/
sampling_ops.py
422 lines (362 loc) · 16.8 KB
/
sampling_ops.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
# Copyright 2016 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.
# ==============================================================================
"""Sampling functions."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import check_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import logging_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import random_ops
from tensorflow.python.ops import variable_scope
from tensorflow.python.training import input as input_ops
__all__ = [
'rejection_sample',
'stratified_sample',
]
def rejection_sample(tensors,
accept_prob_fn,
batch_size,
queue_threads=1,
enqueue_many=False,
prebatch_capacity=16,
prebatch_threads=1,
runtime_checks=False,
name=None):
"""Stochastically creates batches by rejection sampling.
Each list of non-batched tensors is evaluated by `accept_prob_fn`, to produce
a scalar tensor between 0 and 1. This tensor corresponds to the probability of
being accepted. When `batch_size` tensor groups have been accepted, the batch
queue will return a mini-batch.
Args:
tensors: List of tensors for data. All tensors are either one item or a
batch, according to enqueue_many.
accept_prob_fn: A python lambda that takes a non-batch tensor from each
item in `tensors`, and produces a scalar tensor.
batch_size: Size of batch to be returned.
queue_threads: The number of threads for the queue that will hold the final
batch.
enqueue_many: Bool. If true, interpret input tensors as having a batch
dimension.
prebatch_capacity: Capacity for the large queue that is used to convert
batched tensors to single examples.
prebatch_threads: Number of threads for the large queue that is used to
convert batched tensors to single examples.
runtime_checks: Bool. If true, insert runtime checks on the output of
`accept_prob_fn`. Using `True` might have a performance impact.
name: Optional prefix for ops created by this function.
Raises:
ValueError: enqueue_many is True and labels doesn't have a batch
dimension, or if enqueue_many is False and labels isn't a scalar.
ValueError: enqueue_many is True, and batch dimension on data and labels
don't match.
ValueError: if a zero initial probability class has a nonzero target
probability.
Returns:
A list of tensors of the same length as `tensors`, with batch dimension
`batch_size`.
Example:
# Get tensor for a single data and label example.
data, label = data_provider.Get(['data', 'label'])
# Get stratified batch according to data tensor.
accept_prob_fn = lambda x: (tf.tanh(x[0]) + 1) / 2
data_batch = tf.contrib.training.rejection_sample(
[data, label], accept_prob_fn, 16)
# Run batch through network.
...
"""
with variable_scope.variable_scope(name, 'rejection_sample', tensors):
tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensors)
# Reduce the case of a batched example to that of a batch of a single
# example by taking a batch of size one.
if enqueue_many:
# Validate that batch dimension of the input is consistent.
tensor_list = _verify_data_inputs(tensor_list)
# Make a single queue to hold input examples. Reshape output so examples
# don't have singleton batch dimension.
batched = input_ops.batch(
tensor_list,
batch_size=1,
num_threads=prebatch_threads,
capacity=prebatch_capacity,
enqueue_many=True)
tensor_list = [array_ops.squeeze(x, [0]) for x in batched]
# Set up a queue containing batches that have the distribution.
cur_prob = accept_prob_fn(tensor_list)
if runtime_checks:
cur_prob = array_ops.identity(
control_flow_ops.with_dependencies([
check_ops.assert_less_equal(0.0, cur_prob),
check_ops.assert_less_equal(cur_prob, 1.0)
], cur_prob),
name='prob_with_checks')
minibatch = input_ops.maybe_batch(
tensor_list,
keep_input=random_ops.random_uniform([]) < cur_prob,
batch_size=batch_size,
num_threads=queue_threads)
# Queues return a single tensor if the list of enqueued tensors is one. Since
# we want the type to always be the same, always return a list.
if isinstance(minibatch, ops.Tensor):
minibatch = [minibatch]
return minibatch
def stratified_sample(tensors,
labels,
target_probs,
batch_size,
init_probs=None,
enqueue_many=False,
queue_capacity=16,
threads_per_queue=1,
name=None):
"""Stochastically creates batches based on per-class probabilities.
This method discards examples. Internally, it creates one queue to amortize
the cost of disk reads, and one queue to hold the properly-proportioned
batch.
Args:
tensors: List of tensors for data. All tensors are either one item or a
batch, according to enqueue_many.
labels: Tensor for label of data. Label is a single integer or a batch,
depending on `enqueue_many`. It is not a one-hot vector.
target_probs: Target class proportions in batch. An object whose type has a
registered Tensor conversion function.
batch_size: Size of batch to be returned.
init_probs: Class proportions in the data. An object whose type has a
registered Tensor conversion function, or `None` for estimating the
initial distribution.
enqueue_many: Bool. If true, interpret input tensors as having a batch
dimension.
queue_capacity: Capacity of the large queue that holds input examples.
threads_per_queue: Number of threads for the large queue that holds input
examples and for the final queue with the proper class proportions.
name: Optional prefix for ops created by this function.
