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
|
# 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.
# ==============================================================================
"""Tests for atrous convolution functionality in tensorflow.ops.nn."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow as tf
def upsample_filters(filters, rate):
"""Upsamples the filters by a factor of rate along the spatial dimensions.
Args:
filters: spatial_shape + [in_channels, out_channels]
Original filters.
rate: A list of len(spatial_shape) positive ints, specifying the
upsampling rate.
Returns:
filters_up: output_spatial_shape + [in_channels, out_channels].
Upsampled filters with
output_spatial_shape[i] = (spatial_shape[i] - 1) * rate[i] + 1
containing (rate[i] - 1) zeros between consecutive filter values along
spatial dimension i.
"""
num_spatial_dims = len(rate)
spatial_shape = np.array(filters.shape[:num_spatial_dims])
output_spatial_shape = (spatial_shape - 1) * rate + 1
output = np.zeros(tuple(output_spatial_shape) + tuple(filters.shape[-2:]),
filters.dtype)
output[tuple(np.s_[::rate[i]] for i in range(num_spatial_dims))] = filters
return output
class AtrousConvolutionTest(tf.test.TestCase):
def _test_atrous_convolution(self, input_shape, filter_shape, dilation_rate,
**kwargs):
filters = np.arange(
np.prod(filter_shape), dtype=np.float32).reshape(filter_shape)
filters_upsampled = upsample_filters(filters, dilation_rate)
x = np.arange(np.prod(input_shape), dtype=np.float32).reshape(input_shape)
y1 = tf.nn.convolution(
input=x, filter=filters, dilation_rate=dilation_rate, **kwargs)
y2 = tf.nn.convolution(input=x, filter=filters_upsampled, **kwargs)
self.assertAllClose(y1.eval(), y2.eval(), rtol=1e-2, atol=1e-2)
def testAtrousConvolution2D(self):
with self.test_session():
for padding in ["SAME", "VALID"]:
for height, width in [[9, 9], [9, 10]]:
for kernel_height, kernel_width in [[1, 1], [2, 2], [2, 3]]:
for dilation_rate in [[1, 1], [3, 2], [2, 1]]:
self._test_atrous_convolution(
input_shape=[2, height, width, 2],
filter_shape=[kernel_height, kernel_width, 2, 2],
padding=padding,
dilation_rate=dilation_rate)
def testAtrousConvolution3D(self):
with self.test_session():
for padding in ["SAME", "VALID"]:
for depth, height, width in [[9, 9, 10], [9, 10, 9]]:
for kernel_depth, kernel_height, kernel_width in [[3, 3, 3],
[3, 2, 2],
[2, 1, 3]]:
for dilation_rate in [[1, 1, 1], [3, 3, 3], [3, 2, 3],
[3, 1, 2]]:
self._test_atrous_convolution(
input_shape=[2, depth, height, width, 2],
filter_shape=[kernel_depth, kernel_height, kernel_width,
2, 2],
padding=padding,
dilation_rate=dilation_rate)
def testAtrousConvolution1D(self):
with self.test_session():
for padding in ["SAME", "VALID"]:
for width in [9, 10]:
for kernel_width in range(1, 4):
for rate in range(1, 4):
self._test_atrous_convolution(
input_shape=[2, width, 2],
filter_shape=[kernel_width, 2, 2],
padding=padding,
dilation_rate=[rate])
def testAtrousConvolutionNC(self):
if tf.test.is_gpu_available():
# "NCW" and "NCHW" formats are not currently supported on CPU.
with self.test_session(use_gpu=True):
for padding in ["SAME", "VALID"]:
self._test_atrous_convolution(
input_shape=[2, 2, 9],
padding=padding,
filter_shape=[3, 2, 2],
dilation_rate=[2],
data_format="NCW")
self._test_atrous_convolution(
input_shape=[2, 2, 9, 5],
padding=padding,
filter_shape=[3, 3, 2, 2],
dilation_rate=[2, 1],
data_format="NCHW")
def testAtrousSequence(self):
"""Tests optimization of sequence of atrous convolutions.
See the documentation of with_space_to_batch.
"""
with self.test_session():
for padding in ["SAME", "VALID"]:
for height in range(15, 17):
for width in range(15, 17):
x_shape = [3, height, width, 2]
x = np.random.random_sample(x_shape).astype(np.float32)
kernel_sizes = [1, 3] if padding == "SAME" else range(1, 3)
for kernel in kernel_sizes:
f_shape = [kernel, kernel, 2, 2]
f1 = 1e-2 * np.random.random_sample(f_shape).astype(np.float32)
f2 = 1e-2 * np.random.random_sample(f_shape).astype(np.float32)
def combined_op(converted_input, num_spatial_dims, padding_arg): # pylint: disable=unused-argument
result = tf.nn.convolution(
input=converted_input, filter=f1, padding=padding) # pylint: disable=cell-var-from-loop
result = tf.nn.convolution(
input=result, filter=f2, padding=padding) # pylint: disable=cell-var-from-loop
return result
for rate_height in range(2, 4):
for rate_width in range(2, 4):
dilation_rate = [rate_height, rate_width]
y1 = tf.nn.convolution(
input=x,
filter=f1,
padding=padding,
dilation_rate=dilation_rate)
y1 = tf.nn.convolution(
input=y1,
filter=f2,
padding=padding,
dilation_rate=dilation_rate)
y2 = tf.nn.with_space_to_batch(
input=x,
dilation_rate=dilation_rate,
op=combined_op,
padding="VALID")
self.assertAllClose(
y1.eval(), y2.eval(), rtol=1e-2, atol=1e-2)
def _test_gradient(self, x_shape, f_shape, dilation_rate, padding):
x_val = np.random.random_sample(x_shape).astype(np.float32)
f_val = np.random.random_sample(f_shape).astype(np.float32)
x = tf.constant(x_val, name="x", dtype=tf.float32)
f = tf.constant(f_val, name="f", dtype=tf.float32)
output = tf.nn.convolution(
input=x, filter=f, dilation_rate=dilation_rate, padding=padding)
y_shape = output.get_shape().as_list()
err = tf.test.compute_gradient_error([x, f], [x_shape, f_shape], output,
y_shape)
err_tolerance = 1e-3
self.assertLess(err, err_tolerance)
def testGradient(self):
with self.test_session():
for padding in ["SAME", "VALID"]:
for rate_width in range(1, 3):
for rate_height in range(1, 3):
self._test_gradient(
x_shape=[2, 5, 6, 2],
f_shape=[3, 3, 2, 2],
dilation_rate=[rate_height, rate_width],
padding=padding)
if __name__ == "__main__":
tf.test.main()
|
