-
Notifications
You must be signed in to change notification settings - Fork 276
/
types.py
572 lines (484 loc) · 22.5 KB
/
types.py
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
# Copyright 2018 Google LLC
#
# 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
#
# https://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.
"""Types."""
import abc
import datetime
import operator
from typing import Any, Callable, Dict, Iterable, List, MutableMapping, NamedTuple, Optional, TypeVar, Tuple, Union
from apache_beam.utils import shared
import numpy as np
import six
import tensorflow as tf
from tensorflow_model_analysis.proto import metrics_for_slice_pb2
class RaggedTensorValue(
NamedTuple('RaggedTensorValue', [('values', np.ndarray),
('nested_row_splits', List[np.ndarray])])):
"""RaggedTensorValue encapsulates a batch of ragged tensor values.
Attributes:
values: A np.ndarray of values.
nested_row_splits: A list of np.ndarray values representing the row splits
(one per dimension including the batch dimension).
"""
class SparseTensorValue(
NamedTuple('SparseTensorValue', [('values', np.ndarray),
('indices', np.ndarray),
('dense_shape', np.ndarray)])):
"""SparseTensorValue encapsulates a batch of sparse tensor values.
Attributes:
values: A np.ndarray of values.
indices: A np.ndarray of indices.
dense_shape: A np.ndarray representing the dense shape.
"""
class VarLenTensorValue(
NamedTuple('VarLenTensorValue', [('values', np.ndarray),
('indices', np.ndarray),
('dense_shape', np.ndarray)])):
"""VarLenTensorValue encapsulates a batch of varlen dense tensor values.
Attributes:
values: A np.ndarray of values.
indices: A np.ndarray of indices.
dense_shape: A np.ndarray representing the dense shape of the entire tensor.
Note that each row (i.e. set of values sharing the same value for the
first / batch dimension) is considered to have its own shape based on the
presence of values.
"""
def __new__(cls, values: np.ndarray, indices: np.ndarray,
dense_shape: np.ndarray):
# we keep the sparse representation despite not needing it so that we can
# convert back to TF sparse tensors for free.
if len(dense_shape) != 2:
raise ValueError('A VarLenTensorValue can only be used to represent a '
'2D tensor in which the size of the second dimension '
'varies over rows. However, the provided dense_shape '
f'({dense_shape}) implies a {dense_shape.size}D tensor')
row_index_diffs = np.diff(indices[:, 0])
column_index_diffs = np.diff(indices[:, 1])
# Enforce row-major ordering of indices by checking that row indices are
# always increasing, and column indices within the same row are also always
# increasing.
bad_index_mask = ((row_index_diffs < 0) | ((row_index_diffs == 0) &
(column_index_diffs < 0)))
if np.any(bad_index_mask):
raise ValueError('The values and indices arrays must be provided in a '
'row major order, and represent a set of variable '
'length dense lists. However, indices['
f'{np.nonzero(bad_index_mask)[0] + 1}, :] did not '
'follow this pattern. The full indices array was: '
f'{indices}.')
return super().__new__(
cls, values=values, indices=indices, dense_shape=dense_shape)
class DenseRowIterator():
"""An Iterator over rows of a VarLenTensorValue as dense np.arrays.
Because the VarLenTensorValue was created from a set of variable length
(dense) arrays, we can invert this process to turn a VarLenTensorValue back
into the original dense arrays.
"""
def __init__(self, tensor):
self._tensor = tensor
self._offset = 0
def __iter__(self):
return self
def __next__(self):
if (not self._tensor.indices.size or
self._offset >= self._tensor.dense_shape[0]):
raise StopIteration
row_mask = self._tensor.indices[:, 0] == self._offset
self._offset += 1
if not row_mask.any():
# handle empty rows
return np.array([])
# we rely on slice indexing (a[start:end] rather than fancy indexing
# (a[mask]) to avoid making a copy of each row. For details, see:
# https://scipy-cookbook.readthedocs.io/items/ViewsVsCopies.html
row_mask_indices = np.nonzero(row_mask)[0]
row_start_index = row_mask_indices[0]
row_end = row_mask_indices[-1] + 1
assert (row_end - row_start_index) == len(row_mask_indices), (
'The values for each row in the represented tensor must be '
'contiguous in the values and indices arrays but found '
f'row_start_index: {row_start_index}, row_end: {row_end}'
f'len(row_mask_indices): {len(row_mask_indices)}')
return self._tensor.values[row_start_index:row_end]
def dense_rows(self):
return self.DenseRowIterator(self)
@classmethod
def from_dense_rows(cls,
dense_rows: Iterable[np.ndarray]) -> 'VarLenTensorValue':
"""Converts a collection of variable length dense arrays into a tensor.
