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
|
"""Tests for tensorflow.ops.math_ops.matmul."""
import tensorflow.python.platform
import numpy as np
import tensorflow as tf
from tensorflow.python.kernel_tests import gradient_checker as gc
class MatMulTest(tf.test.TestCase):
def _testCpuMatmul(self, x, y, transpose_x=False, transpose_y=False):
x_mat = np.matrix(x).T if transpose_x else np.matrix(x)
y_mat = np.matrix(y).T if transpose_y else np.matrix(y)
np_ans = x_mat * y_mat
with self.test_session(use_gpu=False):
tf_ans = tf.matmul(x, y, transpose_x, transpose_y).eval()
self.assertAllClose(np_ans, tf_ans)
self.assertAllEqual(np_ans.shape, tf_ans.shape)
def _testGpuMatmul(self, x, y, transpose_x=False, transpose_y=False):
x_mat = np.matrix(x).T if transpose_x else np.matrix(x)
y_mat = np.matrix(y).T if transpose_y else np.matrix(y)
np_ans = x_mat * y_mat
with self.test_session(use_gpu=True):
tf_ans = tf.matmul(x, y, transpose_x, transpose_y).eval()
self.assertAllClose(np_ans, tf_ans)
self.assertAllEqual(np_ans.shape, tf_ans.shape)
def _randMatrix(self, rows, cols, dtype):
if dtype is np.complex64:
real = self._randMatrix(rows, cols, np.float32)
imag = self._randMatrix(rows, cols, np.float32)
return real + np.complex(0, 1) * imag
else:
return np.random.uniform(low=1.0, high=100.0, size=rows * cols).reshape(
[rows, cols]).astype(dtype)
# Basic test:
# [ [1],
# [2],
# [3], * [1, 2]
# [4] ]
def testFloatBasic(self):
x = np.arange(1., 5.).reshape([4, 1]).astype(np.float32)
y = np.arange(1., 3.).reshape([1, 2]).astype(np.float32)
self._testCpuMatmul(x, y)
self._testGpuMatmul(x, y)
def testDoubleBasic(self):
x = np.arange(1., 5.).reshape([4, 1]).astype(np.float64)
y = np.arange(1., 3.).reshape([1, 2]).astype(np.float64)
self._testCpuMatmul(x, y)
def testInt32Basic(self):
x = np.arange(1., 5.).reshape([4, 1]).astype(np.int32)
y = np.arange(1., 3.).reshape([1, 2]).astype(np.int32)
self._testCpuMatmul(x, y)
def testSComplexBasic(self):
x = np.arange(1., 5.).reshape([4, 1]).astype(np.complex64)
y = np.arange(1., 3.).reshape([1, 2]).astype(np.complex64)
self._testCpuMatmul(x, y)
# Tests testing random sized matrices.
def testFloatRandom(self):
for _ in range(10):
n, k, m = np.random.randint(1, 100, size=3)
x = self._randMatrix(n, k, np.float32)
y = self._randMatrix(k, m, np.float32)
self._testCpuMatmul(x, y)
self._testGpuMatmul(x, y)
def testDoubleRandom(self):
for _ in range(10):
n, k, m = np.random.randint(1, 100, size=3)
x = self._randMatrix(n, k, np.float64)
y = self._randMatrix(k, m, np.float64)
self._testCpuMatmul(x, y)
def testInt32Random(self):
for _ in range(10):
n, k, m = np.random.randint(1, 100, size=3)
x = self._randMatrix(n, k, np.int32)
y = self._randMatrix(k, m, np.int32)
self._testCpuMatmul(x, y)
def testSComplexRandom(self):
for _ in range(10):
n, k, m = np.random.randint(1, 100, size=3)
x = self._randMatrix(n, k, np.complex64)
y = self._randMatrix(k, m, np.complex64)
self._testCpuMatmul(x, y)
