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
|
# Copyright 2017 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.
# ==============================================================================
"""Deep Neural Network estimators with layer annotations."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import contextlib
import pickle
from google.protobuf.any_pb2 import Any
from tensorflow.python.estimator import estimator
from tensorflow.python.estimator import model_fn
from tensorflow.python.estimator.canned import dnn
from tensorflow.python.feature_column import feature_column as feature_column_lib
from tensorflow.python.framework import ops
from tensorflow.python.ops import nn
from tensorflow.python.ops.losses import losses
from tensorflow.python.saved_model import utils as saved_model_utils
class LayerAnnotationsCollectionNames(object):
"""Names for the collections containing the annotations."""
UNPROCESSED_FEATURES = 'layer_annotations/unprocessed_features'
PROCESSED_FEATURES = 'layer_annotatons/processed_features'
FEATURE_COLUMNS = 'layer_annotations/feature_columns'
@classmethod
def keys(cls, collection_name):
return '%s/keys' % collection_name
@classmethod
def values(cls, collection_name):
return '%s/values' % collection_name
def serialize_feature_column(feature_column):
if isinstance(feature_column, feature_column_lib._EmbeddingColumn): # pylint: disable=protected-access
# We can't pickle nested functions, and we don't need the value of
# layer_creator in most cases anyway, so just discard its value.
args = feature_column._asdict()
args['layer_creator'] = None
temp = type(feature_column)(**args)
return pickle.dumps(temp)
return pickle.dumps(feature_column)
def _to_any_wrapped_tensor_info(tensor):
"""Converts a `Tensor` to a `TensorInfo` wrapped in a proto `Any`."""
any_buf = Any()
tensor_info = saved_model_utils.build_tensor_info(tensor)
any_buf.Pack(tensor_info)
return any_buf
def make_input_layer_with_layer_annotations(original_input_layer, mode):
"""Make an input_layer replacement function that adds layer annotations."""
def input_layer_with_layer_annotations(features,
feature_columns,
weight_collections=None,
trainable=True,
cols_to_vars=None,
cols_to_output_tensors=None):
"""Returns a dense `Tensor` as input layer based on given `feature_columns`.
Generally a single example in training data is described with
FeatureColumns.
At the first layer of the model, this column oriented data should be
converted
to a single `Tensor`.
This is like tf.feature_column.input_layer, except with added
Integrated-Gradient annotations.
Args:
features: A mapping from key to tensors. `_FeatureColumn`s look up via
these keys. For example `numeric_column('price')` will look at 'price'
key in this dict. Values can be a `SparseTensor` or a `Tensor` depends
on corresponding `_FeatureColumn`.
feature_columns: An iterable containing the FeatureColumns to use as
inputs to your model. All items should be instances of classes derived
from `_DenseColumn` such as `numeric_column`, `embedding_column`,
`bucketized_column`, `indicator_column`. If you have categorical
features, you can wrap them with an `embedding_column` or
`indicator_column`.
weight_collections: A list of collection names to which the Variable will
be added. Note that variables will also be added to collections
`tf.GraphKeys.GLOBAL_VARIABLES` and `ops.GraphKeys.MODEL_VARIABLES`.
trainable: If `True` also add the variable to the graph collection
`GraphKeys.TRAINABLE_VARIABLES` (see `tf.Variable`).
cols_to_vars: If not `None`, must be a dictionary that will be filled with
a mapping from `_FeatureColumn` to list of `Variable`s. For example,
after the call, we might have cols_to_vars = {_EmbeddingColumn(
categorical_column=_HashedCategoricalColumn( key='sparse_feature',
hash_bucket_size=5, dtype=tf.string), dimension=10): [<tf.Variable
'some_variable:0' shape=(5, 10), <tf.Variable 'some_variable:1'
shape=(5, 10)]} If a column creates no variables, its value will be an
empty list.
cols_to_output_tensors: If not `None`, must be a dictionary that will be
filled with a mapping from '_FeatureColumn' to the associated output
`Tensor`s.