Args:
dense_rows: A sequence of possibly variable length 1D arrays.
Returns:
A new VarLenTensorValue containing the sparse representation of the
vertically stacked dense rows. The dense_shape attribute on the result
will be (num_rows, max_row_len).
"""
rows = []
index_arrays = []
max_row_len = 0
num_rows = 0
for i, row in enumerate(dense_rows):
num_rows += 1
if row.size:
if row.ndim <= 1:
# Add a dimension for unsized numpy array. This will solve the problem
# where scalar numpy arrays like np.array(None), np.array(0) can not
# be merged with other numpy arrays.
row = row.reshape(-1)
rows.append(row)
else:
raise ValueError(
'Each non-empty dense row should be 1D or scalar but'
f' found row with shape {row.shape}.'
)
index_arrays.append(np.array([[i, j] for j in range(len(row))]))
max_row_len = max(max_row_len, row.size)
if index_arrays:
values = np.concatenate(rows, axis=0)
indices = np.concatenate(index_arrays, axis=0)
else:
# empty case
values = np.array([])
indices = np.empty((0, 2))
dense_shape = np.array([num_rows, max_row_len])
return cls.__new__(
cls, values=values, indices=indices, dense_shape=dense_shape)
# pylint: disable=invalid-name
TensorType = Union[tf.Tensor, tf.SparseTensor, tf.RaggedTensor]
TensorOrOperationType = Union[TensorType, tf.Operation]
DictOfTensorType = Dict[str, TensorType]
TensorTypeMaybeDict = Union[TensorType, DictOfTensorType]
DictOfTensorTypeMaybeDict = Dict[str, TensorTypeMaybeDict]
TensorTypeMaybeMultiLevelDict = Union[TensorTypeMaybeDict,
DictOfTensorTypeMaybeDict]
DictOfTypeSpec = Dict[str, tf.TypeSpec]
TypeSpecMaybeDict = Union[tf.TypeSpec, DictOfTypeSpec]
DictOfTypeSpecMaybeDict = Dict[str, TypeSpecMaybeDict]
TypeSpecMaybeMultiLevelDict = Union[TypeSpecMaybeDict, DictOfTypeSpecMaybeDict]
# TODO(b/171992041): Remove tf.compat.v1.SparseTensorValue.
TensorValue = Union[np.ndarray, SparseTensorValue, RaggedTensorValue,
tf.compat.v1.SparseTensorValue]
DictOfTensorValue = Dict[str, TensorValue]
TensorValueMaybeDict = Union[TensorValue, DictOfTensorValue]
DictOfTensorValueMaybeDict = Dict[str, TensorValueMaybeDict]
TensorValueMaybeMultiLevelDict = Union[TensorValueMaybeDict,
DictOfTensorValueMaybeDict]
MetricVariablesType = List[Any]
PrimitiveMetricValueType = Union[float, int, np.number]
ConcreteStructuredMetricValue = TypeVar(
'ConcreteStructuredMetricValue', bound='StructuredMetricValue')
class StructuredMetricValue(abc.ABC):
"""The base class for all structured metrics used within TFMA.
This class allows custom metrics to control how proto serialization happens,
and how to handle basic algebraic operations used in computing confidence
intervals and model diffs. By implementing the _apply_binary_op methods,
subclasses can then be treated like primitive numeric types.
"""
@abc.abstractmethod
def to_proto(self) -> metrics_for_slice_pb2.MetricValue:
...
