-
Notifications
You must be signed in to change notification settings - Fork 190
/
configs.py
438 lines (377 loc) · 18.1 KB
/
configs.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
# Copyright 2019 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.
"""Classes for configuring modules in Neural Structured Learning (NSL)."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import enum
import attr
import tensorflow as tf
class NormType(enum.Enum):
"""Types of norms."""
L1 = 'l1'
L2 = 'l2'
INFINITY = 'infinity'
@classmethod
def all(cls):
return list(cls)
@attr.s
class AdvNeighborConfig(object):
"""Contains configuration for generating adversarial neighbors.
Attributes:
feature_mask: mask w/ values in `[0, 1]` applied on the gradient. Its shape
should be the same as (or broadcastable to) that of the input features.
If the input features are in a collection (e.g. list or dictionary), this
field should also be a collection of the same structure. Input features
corresponding to mask values of 0.0 are *not* perturbed. Setting this
field to `None` is equivalent to setting a mask value of 1.0 for all input
features.
adv_step_size: step size to find the adversarial sample. Default set to
0.001.
adv_grad_norm: type of tensor norm to normalize the gradient. Input will be
converted to `nsl.configs.NormType` when applicable (e.g., `'l2'` ->
`nls.configs.NormType.L2`). Default set to L2 norm.
clip_value_min: minimum value to clip the features after perturbation. The
shape should be the same as (or broadcastable to) input features. If the
input features are in a collection (e.g. list or dictionary), this field
should also be a collection with the same structure. An omitted or
`None`-valued entry in the collection indicates no constraint on the
corresponding feature.
clip_value_max: maximum value to clip the feature after perturbation. (See
`clip_value_min` for the structure and shape limitations.)
pgd_iterations: number of attack iterations for Projected Gradient Descent
(PGD) attack. Defaults to 1, which resembles the Fast Gradient Sign Method
(FGSM) attack.
pgd_epsilon: radius of the epsilon ball to project back to. Only used in
Projected Gradient Descent (PGD) attack.
random_init: Apply a random perturbation before FGSM/PGD steps. Default set
to `False` for no random initialization being applied.
"""
feature_mask = attr.ib(default=None)
adv_step_size = attr.ib(default=0.001)
adv_grad_norm = attr.ib(converter=NormType, default='l2')
clip_value_min = attr.ib(default=None)
clip_value_max = attr.ib(default=None)
pgd_iterations = attr.ib(default=1) # 1 is the FGSM attack.
pgd_epsilon = attr.ib(default=None)
random_init = attr.ib(default=False)
@attr.s
class AdvRegConfig(object):
"""Contains configuration for adversarial regularization.
Attributes:
multiplier: multiplier to adversarial regularization loss. Default set to
0.2.
adv_neighbor_config: an `nsl.configs.AdvNeighborConfig` object for
generating adversarial neighbor examples.
"""
multiplier = attr.ib(default=0.2)
adv_neighbor_config = attr.ib(default=AdvNeighborConfig())
def make_adv_reg_config(
multiplier=attr.fields(AdvRegConfig).multiplier.default,
feature_mask=attr.fields(AdvNeighborConfig).feature_mask.default,
adv_step_size=attr.fields(AdvNeighborConfig).adv_step_size.default,
adv_grad_norm=attr.fields(AdvNeighborConfig).adv_grad_norm.default,
clip_value_min=attr.fields(AdvNeighborConfig).clip_value_min.default,
clip_value_max=attr.fields(AdvNeighborConfig).clip_value_max.default,
pgd_iterations=attr.fields(AdvNeighborConfig).pgd_iterations.default,
pgd_epsilon=attr.fields(AdvNeighborConfig).pgd_epsilon.default,
random_init=attr.fields(AdvNeighborConfig).random_init.default):
"""Creates an `nsl.configs.AdvRegConfig` object.
Args:
multiplier: multiplier to adversarial regularization loss. Defaults to 0.2.
feature_mask: mask w/ values in `[0, 1]` applied on the gradient. Its shape
should be the same as (or broadcastable to) that of the input features.