Returns:
A `Tensor` which represents input layer of a model. Its shape
is (batch_size, first_layer_dimension) and its dtype is `float32`.
first_layer_dimension is determined based on given `feature_columns`.
Raises:
ValueError: features and feature_columns have different lengths.
"""
local_cols_to_output_tensors = {}
input_layer = original_input_layer(
features=features,
feature_columns=feature_columns,
weight_collections=weight_collections,
trainable=trainable,
cols_to_vars=cols_to_vars,
cols_to_output_tensors=local_cols_to_output_tensors)
if cols_to_output_tensors is not None:
cols_to_output_tensors = local_cols_to_output_tensors
if mode and mode == model_fn.ModeKeys.PREDICT:
# Only annotate in PREDICT mode.
# Annotate features.
# These are the parsed Tensors, before embedding.
# Only annotate features used by FeatureColumns.
# We figure which ones are used by FeatureColumns by creating a parsing
# spec and looking at the keys.
spec = feature_column_lib.make_parse_example_spec(feature_columns)
for key in spec.keys():
tensor = features[key]
ops.add_to_collection(
LayerAnnotationsCollectionNames.keys(
LayerAnnotationsCollectionNames.UNPROCESSED_FEATURES), key)
ops.add_to_collection(
LayerAnnotationsCollectionNames.values(
LayerAnnotationsCollectionNames.UNPROCESSED_FEATURES),
_to_any_wrapped_tensor_info(tensor))
# Annotate feature columns.
for column in feature_columns:
# TODO(cyfoo): Find a better way to serialize and deserialize
# _FeatureColumn.
ops.add_to_collection(LayerAnnotationsCollectionNames.FEATURE_COLUMNS,
serialize_feature_column(column))
for column, tensor in local_cols_to_output_tensors.items():
ops.add_to_collection(
LayerAnnotationsCollectionNames.keys(
LayerAnnotationsCollectionNames.PROCESSED_FEATURES),
column.name)
ops.add_to_collection(
LayerAnnotationsCollectionNames.values(
LayerAnnotationsCollectionNames.PROCESSED_FEATURES),
_to_any_wrapped_tensor_info(tensor))
return input_layer
return input_layer_with_layer_annotations
@contextlib.contextmanager
def _monkey_patch(module, function, replacement):
old_function = getattr(module, function)
setattr(module, function, replacement)
yield
setattr(module, function, old_function)
def DNNClassifierWithLayerAnnotations( # pylint: disable=invalid-name
hidden_units,
feature_columns,
model_dir=None,
n_classes=2,
weight_column=None,
label_vocabulary=None,
optimizer='Adagrad',
activation_fn=nn.relu,
dropout=None,
input_layer_partitioner=None,
config=None,
warm_start_from=None,
loss_reduction=losses.Reduction.SUM):
"""A classifier for TensorFlow DNN models with layer annotations.
This classifier is fuctionally identical to estimator.DNNClassifier as far as
training and evaluating models is concerned. The key difference is that this
classifier adds additional layer annotations, which can be used for computing
Integrated Gradients.
Integrated Gradients is a method for attributing a classifier's predictions
to its input features (https://arxiv.org/pdf/1703.01365.pdf). Given an input
instance, the method assigns attribution scores to individual features in
proportion to the feature's importance to the classifier's prediction.
See estimator.DNNClassifer for example code for training and evaluating models
using this classifier.
This classifier is checkpoint-compatible with estimator.DNNClassifier and
therefore the following should work seamlessly:
# Instantiate ordinary estimator as usual.
estimator = tf.estimator.DNNClassifier(
config, feature_columns, hidden_units, ...)
# Train estimator, export checkpoint.
tf.estimator.train_and_evaluate(estimator, ...)
# Instantiate estimator with annotations with the same configuration as the
# ordinary estimator.
estimator_with_annotations = (
tf.contrib.estimator.DNNClassifierWithLayerAnnotations(
config, feature_columns, hidden_units, ...))