@abc.abstractmethod
def _apply_binary_op_elementwise(
self: ConcreteStructuredMetricValue, other: ConcreteStructuredMetricValue,
op: Callable[[float, float], float]) -> ConcreteStructuredMetricValue:
"""Applies the binary operator elementwise on self and `other`.
Given two structures of the same type, this function's job is to find
corresponding pairs of elements within both structures, invoke `op` on each
pair, and store the result in a corresponding location within a new
structure. For example, to implement for a list, this function could be
implemented as:
return [op(elem, other_elem) for elem, other_elem in zip(self, other)]
Args:
other: A structure containing elements which should be the second operand
when applying `op`. `Other` must be a structured metric of the same type
as self.
op: A binary operator which should be applied elementwise to corresponding
primitive values in self and `other`.
Returns:
A new structured metric that is the result of elementwise applying `op`
on corresponding elements within self and `other`.
"""
...
@abc.abstractmethod
def _apply_binary_op_broadcast(
self: ConcreteStructuredMetricValue, other: float,
op: Callable[[float, float], float]) -> ConcreteStructuredMetricValue:
"""Applies the binary operator on each element in self and a single float.
This function supports broadcasting operations on the structured metric by
applying `op` on each element in self, paired with the primitive value
`other`. This makes it possible do things like add a fixed quantity to every
element in a structure. For example, to implement for a list, this function
could be implemented as:
return [op(elem, other) for elem in self]
Args:
other: The value to be used as the second operand when applying `op`.
op: A binary operator which should be applied elementwise to each element
in self and `other`.
Returns:
A new structured metric that is the result of applying `op` on each
element within self and a single value, `other`.
"""
...
def _apply_binary_op(
self: ConcreteStructuredMetricValue,
other: Union[PrimitiveMetricValueType, ConcreteStructuredMetricValue],
op: Callable[[float, float], float]) -> ConcreteStructuredMetricValue:
if type(other) is type(self): # pylint: disable=unidiomatic-typecheck
return self._apply_binary_op_elementwise(other, op)
elif isinstance(other, (float, int, np.number)):
return self._apply_binary_op_broadcast(float(other), op)
else:
raise ValueError(
'Binary ops can only be applied elementwise on two instances of the '
'same StructuredMetricValue subclass or using broadcasting with one '
'StructuredMetricValue and a primitive numeric type (int, float, '
'np.number). Cannot apply binary op on objects of type '
'{} and {}'.format(type(self), type(other)))
def __add__(self: ConcreteStructuredMetricValue,
other: Union[ConcreteStructuredMetricValue, float]):
return self._apply_binary_op(other, operator.add)
def __sub__(self: ConcreteStructuredMetricValue,
other: Union[ConcreteStructuredMetricValue, float]):
return self._apply_binary_op(other, operator.sub)
def __mul__(self: ConcreteStructuredMetricValue,
other: Union[ConcreteStructuredMetricValue, float]):
return self._apply_binary_op(other, operator.mul)
def __truediv__(self: ConcreteStructuredMetricValue,
other: Union[ConcreteStructuredMetricValue, float]):
return self._apply_binary_op(other, operator.truediv)
def __pow__(self: ConcreteStructuredMetricValue,
other: Union[ConcreteStructuredMetricValue, float]):
return self._apply_binary_op(other, operator.pow)
MetricValueType = Union[PrimitiveMetricValueType, np.ndarray,
StructuredMetricValue]
class ValueWithTDistribution(
NamedTuple('ValueWithTDistribution', [
('sample_mean', MetricValueType),
('sample_standard_deviation', MetricValueType),
('sample_degrees_of_freedom', int),
('unsampled_value', MetricValueType),
])):
r"""Represents the t-distribution value.
It includes sample_mean, sample_standard_deviation,
sample_degrees_of_freedom. And also unsampled_value is also stored here to
record the value calculated without bootstrapping.