If the input features are in a collection (e.g. list or dictionary), this
field should also be a collection of the same structure. Input features
corresponding to mask values of 0.0 are *not* perturbed. Setting this
field to `None` is equivalent to setting a mask value of 1.0 for all input
features.
adv_step_size: step size to find the adversarial sample. Defaults to 0.001.
adv_grad_norm: type of tensor norm to normalize the gradient. Input will be
converted to `NormType` when applicable (e.g., a value of 'l2' will be
converted to `nsl.configs.NormType.L2`). Defaults to L2 norm.
clip_value_min: minimum value to clip the features after perturbation. The
shape should be the same as (or broadcastable to) input features. If the
input features are in a collection (e.g. list or dictionary), this field
should also be a collection with the same structure. An omitted or
`None`-valued entry in the collection indicates no constraint on the
corresponding feature.
clip_value_max: maximum value to clip the feature after perturbation. (See
`clip_value_min` for the structure and shape limitations.)
pgd_iterations: number of attack iterations for Projected Gradient Descent
(PGD) attack. Defaults to 1, which resembles the Fast Gradient Sign Method
(FGSM) attack.
pgd_epsilon: radius of the epsilon ball to project back to. Only used in
Projected Gradient Descent (PGD) attack.
random_init: Apply a random perturbation before FGSM/PGD steps. Default set
to `False` for no random initialization being applied.
Returns:
An `nsl.configs.AdvRegConfig` object.
"""
return AdvRegConfig(
multiplier=multiplier,
adv_neighbor_config=AdvNeighborConfig(
feature_mask=feature_mask,
adv_step_size=adv_step_size,
adv_grad_norm=adv_grad_norm,
clip_value_min=clip_value_min,
clip_value_max=clip_value_max,
pgd_iterations=pgd_iterations,
pgd_epsilon=pgd_epsilon,
random_init=random_init))
class AdvTargetType(enum.Enum):
"""Types of adversarial targeting."""
SECOND = 'second'
LEAST = 'least'
RANDOM = 'random'
GROUND_TRUTH = 'ground_truth'
@classmethod
def all(cls):
return list(cls)
@attr.s
class AdvTargetConfig(object):
"""Contains configuration for selecting targets to be attacked.
Attributes:
target_method: type of adversarial targeting method. The value needs to be
one of the enums from `nsl.configs.AdvTargetType` (e.g.,
`nsl.configs.AdvTargetType.LEAST`).
random_seed: a Python integer as seed in 'random_uniform' op.
"""
target_method = attr.ib(default=AdvTargetType.GROUND_TRUTH)
random_seed = attr.ib(default=0.0)
class TransformType(enum.Enum):
"""Types of nonlinear functions to be applied ."""
SOFTMAX = 'softmax'
NONE = 'none'
class DistanceType(enum.Enum):
"""Types of distance."""
L1 = 'l1'
L2 = 'l2'
COSINE = 'cosine'
JENSEN_SHANNON_DIVERGENCE = 'jensen_shannon_divergence'
KL_DIVERGENCE = 'kl_divergence'
@classmethod
def all(cls):
return list(cls)
@attr.s
class DistanceConfig(object):
"""Contains configuration for computing distances between tensors.
Attributes:
distance_type: type of distance function. Input type will be converted to
the appropriate `nsl.configs.DistanceType` value (e.g., the value 'l2' is
converted to `nsl.configs.DistanceType.L2`). Defaults to the L2 norm.
reduction: type of distance reduction. See `tf.compat.v1.losses.Reduction`
for details. Defaults to `tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS`.
sum_over_axis: the distance is the sum over the difference along the axis.
See `nsl.lib.pairwise_distance_wrapper` for how this field is used.
Defaults to `None`.
transform_fn: type of transform function to be applied on each side before
computing the pairwise distance. Input type will be converted to
`nsl.configs.TransformType` when applicable (e.g., the value 'softmax'
maps to `nsl.configs.TransformType.SOFTMAX`). Defaults to 'none'.
"""
distance_type = attr.ib(converter=DistanceType, default=DistanceType.L2)
reduction = attr.ib(
default=tf.compat.v1.losses.Reduction.SUM_BY_NONZERO_WEIGHTS)
sum_over_axis = attr.ib(default=None)
transform_fn = attr.ib(converter=TransformType, default='none')
class DecayType(enum.Enum):
"""Types of decay."""
EXPONENTIAL_DECAY = 'exponential_decay'
INVERSE_TIME_DECAY = 'inverse_time_decay'
NATURAL_EXP_DECAY = 'natural_exp_decay'
@classmethod
def all(cls):
return list(cls)
@attr.s
class DecayConfig(object):
"""Contains configuration for decaying a value during training.