# Call export_savedmodel with the same arguments as the ordinary estimator,
# using the checkpoint produced for the ordinary estimator.
estimator_with_annotations.export_saved_model(
export_dir_base, serving_input_receiver, ...
checkpoint_path='/path/to/ordinary/estimator/checkpoint/model.ckpt-1234')
Args:
hidden_units: Iterable of number hidden units per layer. All layers are
fully connected. Ex. `[64, 32]` means first layer has 64 nodes and second
one has 32.
feature_columns: An iterable containing all the feature columns used by the
model. All items in the set should be instances of classes derived from
`_FeatureColumn`.
model_dir: Directory to save model parameters, graph and etc. This can also
be used to load checkpoints from the directory into a estimator to
continue training a previously saved model.
n_classes: Number of label classes. Defaults to 2, namely binary
classification. Must be > 1.
weight_column: A string or a `_NumericColumn` created by
`tf.feature_column.numeric_column` defining feature column representing
weights. It is used to down weight or boost examples during training. It
will be multiplied by the loss of the example. If it is a string, it is
used as a key to fetch weight tensor from the `features`. If it is a
`_NumericColumn`, raw tensor is fetched by key `weight_column.key`, then
weight_column.normalizer_fn is applied on it to get weight tensor.
label_vocabulary: A list of strings represents possible label values. If
given, labels must be string type and have any value in
`label_vocabulary`. If it is not given, that means labels are already
encoded as integer or float within [0, 1] for `n_classes=2` and encoded as
integer values in {0, 1,..., n_classes-1} for `n_classes`>2 . Also there
will be errors if vocabulary is not provided and labels are string.
optimizer: An instance of `tf.Optimizer` used to train the model. Defaults
to Adagrad optimizer.
activation_fn: Activation function applied to each layer. If `None`, will
use `tf.nn.relu`.
dropout: When not `None`, the probability we will drop out a given
coordinate.
input_layer_partitioner: Optional. Partitioner for input layer. Defaults to
`min_max_variable_partitioner` with `min_slice_size` 64 << 20.
config: `RunConfig` object to configure the runtime settings.
warm_start_from: A string filepath to a checkpoint to warm-start from, or a
`WarmStartSettings` object to fully configure warm-starting. If the
string filepath is provided instead of a `WarmStartSettings`, then all
weights are warm-started, and it is assumed that vocabularies and Tensor
names are unchanged.
loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how to
reduce training loss over batch. Defaults to `SUM`.
Returns:
DNNClassifier with layer annotations.
"""
original = dnn.DNNClassifier(
hidden_units=hidden_units,
feature_columns=feature_columns,
model_dir=model_dir,
n_classes=n_classes,
weight_column=weight_column,
label_vocabulary=label_vocabulary,
optimizer=optimizer,
activation_fn=activation_fn,
dropout=dropout,
input_layer_partitioner=input_layer_partitioner,
config=config,
warm_start_from=warm_start_from,
loss_reduction=loss_reduction)
def _model_fn(features, labels, mode, config):
with _monkey_patch(
feature_column_lib, 'input_layer',
make_input_layer_with_layer_annotations(feature_column_lib.input_layer,
mode)):
return original.model_fn(features, labels, mode, config)
return estimator.Estimator(
model_fn=_model_fn,
model_dir=model_dir,
config=config,
warm_start_from=warm_start_from)
def DNNRegressorWithLayerAnnotations( # pylint: disable=invalid-name
hidden_units,
feature_columns,
model_dir=None,
label_dimension=1,
weight_column=None,
optimizer='Adagrad',
activation_fn=nn.relu,
dropout=None,
input_layer_partitioner=None,
config=None,
warm_start_from=None,
loss_reduction=losses.Reduction.SUM,
):
"""A regressor for TensorFlow DNN models with layer annotations.
This regressor is fuctionally identical to estimator.DNNRegressor as far as
training and evaluating models is concerned. The key difference is that this
classifier adds additional layer annotations, which can be used for computing
Integrated Gradients.
