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
|
# Copyright 2015 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 SoftmaxCrossEntropyWithLogits op."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow as tf
from tensorflow.python.ops import gen_nn_ops
class XentTest(tf.test.TestCase):
def _npXent(self, features, labels):
batch_dim = 0
class_dim = 1
batch_size = features.shape[batch_dim]
e = np.exp(features -
np.reshape(np.amax(features, axis=class_dim), [batch_size, 1]))
probs = e / np.reshape(np.sum(e, axis=class_dim), [batch_size, 1])
bp = (probs - labels)
l = -np.sum(labels * np.log(probs + 1.0e-20), axis=1)
return l, bp
def _testXent(self, np_features, np_labels, use_gpu=False):
np_loss, np_backprop = self._npXent(np_features, np_labels)
with self.test_session(use_gpu=use_gpu) as sess:
loss, backprop = gen_nn_ops._softmax_cross_entropy_with_logits(
np_features, np_labels)
tf_loss, tf_backprop = sess.run([loss, backprop])
self.assertAllCloseAccordingToType(np_loss, tf_loss)
self.assertAllCloseAccordingToType(np_backprop, tf_backprop)
def _testAll(self, features, labels):
self._testXent(features, labels, use_gpu=False)
self._testXent(features, labels, use_gpu=True)
def _testSingleClass(self, use_gpu=False):
for dtype in np.float16, np.float32:
with self.test_session(use_gpu=use_gpu) as sess:
loss, backprop = gen_nn_ops._softmax_cross_entropy_with_logits(
np.array([[1.], [-1.], [0.]]).astype(dtype),
np.array([[-1.], [0.], [1.]]).astype(dtype))
tf_loss, tf_backprop = sess.run([loss, backprop])
self.assertAllClose([0.0, 0.0, 0.0], tf_loss)
self.assertAllClose([[2.0], [1.0], [0.0]], tf_backprop)
def testSingleClass(self):
self._testSingleClass(True)
self._testSingleClass(False)
def testRankTooLarge(self):
for dtype in np.float16, np.float32:
np_features = np.array(
[[[1., 1., 1., 1.]], [[1., 2., 3., 4.]]]).astype(dtype)
np_labels = np.array(
[[[0., 0., 0., 1.]], [[0., .5, .5, 0.]]]).astype(dtype)
self.assertRaisesRegexp(ValueError, "must have rank 2",
gen_nn_ops._softmax_cross_entropy_with_logits,
np_features, np_labels)
def testNpXent(self):
# We create 2 batches of logits for testing.
# batch 0 is the boring uniform distribution: 1, 1, 1, 1, with target 3.
# batch 1 has a bit of difference: 1, 2, 3, 4, with soft targets (1, 2).
features = [[1., 1., 1., 1.], [1., 2., 3., 4.]]
labels = [[0., 0., 0., 1.], [0., .5, .5, 0.]]
# For batch 0, we expect the uniform distribution: 0.25, 0.25, 0.25, 0.25
# With a hard target 3, the backprop is [0.25, 0.25, 0.25, -0.75]
# The loss for this batch is -log(0.25) = 1.386
#
# For batch 1, we have:
# exp(0) = 1
# exp(1) = 2.718
# exp(2) = 7.389
# exp(3) = 20.085
# SUM = 31.192
# So we have as probabilities:
# exp(0) / SUM = 0.032
# exp(1) / SUM = 0.087
# exp(2) / SUM = 0.237
# exp(3) / SUM = 0.644
# With a soft target (1, 2), the backprop is
# [0.032, 0.087 - 0.5 = -0.413, 0.237 - 0.5 = -0.263, 0.644]
# The loss for this batch is [0.5 * -log(0.087), 0.5 * -log(0.237)]
# = [1.3862, 1.9401]
np_loss, np_backprop = self._npXent(np.array(features), np.array(labels))
self.assertAllClose(np.array([[0.25, 0.25, 0.25, -0.75],
[0.0321, -0.4129, -0.2632, 0.6439]]),
np_backprop,
rtol=1.e-3, atol=1.e-3)
self.assertAllClose(np.array([1.3862, 1.9401]), np_loss,
rtol=1.e-3, atol=1.e-3)
def testShapeMismatch(self):
with self.test_session():
with self.assertRaises(ValueError):
gen_nn_ops._softmax_cross_entropy_with_logits(
[[0., 1.], [2., 3.]], [[0., 1., 0.], [1., 0., 0.]])
def testNotMatrix(self):
with self.test_session():
with self.assertRaises(ValueError):
gen_nn_ops._softmax_cross_entropy_with_logits([0., 1., 2., 3.],
[0., 1., 0., 1.])
def testHalf(self):
self._testAll(
np.array([[1., 1., 1., 1.], [1., 2., 3., 4.]]).astype(np.float16),
np.array([[0., 0., 0., 1.], [0., .5, .5, 0.]]).astype(np.float16))
def testFloat(self):
self._testAll(
np.array([[1., 1., 1., 1.], [1., 2., 3., 4.]]).astype(np.float32),
np.array([[0., 0., 0., 1.], [0., .5, .5, 0.]]).astype(np.float32))
def testDouble(self):
self._testAll(
np.array([[1., 1., 1., 1.], [1., 2., 3., 4.]]).astype(np.float64),
np.array([[0., 0., 0., 1.], [0., .5, .5, 0.]]).astype(np.float64))
def testGradient(self):
with self.test_session():
l = tf.constant([0.0, 0.0, 1.0, 0.0,
1.0, 0.0, 0.0, 0.0,
0.0, 0.5, 0.0, 0.5], shape=[3, 4],
dtype=tf.float64, name="l")
f = tf.constant([0.1, 0.2, 0.3, 0.4,
0.1, 0.4, 0.9, 1.6,
0.1, 0.8, 2.7, 6.4], shape=[3, 4],
dtype=tf.float64, name="f")
x = tf.nn.softmax_cross_entropy_with_logits(f, l, name="xent")
err = tf.test.compute_gradient_error(f, [3, 4], x, [3])
print("cross entropy gradient err = ", err)
self.assertLess(err, 5e-8)
if __name__ == "__main__":
tf.test.main()
|