Attributes:
decay_steps: A scalar `int32` or `int64` Tensor or a Python number that
specifies the decay frequency, specied in units of training steps. Must be
positive.
decay_rate: A scalar `float32` or `float64` Tensor or a Python number.
Defaults to 0.96.
min_value: minimal acceptable value after applying decay. Defaults to 0.0.
decay_type: Type of decay function to apply. Defaults to
`nsl.configs.DecayType.EXPONENTIAL_DECAY`.
"""
decay_steps = attr.ib()
decay_rate = attr.ib(default=0.96)
min_value = attr.ib(default=0.0)
decay_type = attr.ib(default=DecayType.EXPONENTIAL_DECAY)
class IntegrationType(enum.Enum):
"""Types of integration for multimodal fusion."""
ADD = 'additive'
MUL = 'multiplicative'
TUCKER_DECOMP = 'tucker_decomp'
@classmethod
def all(cls):
return list(cls)
@attr.s
class IntegrationConfig(object):
"""Contains configuration for computing multimodal integration.
Attributes:
integration_type: Type of integration function to apply.
hidden_dims: Integer or a list of Integer, the number of hidden units in the
fully-connected layer(s) before the output layer.
activation_fn: Activation function to be applied to.
"""
integration_type = attr.ib(converter=IntegrationType)
hidden_dims = attr.ib()
activation_fn = attr.ib(default=tf.nn.tanh)
@attr.s
class VirtualAdvConfig(object):
"""Contains configuration for virtual adversarial training.
Attributes:
adv_neighbor_config: an `nsl.configs.AdvNeighborConfig` object for
generating virtual adversarial examples. Defaults to
`nsl.configs.AdvNeighborConfig()`.
distance_config: a `nsl.configs.DistanceConfig` object for calculating
virtual adversarial loss. Defaults to `nsl.configs.DistanceConfig()`.
num_approx_steps: number of steps used to approximate the calculation of
Hessian matrix required for creating virtual adversarial examples.
Defaults to 1.
approx_difference: the finite difference to approximate the calculation of
the Hessian matrix required for creating virtual adversarial examples,
namely, the `xi` in Equation 12 in the paper:
https://arxiv.org/pdf/1704.03976.pdf. Defaults to 1e-6.
"""
adv_neighbor_config = attr.ib(default=AdvNeighborConfig())
distance_config = attr.ib(default=DistanceConfig())
num_approx_steps = attr.ib(default=1)
approx_difference = attr.ib(default=1e-6)
@attr.s
class GraphBuilderConfig(object):
"""Encapsulates configuration parameters for building a graph.
For more information, see `nsl.tools.build_graph_from_config`.
Attributes:
id_feature_name: The name of the feature in the input `tf.train.Example`
objects representing the ID of examples. Defaults to `'id'`.
embedding_feature_name: The name of the feature in the input
`tf.train.Example` objects representing the embedding of examples.
Defaults to `'embedding'`.
similarity_threshold: Threshold used to determine which edges to retain in
the resulting graph. Defaults to 0.8.
lsh_splits: On each LSH bucketing round, the space containing the input
instances will be randomly split/partitioned this many times for better
graph builder performance. See the `nsl.tools.build_graph_from_config`
documentation for details. Defaults to 0, in which case all pairs of
inputs will be compared, probably resulting in slow running times on
larger input sets.
lsh_rounds: The number of rounds of LSH bucketing to perform when
`lsh_splits > 0`. This is also the number of LSH buckets each point will
be hashed into. Defaults to 1.
random_seed: Value used to seed the random number generator used to perform
randomized LSH bucketing of the inputs when `lsh_splits > 0`. By default,
the generator will be initialized randomly, but setting this to any
integer will initialize it deterministically. Defaults to `None`.
"""
id_feature_name = attr.ib(default='id')
embedding_feature_name = attr.ib(default='embedding')
similarity_threshold = attr.ib(default=0.8)
lsh_splits = attr.ib(default=0)
lsh_rounds = attr.ib(default=1)
random_seed = attr.ib(default=None)
@attr.s
class GraphNeighborConfig(object):
"""Specifies neighbor attributes for graph regularization.