The sample_standard_deviation is calculated as:
\sqrt{ \frac{1}{N-1} \sum_{i=1}^{N}{(x_i - \bar{x})^2} }
"""
def __new__(
cls,
sample_mean: float,
sample_standard_deviation: Optional[float] = None,
sample_degrees_of_freedom: Optional[int] = None,
unsampled_value: Optional[float] = None,
):
return super(ValueWithTDistribution,
cls).__new__(cls, sample_mean, sample_standard_deviation,
sample_degrees_of_freedom, unsampled_value)
def __float__(self):
# unsampled_value can be numpy.float which is a subclass of float, but here
# need to return a strict float.
return float(self.unsampled_value)
# AddMetricsCallback should have the following prototype:
# def add_metrics_callback(features_dict, predictions_dict, labels_dict):
#
# It should create and return a metric_ops dictionary, such that
# metric_ops['metric_name'] = (value_op, update_op), just as in the Trainer.
#
# Note that features_dict, predictions_dict and labels_dict are not
# necessarily dictionaries - they might also be Tensors, depending on what the
# model's eval_input_receiver_fn returns.
# pyformat: disable
AddMetricsCallbackType = Callable[[
TensorTypeMaybeDict, TensorTypeMaybeDict, TensorTypeMaybeDict
], Dict[str, Tuple[TensorType, TensorType]]]
# pyformat: enable
# Type of keys we support for prediction, label and features dictionaries.
FPLKeyType = Union[str, Tuple[str, ...]]
# Dictionary of Tensor values fetched. The dictionary maps original dictionary
# keys => ('node' => value). This type exists for backward compatibility with
# FeaturesPredictionsLabels, new code should use DictOfTensorValue instead.
DictOfFetchedTensorValues = Dict[FPLKeyType, Dict[str, TensorValue]]
FeaturesPredictionsLabels = NamedTuple(
'FeaturesPredictionsLabels', [('input_ref', int),
('features', DictOfFetchedTensorValues),
('predictions', DictOfFetchedTensorValues),
('labels', DictOfFetchedTensorValues)])
# Used in building the model diagnostics table, a MaterializedColumn is a value
# inside of Extracts that will be emitted to file. Note that for strings, the
# values are raw byte strings rather than unicode strings. This is by design, as
# features can have arbitrary bytes values.
MaterializedColumn = NamedTuple(
'MaterializedColumn',
[('name', str),
('value', Union[List[bytes], List[int], List[float], bytes, int, float])])
# Extracts represent data extracted during pipeline processing. In order to
# provide a flexible API, these types are just dicts where the keys are defined
# (reserved for use) by different extractor implementations. For example, the
# FeaturesExtractor stores the data for the features under the key "features",
# LabelsExtractor stores the data for the labels under the key "labels", etc.
Extracts = MutableMapping[str, Any]
# pylint: enable=invalid-name
class ModelLoader:
"""Model loader is responsible for loading shared model types.
Attributes:
construct_fn: A callable which creates the model instance. The callable
should take no args as input (typically a closure is used to capture
necessary parameters).
tags: Optional model tags (e.g. 'serve' for serving or 'eval' for
EvalSavedModel).
"""
__slots__ = ['construct_fn', 'tags', '_shared_handle']
def __init__(self,
construct_fn: Callable[[], Any],
tags: Optional[List[str]] = None):
self.construct_fn = construct_fn
self.tags = tags
self._shared_handle = shared.Shared()
def load(
self,
model_load_time_callback: Optional[Callable[[int], None]] = None) -> Any:
"""Returns loaded model.
Args:
model_load_time_callback: Optional callback to track load time.
"""
if model_load_time_callback:
construct_fn = self._construct_fn_with_load_time(model_load_time_callback)
else:
construct_fn = self.construct_fn
return self._shared_handle.acquire(construct_fn)
def _construct_fn_with_load_time(
self, model_load_time_callback: Callable[[int],
None]) -> Callable[[], Any]:
"""Wraps actual construct fn to allow for load time metrics."""
def with_load_times():
start_time = datetime.datetime.now()
model = self.construct_fn()
end_time = datetime.datetime.now()
model_load_time_callback(int((end_time - start_time).total_seconds()))
return model
return with_load_times
class EvalSharedModel(
NamedTuple(
'EvalSharedModel',
[
('model_path', str),
(
'add_metrics_callbacks',
List[Callable],
), # List[AnyMetricsCallbackType]
('include_default_metrics', bool),
('example_weight_key', Union[str, Dict[str, str]]),
('additional_fetches', List[str]),
('model_loader', ModelLoader),
('model_name', str),
('model_type', str),
('rubber_stamp', bool),
('is_baseline', bool),
('resource_hints', Optional[Dict[str, Any]]),
('backend_config', Optional[Any]),
],
)
):
# pyformat: disable
"""Shared model used during extraction and evaluation.