Integrated Gradients is a method for attributing a classifier's predictions
to its input features (https://arxiv.org/pdf/1703.01365.pdf). Given an input
instance, the method assigns attribution scores to individual features in
proportion to the feature's importance to the classifier's prediction.
See estimator.DNNRegressor for example code for training and evaluating models
using this regressor.
This regressor is checkpoint-compatible with estimator.DNNRegressor and
therefore the following should work seamlessly:
# Instantiate ordinary estimator as usual.
estimator = tf.estimator.DNNRegressor(
config, feature_columns, hidden_units, ...)
# Train estimator, export checkpoint.
tf.estimator.train_and_evaluate(estimator, ...)
# Instantiate estimator with annotations with the same configuration as the
# ordinary estimator.
estimator_with_annotations = (
tf.contrib.estimator.DNNRegressorWithLayerAnnotations(
config, feature_columns, hidden_units, ...))
# Call export_savedmodel with the same arguments as the ordinary estimator,
# using the checkpoint produced for the ordinary estimator.
estimator_with_annotations.export_saved_model(
export_dir_base, serving_input_receiver, ...
checkpoint_path='/path/to/ordinary/estimator/checkpoint/model.ckpt-1234')
Args:
hidden_units: Iterable of number hidden units per layer. All layers are
fully connected. Ex. `[64, 32]` means first layer has 64 nodes and second
one has 32.
feature_columns: An iterable containing all the feature columns used by the
model. All items in the set should be instances of classes derived from
`_FeatureColumn`.
model_dir: Directory to save model parameters, graph and etc. This can also
be used to load checkpoints from the directory into a estimator to
continue training a previously saved model.
label_dimension: Number of regression targets per example. This is the size
of the last dimension of the labels and logits `Tensor` objects
(typically, these have shape `[batch_size, label_dimension]`).
weight_column: A string or a `_NumericColumn` created by
`tf.feature_column.numeric_column` defining feature column representing
weights. It is used to down weight or boost examples during training. It
will be multiplied by the loss of the example. If it is a string, it is
used as a key to fetch weight tensor from the `features`. If it is a
`_NumericColumn`, raw tensor is fetched by key `weight_column.key`, then
weight_column.normalizer_fn is applied on it to get weight tensor.
optimizer: An instance of `tf.Optimizer` used to train the model. Defaults
to Adagrad optimizer.
activation_fn: Activation function applied to each layer. If `None`, will
use `tf.nn.relu`.
dropout: When not `None`, the probability we will drop out a given
coordinate.
input_layer_partitioner: Optional. Partitioner for input layer. Defaults to
`min_max_variable_partitioner` with `min_slice_size` 64 << 20.
config: `RunConfig` object to configure the runtime settings.
warm_start_from: A string filepath to a checkpoint to warm-start from, or a
`WarmStartSettings` object to fully configure warm-starting. If the
string filepath is provided instead of a `WarmStartSettings`, then all
weights are warm-started, and it is assumed that vocabularies and Tensor
names are unchanged.
loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how to
reduce training loss over batch. Defaults to `SUM`.
Returns:
DNNRegressor with layer annotations.
"""
original = dnn.DNNRegressor(
hidden_units=hidden_units,
feature_columns=feature_columns,
model_dir=model_dir,
label_dimension=label_dimension,
weight_column=weight_column,
optimizer=optimizer,
activation_fn=activation_fn,
dropout=dropout,
input_layer_partitioner=input_layer_partitioner,
config=config,
warm_start_from=warm_start_from,
loss_reduction=loss_reduction,
)
def _model_fn(features, labels, mode, config):
with _monkey_patch(
feature_column_lib, 'input_layer',
make_input_layer_with_layer_annotations(feature_column_lib.input_layer,
mode)):
return original.model_fn(features, labels, mode, config)
return estimator.Estimator(
model_fn=_model_fn,
model_dir=model_dir,
config=config,
warm_start_from=warm_start_from)
|