1 # Copyright 2015 The TensorFlow Authors. All Rights Reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # ============================================================================== 15 """Tests for tensorflow.ops.tf.matmul.""" 16 17 from __future__ import absolute_import 18 from __future__ import division 19 from __future__ import print_function 20 21 import numpy as np 22 23 from tensorflow.python.framework import constant_op 24 from tensorflow.python.framework import dtypes 25 from tensorflow.python.ops import gradient_checker 26 from tensorflow.python.ops import math_ops 27 from tensorflow.python.platform import test 28 29 30 def RandMatrix(rows, cols, tr, round_bfloat=False): 31 if tr: 32 rows, cols = cols, rows 33 rand_func = np.random.randint if round_bfloat else np.random.uniform 34 return (np.clip( 35 rand_func( 36 low=-256.0, high=256.0, size=rows * cols), -64, 37 64) / 128.0).reshape([rows, cols]).astype(np.float32) 38 39 40 class SparseMatMulTest(test.TestCase): 41 42 def _testCpuMatmul(self, 43 x, 44 y, 45 tr_a=False, 46 tr_b=False, 47 sp_a=True, 48 sp_b=False, 49 x_dtype=dtypes.float32, 50 y_dtype=dtypes.float32): 51 with self.test_session(use_gpu=False): 52 tf_x = math_ops.cast(x, x_dtype) 53 tf_y = math_ops.cast(y, y_dtype) 54 tf_ans = math_ops.matmul( 55 tf_x, 56 tf_y, 57 transpose_a=tr_a, 58 transpose_b=tr_b, 59 a_is_sparse=sp_a, 60 b_is_sparse=sp_b) 61 out = tf_ans.eval() 62 np_x = math_ops.cast(tf_x, dtypes.float32).eval() 63 np_y = math_ops.cast(tf_y, dtypes.float32).eval() 64 65 if tr_a: 66 np_x = np.transpose(np_x) 67 if tr_b: 68 np_y = np.transpose(np_y) 69 70 np_ans = np.matrix(np_x) * np.matrix(np_y) 71 self.assertShapeEqual(np_ans, tf_ans) 72 self.assertAllCloseAccordingToType(np_ans, out, rtol=1e-4, atol=1e-4) 73 74 def testBasic(self): 75 x = np.arange(0., 4.).reshape([4, 1]).astype(np.float32) 76 y = np.arange(-1., 1.).reshape([1, 2]).astype(np.float32) 77 for x_dtype in (dtypes.float32, dtypes.bfloat16): 78 for y_dtype in (dtypes.float32, dtypes.bfloat16): 79 self._testCpuMatmul(x, y, x_dtype=x_dtype, y_dtype=y_dtype) 80 81 def testZeroDim(self): 82 x = np.ones((4, 0)).astype(np.float32) 83 y = np.ones((0, 3)).astype(np.float32) 84 for x_dtype in (dtypes.float32, dtypes.bfloat16): 85 for y_dtype in (dtypes.float32, dtypes.bfloat16): 86 self._testCpuMatmul(x, y, x_dtype=x_dtype, y_dtype=y_dtype) 87 88 def testEmpty(self): 89 x = np.ones((0, 0)).astype(np.float32) 90 y = np.ones((0, 0)).astype(np.float32) 91 for x_dtype in (dtypes.float32, dtypes.bfloat16): 92 for y_dtype in (dtypes.float32, dtypes.bfloat16): 93 self._testCpuMatmul(x, y, x_dtype=x_dtype, y_dtype=y_dtype) 94 95 # Tests setting one dimension to be a high value. 96 def testLarge(self): 97 r1 = np.random.randint(6000, 20000) 98 r2 = np.random.randint(1, 10) 99 r3 = np.random.randint(1, 10) 100 for m, k, n in [(r1, r2, r3), (r2, r1, r3), (r2, r3, r1)]: 101 for x_dtype in (dtypes.float32, dtypes.bfloat16): 102 for y_dtype in (dtypes.float32, dtypes.bfloat16): 103 x = RandMatrix(m, k, False) 104 y = RandMatrix(k, n, False) 105 self._testCpuMatmul(x, y, x_dtype=x_dtype, y_dtype=y_dtype) 106 107 # Tests random sized matrices. 108 def testRandom(self): 109 for tr_a in [True, False]: 110 for tr_b in [True, False]: 111 for sp_a in [True, False]: 112 for sp_b in [True, False]: 113 for x_dtype in (dtypes.float32, dtypes.bfloat16): 114 for y_dtype in (dtypes.float32, dtypes.bfloat16): 115 n, k, m = np.random.randint(1, 100, size=3) 116 x = RandMatrix(n, k, tr_a) 117 y = RandMatrix(k, m, tr_b) 118 self._testCpuMatmul( 119 x, 120 y, 121 tr_a, 122 tr_b, 123 sp_a, 124 sp_b, 125 x_dtype=x_dtype, 126 y_dtype=y_dtype) 127 128 129 class MatMulGradientTest(test.TestCase): 130 131 def _testGradients(self, tr_a, tr_b, sp_a, sp_b, a_dtype, b_dtype, delta, 132 name): 133 with self.test_session(): 134 a = constant_op.constant( 135 RandMatrix( 136 3, 2, tr_a, round_bfloat=True), dtype=dtypes.float32) 137 b = constant_op.constant( 138 RandMatrix( 139 2, 4, tr_b, round_bfloat=True), dtype=dtypes.float32) 140 tf_a = math_ops.cast(a, a_dtype) if a_dtype != dtypes.float32 else a 141 tf_b = math_ops.cast(b, b_dtype) if b_dtype != dtypes.float32 else b 142 143 m = math_ops.matmul( 144 tf_a, 145 tf_b, 146 name=name, 147 transpose_a=tr_a, 148 transpose_b=tr_b, 149 a_is_sparse=sp_a, 150 b_is_sparse=sp_b) 151 err = (gradient_checker.compute_gradient_error( 152 a, [2, 3] if tr_a else [3, 2], 153 m, [3, 4], 154 x_init_value=a.eval(), 155 delta=delta) + gradient_checker.compute_gradient_error( 156 b, [4, 2] if tr_b else [2, 4], 157 m, [3, 4], 158 x_init_value=b.eval(), 159 delta=delta)) 160 self.assertLess(err, delta / 2.) 161 162 def testGradientInput(self): 163 for tr_a in [True, False]: 164 for tr_b in [True, False]: 165 for sp_a in [True, False]: 166 for sp_b in [True, False]: 167 for a_dtype in (dtypes.float32, dtypes.bfloat16): 168 for b_dtype in (dtypes.float32, dtypes.bfloat16): 169 # Note: bfloat16 only has 7 mantissa bits, versus float32 with 170 # 10. Hence, we shift by 2 bits to pass the test. 171 if a_dtype == dtypes.bfloat16 and b_dtype == dtypes.bfloat16: 172 delta = 1 / 16. 173 else: 174 delta = 1 / 64. 175 name = "sparse_matmul_%s_%s_%s_%s" % (tr_a, tr_b, sp_a, sp_b) 176 self._testGradients(tr_a, tr_b, sp_a, sp_b, a_dtype, b_dtype, 177 delta, name) 178 179 180 if __name__ == "__main__": 181 test.main() 182