# Test the cases that transpose the matrices before multiplying.
# NOTE(keveman): The cases where only one of the inputs is
# transposed are covered by tf.matmul's gradient function.
def testFloatRandomTransposeBoth(self):
for _ in range(10):
n, k, m = np.random.randint(1, 100, size=3)
x = self._randMatrix(k, n, np.float32)
y = self._randMatrix(m, k, np.float32)
self._testCpuMatmul(x, y, True, True)
self._testGpuMatmul(x, y, True, True)
def testDoubleRandomTranposeBoth(self):
for _ in range(10):
n, k, m = np.random.randint(1, 100, size=3)
x = self._randMatrix(k, n, np.float64)
y = self._randMatrix(m, k, np.float64)
self._testCpuMatmul(x, y, True, True)
def testMatMul_OutEmpty_A(self):
n, k, m = 0, 8, 3
x = self._randMatrix(n, k, np.float32)
y = self._randMatrix(k, m, np.float32)
self._testCpuMatmul(x, y)
self._testGpuMatmul(x, y)
def testMatMul_OutEmpty_B(self):
n, k, m = 3, 8, 0
x = self._randMatrix(n, k, np.float32)
y = self._randMatrix(k, m, np.float32)
self._testCpuMatmul(x, y)
self._testGpuMatmul(x, y)
def testMatMul_Inputs_Empty(self):
n, k, m = 3, 0, 4
x = self._randMatrix(n, k, np.float32)
y = self._randMatrix(k, m, np.float32)
self._testCpuMatmul(x, y)
self._testGpuMatmul(x, y)
# TODO(zhifengc): Figures out how to test matmul gradients on GPU.
class MatMulGradientTest(tf.test.TestCase):
def testGradientInput0(self):
with self.test_session(use_gpu=False):
x = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2],
dtype=tf.float64, name="x")
y = tf.constant([1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7],
shape=[2, 4], dtype=tf.float64, name="y")
m = tf.matmul(x, y, name="matmul")
err = gc.ComputeGradientError(x, [3, 2], m, [3, 4])
print "matmul input0 gradient err = ", err
self.assertLess(err, 1e-10)
def testGradientInput1(self):
with self.test_session(use_gpu=False):
x = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2],
dtype=tf.float64, name="x")
y = tf.constant([1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7],
shape=[2, 4], dtype=tf.float64, name="y")
m = tf.matmul(x, y, name="matmul")
err = gc.ComputeGradientError(y, [2, 4], m, [3, 4])
print "matmul input1 gradient err = ", err
self.assertLess(err, 1e-10)
def _VerifyInput0(self, transpose_a, transpose_b):
shape_x = [3, 2]
shape_y = [2, 4]
if transpose_a:
shape_x = list(reversed(shape_x))
if transpose_b:
shape_y = list(reversed(shape_y))
with self.test_session(use_gpu=False):
x = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=shape_x,
dtype=tf.float64, name="x")
y = tf.constant([1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7],
shape=shape_y, dtype=tf.float64, name="y")
m = tf.matmul(x, y, transpose_a, transpose_b, name="matmul")
err = gc.ComputeGradientError(x, shape_x, m, [3, 4])
print "matmul input0 gradient err = ", err
self.assertLess(err, 1e-10)
def testGradientInput0WithTranspose(self):
self._VerifyInput0(transpose_a=True, transpose_b=False)
self._VerifyInput0(transpose_a=False, transpose_b=True)
self._VerifyInput0(transpose_a=True, transpose_b=True)
def _VerifyInput1(self, transpose_a, transpose_b):
shape_x = [3, 2]
shape_y = [2, 4]
if transpose_a:
shape_x = list(reversed(shape_x))
if transpose_b:
shape_y = list(reversed(shape_y))
with self.test_session(use_gpu=False):
x = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=shape_x,
dtype=tf.float64, name="x")
y = tf.constant([1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7],
shape=shape_y, dtype=tf.float64, name="y")
m = tf.matmul(x, y, transpose_a, transpose_b, name="matmul")
err = gc.ComputeGradientError(y, shape_y, m, [3, 4])
print "matmul input1 gradient err = ", err
self.assertLess(err, 1e-10)
def testGradientInput1WithTranspose(self):
self._VerifyInput1(transpose_a=True, transpose_b=False)
self._VerifyInput1(transpose_a=False, transpose_b=True)
self._VerifyInput1(transpose_a=True, transpose_b=True)
if __name__ == "__main__":
tf.test.main()
|