Attributes:
model_path: Path to EvalSavedModel (containing the saved_model.pb file).
add_metrics_callbacks: Optional list of callbacks for adding additional
metrics to the graph. The names of the metrics added by the callbacks
should not conflict with existing metrics. See below for more details
about what each callback should do. The callbacks are only used during
evaluation.
include_default_metrics: True to include the default metrics that are part
of the saved model graph during evaluation.
example_weight_key: Example weight key (single-output model) or dict of
example weight keys (multi-output model) keyed by output_name.
additional_fetches: Prefixes of additional tensors stored in
signature_def.inputs that should be fetched at prediction time. The
"features" and "labels" tensors are handled automatically and should not
be included in this list.
model_loader: Model loader.
model_name: Model name (should align with ModelSpecs.name).
model_type: Model type (tfma.TF_KERAS, tfma.TF_LITE, tfma.TF_ESTIMATOR, ..).
rubber_stamp: True if this model is being rubber stamped. When a
model is rubber stamped diff thresholds will be ignored if an associated
baseline model is not passed.
is_baseline: The model is the baseline for comparison or not.
resource_hints: The beam resource hints to apply to the PTransform which
runs inference for this model.
backend_config: The backend config for running model inference.
More details on add_metrics_callbacks:
Each add_metrics_callback should have the following prototype:
def add_metrics_callback(features_dict, predictions_dict, labels_dict):
Note that features_dict, predictions_dict and labels_dict are not
necessarily dictionaries - they might also be Tensors, depending on what the
model's eval_input_receiver_fn returns.
It should create and return a metric_ops dictionary, such that
metric_ops['metric_name'] = (value_op, update_op), just as in the Trainer.
Short example:
def add_metrics_callback(features_dict, predictions_dict, labels):
metrics_ops = {}
metric_ops['mean_label'] = tf.metrics.mean(labels)
metric_ops['mean_probability'] = tf.metrics.mean(tf.slice(
predictions_dict['probabilities'], [0, 1], [2, 1]))
return metric_ops
"""
# pyformat: enable
def __new__(
cls,
model_path: Optional[str] = None,
add_metrics_callbacks: Optional[List[AddMetricsCallbackType]] = None,
include_default_metrics: Optional[bool] = True,
example_weight_key: Optional[Union[str, Dict[str, str]]] = None,
additional_fetches: Optional[List[str]] = None,
model_loader: Optional[ModelLoader] = None,
model_name: str = '',
model_type: str = '',
rubber_stamp: bool = False,
is_baseline: bool = False,
resource_hints: Optional[Dict[str, Any]] = None,
backend_config: Optional[Any] = None,
construct_fn: Optional[Callable[[], Any]] = None,
):
if not add_metrics_callbacks:
add_metrics_callbacks = []
if model_loader and construct_fn:
raise ValueError(
'only one of model_loader or construct_fn should be used')
if construct_fn:
model_loader = ModelLoader(tags=None, construct_fn=construct_fn)
if model_path is not None:
model_path = six.ensure_str(model_path)
if is_baseline and rubber_stamp:
raise ValueError('Baseline model cannot be rubber stamped.')
return super(EvalSharedModel, cls).__new__(
cls,
model_path,
add_metrics_callbacks,
include_default_metrics,
example_weight_key,
additional_fetches,
model_loader,
model_name,
model_type,
rubber_stamp,
is_baseline,
resource_hints,
backend_config,
)
# MaybeMultipleEvalSharedModels represents a parameter that can take on a single
# model or a list of models.
#
# TODO(b/150416505): Deprecate support for dict.
MaybeMultipleEvalSharedModels = Union[EvalSharedModel, List[EvalSharedModel],
Dict[str, EvalSharedModel]]