Raises:
ValueError: If `tensors` isn't iterable.
ValueError: `enqueue_many` is True and labels doesn't have a batch
dimension, or if `enqueue_many` is False and labels isn't a scalar.
ValueError: `enqueue_many` is True, and batch dimension on data and labels
don't match.
ValueError: if probs don't sum to one.
ValueError: if a zero initial probability class has a nonzero target
probability.
TFAssertion: if labels aren't integers in [0, num classes).
Returns:
(data_batch, label_batch), where data_batch is a list of tensors of the same
length as `tensors`
Example:
# Get tensor for a single data and label example.
data, label = data_provider.Get(['data', 'label'])
# Get stratified batch according to per-class probabilities.
target_probs = [...distribution you want...]
[data_batch], labels = tf.contrib.training.stratified_sample(
[data], label, target_probs)
# Run batch through network.
...
"""
with ops.name_scope(name, 'stratified_sample', list(tensors) + [labels]):
tensor_list = ops.convert_n_to_tensor_or_indexed_slices(tensors)
labels = ops.convert_to_tensor(labels)
target_probs = ops.convert_to_tensor(target_probs, dtype=dtypes.float32)
# Reduce the case of a single example to that of a batch of size 1.
if not enqueue_many:
tensor_list = [array_ops.expand_dims(tensor, 0) for tensor in tensor_list]
labels = array_ops.expand_dims(labels, 0)
# If `init_probs` is `None`, set up online estimation of data distribution.
if init_probs is None:
# We use `target_probs` to get the number of classes, so its shape must be
# fully defined at graph construction time.
target_probs.get_shape().assert_is_fully_defined()
init_probs = _estimate_data_distribution(
labels, target_probs.get_shape().num_elements())
else:
init_probs = ops.convert_to_tensor(init_probs, dtype=dtypes.float32)
# Validate that input is consistent.
tensor_list, labels, [init_probs, target_probs] = _verify_input(
tensor_list, labels, [init_probs, target_probs])
# Check that all zero initial probabilities also have zero target
# probabilities.
assert_op = control_flow_ops.Assert(
math_ops.reduce_all(
math_ops.logical_or(
math_ops.not_equal(init_probs, 0),
math_ops.equal(target_probs, 0))),
['All classes with zero initial probability must also have zero target '
'probability: ', init_probs, target_probs
])
init_probs = control_flow_ops.with_dependencies([assert_op], init_probs)
# Calculate acceptance sampling probabilities.
accept_probs = _calculate_acceptance_probabilities(init_probs, target_probs)
proportion_rejected = math_ops.reduce_sum((1 - accept_probs) * init_probs)
accept_probs = control_flow_ops.cond(
math_ops.less(proportion_rejected, .5),
lambda: accept_probs,
lambda: logging_ops.Print( # pylint: disable=g-long-lambda
accept_probs, [accept_probs],
message='Proportion of examples rejected by sampler is high.',
first_n=10))
# Make a single queue to hold input examples. Reshape output so examples
# don't have singleton batch dimension.
batched = input_ops.batch(
tensor_list + [labels],
batch_size=1,
num_threads=threads_per_queue,
capacity=queue_capacity,
enqueue_many=True)
val_list = [array_ops.squeeze(x, [0]) for x in batched[:-1]]
label = array_ops.squeeze(batched[-1], [0])
# Set up second queue containing batches that have the desired class
# proportions.
cur_prob = array_ops.gather(accept_probs, label)
batched = input_ops.maybe_batch(
val_list + [label],
keep_input=random_ops.random_uniform([]) < cur_prob,
batch_size=batch_size,
num_threads=threads_per_queue)
return batched[:-1], batched[-1]
def _estimate_data_distribution(labels, num_classes, smoothing_constant=10):
"""Estimate data distribution as labels are seen."""
# Variable to track running count of classes. Smooth by a nonzero value to
# avoid division-by-zero. Higher values provide more stability at the cost of
# slower convergence.
if smoothing_constant <= 0:
raise ValueError('smoothing_constant must be nonzero.')
num_examples_per_class_seen = variable_scope.variable(
initial_value=[smoothing_constant] * num_classes,
trainable=False,
name='class_count',
dtype=dtypes.int64)
# Update the class-count based on what labels are seen in batch.
num_examples_per_class_seen = num_examples_per_class_seen.assign_add(
math_ops.reduce_sum(
array_ops.one_hot(
labels, num_classes, dtype=dtypes.int64), 0))
# Normalize count into a probability.
# NOTE: Without the `+= 0` line below, the test
# `testMultiThreadedEstimateDataDistribution` fails. The reason is that
# before this line, `num_examples_per_class_seen` is a Tensor that shares a
# buffer with an underlying `ref` object. When the `ref` is changed by another
# thread, `num_examples_per_class_seen` changes as well. Since this can happen
# in the middle of the normalization computation, we get probabilities that
# are very far from summing to one. Adding `+= 0` copies the contents of the
# tensor to a new buffer, which will be consistent from the start to the end
# of the normalization computation.
num_examples_per_class_seen += 0
init_prob_estimate = math_ops.truediv(
num_examples_per_class_seen,
math_ops.reduce_sum(num_examples_per_class_seen))
# Must return float32 (not float64) to agree with downstream `_verify_input`
# checks.
return math_ops.cast(init_prob_estimate, dtypes.float32)
def _verify_data_inputs(tensor_list):
"""Verify that batched data inputs are well-formed."""
