different objects, thus different
// contents. It is very common for regression models. For example, a model
// to predict the fuel efficiency
// (https://www.tensorflow.org/tutorials/keras/regression).
// The input tensor has shape [1, 9], consisting of 9 features, such as
// "Cylinders", "Displacement", "Weight", etc. In this case, dimension 1
// contains 9 different contents. However, since these sub-dimension objects
// barely need to be specifically processed, their contents are not recorded
// in the metadata. Through, the name of each dimension can be set through
// TensorMetadata.dimension_names.
//
// Note that if it is not case 3, a tensor can only have one content type.
//
// <Codegen usage>:
// Case 1: return a processed single object of certain content type.
// Case 2: return a list of processed objects of certain content type. The
// generated model interface have API to random access those objects from
// the output.
range:ValueRange;
}
// Parameters that are used when normalizing the tensor.
table NormalizationOptions{
// mean and std are normalization parameters. Tensor values are normalized
// on a per-channel basis, by the formula
// (x - mean) / std.
// If there is only one value in mean or std, we'll propagate the value to
// all channels.
//
// Quantized models share the same normalization parameters as their
// corresponding float models. For example, an image input tensor may have
// the normalization parameter of
// mean = 127.5f and std = 127.5f.
// The image value will be normalized from [0, 255] to [-1, 1].
// Then, for quantized models, the image data should be further quantized
// according to the quantization parameters. In the case of uint8, the image
// data will be scaled back to [0, 255], while for int8, the image data will
// be scaled to [-128, 127].
//
// Both the normalization parameters and quantization parameters can be
// retrieved through the metadata extractor library.
// TODO: add link for the metadata extractor library.
// Per-channel mean of the possible values used in normalization.
//
// <Codegen usage>:
// Apply normalization to input tensors accordingly.
mean:[float];
// Per-channel standard dev. of the possible values used in normalization.
//
// <Codegen usage>:
// Apply normalization to input tensors accordingly.
std:[float];
}
// The different possible score transforms to apply to uncalibrated scores
// before applying score calibration.
enum ScoreTransformationType : byte {
// Identity function: g(x) = x.
IDENTITY = 0,
// Log function: g(x) = log(x).
LOG = 1,
// Inverse logistic function: g(x) = log(x) - log(1-x).
INVERSE_LOGISTIC = 2,
}
// Options to perform score calibration on an output tensor through sigmoid
// functions. One of the main purposes of score calibration is to make scores
// across classes comparable, so that a common threshold can be used for all
// output classes. This is meant for models producing class predictions as
// output, e.g. image classification or detection models.
//
// For each index in the output tensor, this applies:
// * `f(x) = scale / (1 + e^-(slope*g(x)+offset))` if `x > min_score` or if no
// `min_score` has been specified,
// * `f(x) = default_score` otherwise or if no scale, slope and offset have been
// specified.
// Where:
// * scale, slope, offset and (optional) min_score are index-specific parameters
// * g(x) is an index-independent transform among those defined in
// ScoreTransformationType
// * default_score is an index-independent parameter.
// An AssociatedFile with type TANSOR_AXIS_SCORE_CALIBRATION specifying the
// index-specific parameters must be associated with the corresponding
// TensorMetadata for score calibration be applied.
//
// See the example score calibration file used in image classification [1].
// TODO: Add github example link.
table ScoreCalibrationOptions {
// The function to use for transforming the uncalibrated score before
// applying score calibration.
score_transformation:ScoreTransformationType;
// The default calibrated score to apply if the uncalibrated score is
// below min_score or if no parameters were specified for a given index.
default_score:float;
}
// Performs thresholding on output tensor values, in order to filter out
// low-confidence results.
table ScoreThresholdingOptions {
// The recommended global threshold below which results are considered
// low-confidence and should be filtered out.
global_score_threshold:float;
}
// Performs Bert tokenization as in tf.text.BertTokenizer
// (https://github.com/tensorflow/text/blob/3599f6fcd2b780a2dc413b90fb9315464f10b314/docs/api_docs/python/text/BertTokenizer.md)
// Added in: 1.1.0
table BertTokenizerOptions {
// The vocabulary files used in the BertTokenizer.
vocab_file:[AssociatedFile];
}
// Performs SentencePiece tokenization as in tf.text.SentencepieceTokenizer
// (https://github.com/tensorflow/text/blob/3599f6fcd2b780a2dc413b90fb9315464f10b314/docs/api_docs/python/text/SentencepieceTokenizer.md).