Attributes:
prefix: The prefix in feature names that identifies neighbor-specific
features. Defaults to 'NL_nbr_'.
weight_suffix: The suffix in feature names that identifies the neighbor
weight value. Defaults to '_weight'. Note that neighbor weight features
will have `prefix` as a prefix and `weight_suffix` as a suffix. For
example, based on the default values of `prefix` and `weight_suffix`, a
valid neighbor weight feature is 'NL_nbr_0_weight', where 0 corresponds to
the first neighbor of the sample.
max_neighbors: The maximum number of neighbors to be used for graph
regularization. Defaults to 0, which disables graph regularization. Note
that this value has to be less than or equal to the actual number of
neighbors in each sample.
"""
prefix = attr.ib(default='NL_nbr_')
weight_suffix = attr.ib(default='_weight')
max_neighbors = attr.ib(default=0)
@attr.s
class GraphRegConfig(object):
"""Contains the configuration for graph regularization.
Attributes:
neighbor_config: A `nsl.configs.GraphNeighborConfig` instance that describes
neighbor attributes for graph regularization. Defaults to
`nsl.configs.GraphNeighborConfig()`.
multiplier: The multiplier or weight factor applied on the graph
regularization loss term. This value has to be non-negative. Defaults to
0.01.
distance_config: An instance of `DistanceConfig` to calculate the graph
regularization loss term. Defaults to `nsl.configs.DistanceConfig()`.
"""
neighbor_config = attr.ib(default=GraphNeighborConfig())
multiplier = attr.ib(default=0.01)
distance_config = attr.ib(default=DistanceConfig())
def make_graph_reg_config(
neighbor_prefix=attr.fields(GraphNeighborConfig).prefix.default,
neighbor_weight_suffix=attr.fields(
GraphNeighborConfig).weight_suffix.default,
max_neighbors=attr.fields(GraphNeighborConfig).max_neighbors.default,
multiplier=attr.fields(GraphRegConfig).multiplier.default,
distance_type=attr.fields(DistanceConfig).distance_type.default,
reduction=attr.fields(DistanceConfig).reduction.default,
sum_over_axis=attr.fields(DistanceConfig).sum_over_axis.default,
transform_fn=attr.fields(DistanceConfig).transform_fn.default):
"""Creates an `nsl.configs.GraphRegConfig` object.
Args:
neighbor_prefix: The prefix in feature names that identifies
neighbor-specific features. Defaults to 'NL_nbr_'.
neighbor_weight_suffix: The suffix in feature names that identifies the
neighbor weight value. Defaults to '_weight'. Note that neighbor weight
features will have `prefix` as a prefix and `weight_suffix` as a suffix.
For example, based on the default values of `prefix` and `weight_suffix`,
a valid neighbor weight feature is 'NL_nbr_0_weight', where 0 corresponds
to the first neighbor of the sample.
max_neighbors: The maximum number of neighbors to be used for graph
regularization. Defaults to 0, which disables graph regularization. Note
that this value has to be less than or equal to the actual number of
neighbors in each sample.
multiplier: The multiplier or weight factor applied on the graph
regularization loss term. This value has to be non-negative. Defaults to
0.01.
distance_type: type of distance function. Input type will be converted to
the appropriate `nsl.configs.DistanceType` value (e.g., the value 'l2' is
converted to `nsl.configs.DistanceType.L2`). Defaults to the L2 norm.
reduction: type of distance reduction. See `tf.compat.v1.losses.Reduction`
for details. Defaults to `tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS`.
sum_over_axis: the distance is the sum over the difference along the axis.
See `nsl.lib.pairwise_distance_wrapper` for how this field is used.
Defaults to `None`.
transform_fn: type of transform function to be applied on each side before
computing the pairwise distance. Input type will be converted to
`nsl.configs.TransformType` when applicable (e.g., the value 'softmax'
maps to `nsl.configs.TransformType.SOFTMAX`). Defaults to 'none'.
Returns:
An `nsl.configs.GraphRegConfig` object.
"""
return GraphRegConfig(
neighbor_config=GraphNeighborConfig(
prefix=neighbor_prefix,
weight_suffix=neighbor_weight_suffix,
max_neighbors=max_neighbors),
multiplier=multiplier,
distance_config=DistanceConfig(
distance_type=distance_type,
reduction=reduction,
sum_over_axis=sum_over_axis,
transform_fn=transform_fn))
DEFAULT_DISTANCE_PARAMS = attr.asdict(DistanceConfig())
DEFAULT_ADVERSARIAL_PARAMS = attr.asdict(AdvNeighborConfig())