for tensor in tensor_list:
# Data tensor should have a batch dimension.
tensor_shape = tensor.get_shape().with_rank_at_least(1)
# Data batch dimensions must be compatible.
tensor_shape[0].assert_is_compatible_with(tensor_list[0].get_shape()[0])
return tensor_list
def _verify_input(tensor_list, labels, probs_list):
"""Verify that batched inputs are well-formed."""
checked_probs_list = []
for probs in probs_list:
# Since number of classes shouldn't change at runtime, probabilities shape
# should be fully defined.
probs.get_shape().assert_is_fully_defined()
# Probabilities must be 1D.
probs.get_shape().assert_has_rank(1)
# Probabilities must be nonnegative and sum to one.
tol = 1e-6
prob_sum = math_ops.reduce_sum(probs)
checked_probs = control_flow_ops.with_dependencies([
check_ops.assert_non_negative(probs),
check_ops.assert_less(prob_sum, 1.0 + tol),
check_ops.assert_less(1.0 - tol, prob_sum)
], probs)
checked_probs_list.append(checked_probs)
# All probabilities should be the same length.
prob_length = checked_probs_list[0].get_shape().num_elements()
for checked_prob in checked_probs_list:
if checked_prob.get_shape().num_elements() != prob_length:
raise ValueError('Probability parameters must have the same length.')
# Labels tensor should only have batch dimension.
labels.get_shape().assert_has_rank(1)
for tensor in tensor_list:
# Data tensor should have a batch dimension.
tensor_shape = tensor.get_shape().with_rank_at_least(1)
# Data and label batch dimensions must be compatible.
tensor_shape[0].assert_is_compatible_with(labels.get_shape()[0])
# Data and labels must have the same, strictly positive batch size. Since we
# can't assume we know the batch size at graph creation, add runtime checks.
labels_batch_size = array_ops.shape(labels)[0]
lbl_assert = check_ops.assert_positive(labels_batch_size)
# Make each tensor depend on its own checks.
labels = control_flow_ops.with_dependencies([lbl_assert], labels)
tensor_list = [
control_flow_ops.with_dependencies([
lbl_assert,
check_ops.assert_equal(array_ops.shape(x)[0], labels_batch_size)
], x) for x in tensor_list
]
# Label's classes must be integers 0 <= x < num_classes.
labels = control_flow_ops.with_dependencies([
check_ops.assert_integer(labels), check_ops.assert_non_negative(labels),
check_ops.assert_less(labels, math_ops.cast(prob_length, labels.dtype))
], labels)
return tensor_list, labels, checked_probs_list
def _calculate_acceptance_probabilities(init_probs, target_probs):
"""Calculate the per-class acceptance rates.
Args:
init_probs: The class probabilities of the data.
target_probs: The desired class proportion in minibatches.
Returns:
A list of the per-class acceptance probabilities.
This method is based on solving the following analysis:
Let F be the probability of a rejection (on any example).
Let p_i be the proportion of examples in the data in class i (init_probs)
Let a_i is the rate the rejection sampler should *accept* class i
Let t_i is the target proportion in the minibatches for class i (target_probs)
```
F = sum_i(p_i * (1-a_i))
= 1 - sum_i(p_i * a_i) using sum_i(p_i) = 1
```
An example with class `i` will be accepted if `k` rejections occur, then an
example with class `i` is seen by the rejector, and it is accepted. This can
be written as follows:
```
t_i = sum_k=0^inf(F^k * p_i * a_i)
= p_i * a_j / (1 - F) using geometric series identity, since 0 <= F < 1
= p_i * a_i / sum_j(p_j * a_j) using F from above
```
Note that the following constraints hold:
```
0 <= p_i <= 1, sum_i(p_i) = 1
0 <= a_i <= 1
0 <= t_i <= 1, sum_i(t_i) = 1
```
A solution for a_i in terms of the other variables is the following:
```a_i = (t_i / p_i) / max_i[t_i / p_i]```
"""
# Make list of t_i / p_i.
ratio_l = target_probs / init_probs
# Replace NaNs with 0s.
ratio_l = array_ops.where(
math_ops.is_nan(ratio_l), array_ops.zeros_like(ratio_l), ratio_l)
# Calculate list of acceptance probabilities.
max_ratio = math_ops.reduce_max(ratio_l)
return ratio_l / max_ratio