// Added in: 1.1.0
table SentencePieceTokenizerOptions {
// The SentencePiece model files used in the SentencePieceTokenizer.
sentencePiece_model:[AssociatedFile];
// The optional vocabulary model files used in the SentencePieceTokenizer.
vocab_file:[AssociatedFile];
}
// Splits strings by the occurrences of delim_regex_pattern and converts the
// tokens into ids. For example, given
// delim_regex_pattern: "\W+",
// string: "Words, words, words.",
// the tokens after split are: "Words", "words", "words", "".
// And then the tokens can be converted into ids according to the vocab_file.
// Added in: 1.2.1
table RegexTokenizerOptions {
delim_regex_pattern:string;
// The vocabulary files used to convert this tokens into ids.
vocab_file:[AssociatedFile];
}
// Options that are used when processing the tensor.
union ProcessUnitOptions {
NormalizationOptions,
ScoreCalibrationOptions,
ScoreThresholdingOptions,
// Added in: 1.1.0
BertTokenizerOptions,
// Added in: 1.1.0
SentencePieceTokenizerOptions,
// Added in: 1.2.1
RegexTokenizerOptions
}
// A process unit that is used to process the tensor out-of-graph.
table ProcessUnit {
options:ProcessUnitOptions;
}
// Statistics to describe a tensor.
table Stats {
// Max and min are not currently used in tflite.support codegen. They mainly
// serve as references for users to better understand the model. They can also
// be used to validate model pre/post processing results.
// If there is only one value in max or min, we'll propagate the value to
// all channels.
// Per-channel maximum value of the tensor.
max:[float];
// Per-channel minimum value of the tensor.
min:[float];
}
// Metadata of a group of tensors. It may contain several tensors that will be
// grouped together in codegen. For example, the TFLite object detection model
// example (https://www.tensorflow.org/lite/models/object_detection/overview)
// has four outputs: classes, scores, bounding boxes, and number of detections.
// If the four outputs are bundled together using TensorGroup (for example,
// named as "detection result"), the codegen tool will generate the class,
// `DetectionResult`, which contains the class, score, and bounding box. And the
// outputs of the model will be converted to a list of `DetectionResults` and
// the number of detection. Note that the number of detection is a single
// number, therefore is inappropriate for the list of `DetectionResult`.
// Added in: 1.2.0
table TensorGroup {
// Name of tensor group.
//
// <codegen usage>:
// Name of the joint class of the tensor group.
name:string;
// Names of the tensors to group together, corresponding to
// TensorMetadata.name.
//
// <codegen usage>:
// Determines which tensors will be added to this group. All tensors in the
// group should have the same number of elements specified by Content.range.
tensor_names:[string];
}
// Detailed information of an input or output tensor.
table TensorMetadata {
// Name of the tensor.
//
// <Codegen usage>:
// The name of this tensor in the generated model interface.
name:string;
// A description of the tensor.
description:string;
// A list of names of the dimensions in this tensor. The length of
// dimension_names need to match the number of dimensions in this tensor.
//
// <Codegen usage>:
// The name of each dimension in the generated model interface. See "Case 2"
// in the comments of Content.range.
dimension_names:[string];
// The content that represents this tensor.
//
// <Codegen usage>:
// Determines how to process this tensor. See each item in ContentProperties
// for the default process units that will be applied to the tensor.
content:Content;
// The process units that are used to process the tensor out-of-graph.
//
// <Codegen usage>:
// Contains the parameters of the default processing pipeline for each content
// type, such as the normalization parameters in all content types. See the
// items under ContentProperties for the details of the default processing
// pipeline.
process_units:[ProcessUnit];
// The statistics of the tensor values.
stats:Stats;
// A list of associated files of this tensor.
//
// <Codegen usage>:
// Contains processing parameters of this tensor, such as normalization.
associated_files:[AssociatedFile];
}
table CustomMetadata {
name:string;
data:[ubyte] (force_align: 16);
}
table SubGraphMetadata {
// Name of the subgraph.
//
// Note that, since TFLite only support one subgraph at this moment, the
// Codegen tool will use the name in ModelMetadata in the generated model
// interface.
name:string;
// A description explains details about what the subgraph does.
description:string;
// Metadata of all input tensors used in this subgraph. It matches exactly
// the input tensors specified by `SubGraph.inputs` in the TFLite
// schema.fbs file[2]. The number of `TensorMetadata` in the array should
// equal to the number of indices in `SubGraph.inputs`.
//
// [2]: tensorflow/lite/schema/schema.fbs
// <Codegen usage>:
// Determines how to process the inputs.
input_tensor_metadata:[TensorMetadata];
// Metadata of all output tensors used in this subgraph. It matches exactly
// the output tensors specified by `SubGraph.outputs` in the TFLite
// schema.fbs file[2]. The number of `TensorMetadata` in the array should
// equal to the number of indices in `SubGraph.outputs`.
//
// <Codegen usage>:
// Determines how to process the outputs.
output_tensor_metadata:[TensorMetadata];
// A list of associated files of this subgraph.
associated_files:[AssociatedFile];
// Input process units of the subgraph. Some models may have complex pre and
// post processing logics where the process units do not work on one tensor at
// a time, but in a similar way of a TFLite graph. For example, in the
// MobileBert model (https://www.tensorflow.org/lite/models/bert_qa/overview),
// the inputs are: ids / mask / segment ids;
// the outputs are: end logits / start logits.
// The preprocessing converts the query string and the context string to the
// model inputs, and the post-processing converts the model outputs to the
// answer string.
// Added in: 1.1.0
input_process_units:[ProcessUnit];
// Output process units of the subgraph.
// Added in: 1.1.0
output_process_units:[ProcessUnit];
// Metadata of all input tensor groups used in this subgraph.
//
// <codegen usage>:
// Bundles the corresponding elements of the underlying input tensors together
// into a class, and converts those individual tensors into a list of the
// class objects.
// Added in: 1.2.0
input_tensor_groups:[TensorGroup];
// Metadata of all output tensor groups used in this subgraph.
//
// <codegen usage>:
// Bundles the corresponding elements of the underlying output tensors
// together into a class, and converts those individual tensors into a list of
// the class objects.
// Added in: 1.2.0
output_tensor_groups:[TensorGroup];
// A list of custom metadata.
// Added in: 1.5.0.
custom_metadata:[CustomMetadata];
}
table ModelMetadata {
// Name of the model.
//
// <Codegen usage>:
// The name of the model in the generated model interface.
name:string;
// Model description in schema.
description:string;
// Version of the model that specified by model creators.
version:string;
// Noted that, the minimum required TFLite runtime version that the model is
// compatible with, has already been added as a metadata entry in tflite
// schema. We'll decide later if we want to move it here, and keep it with
// other metadata entries.
// Metadata of all the subgraphs of the model. The 0th is assumed to be the
// main subgraph.
//
// <Codegen usage>:
// Determines how to process the inputs and outputs.
subgraph_metadata:[SubGraphMetadata];
// The person who creates this model.
author:string;
// Licenses that may apply to this model.
license:string;
// A list of associated files of this model.
associated_files:[AssociatedFile];
// The minimum metadata parser version that can fully understand the fields in
// the metadata flatbuffer. The version is effectively the largest version
// number among the versions of all the fields populated and the smallest
// compatible version indicated by the file identifier.
//
// This field is automatically populated by the MetadataPopulator when
// the metadata is populated into a TFLite model.
min_parser_version:string;
}
root_type ModelMetadata;