1# Copyright 2020 Huawei Technologies Co., Ltd 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 16import numpy as np 17import pytest 18 19import mindspore.context as context 20from mindspore.common.tensor import Tensor 21from mindspore.nn import Cell 22from mindspore.ops import composite as C 23from mindspore.ops import operations as P 24 25 26class MinimumNet(Cell): 27 def __init__(self): 28 super(MinimumNet, self).__init__() 29 self.min = P.Minimum() 30 31 def construct(self, x1, x2): 32 x = self.min(x1, x2) 33 return x 34 35 36class Grad(Cell): 37 def __init__(self, network): 38 super(Grad, self).__init__() 39 self.grad = C.GradOperation(get_all=True, sens_param=True) 40 self.network = network 41 42 def construct(self, x1, x2, sens): 43 gout = self.grad(self.network)(x1, x2, sens) 44 return gout 45 46 47@pytest.mark.level0 48@pytest.mark.platform_x86_gpu_training 49@pytest.mark.env_onecard 50def test_nobroadcast(): 51 context.set_context(mode=context.GRAPH_MODE, device_target='GPU') 52 53 x1_np = np.random.rand(3, 4).astype(np.float32) 54 x2_np = np.random.rand(3, 4).astype(np.float32) 55 dy_np = np.random.rand(3, 4).astype(np.float32) 56 57 net = Grad(MinimumNet()) 58 output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np)) 59 output0_np = np.where(x1_np < x2_np, dy_np, 0) 60 output1_np = np.where(x1_np < x2_np, 0, dy_np) 61 assert np.allclose(output_ms[0].asnumpy(), output0_np) 62 assert np.allclose(output_ms[1].asnumpy(), output1_np) 63 64 65@pytest.mark.level0 66@pytest.mark.platform_x86_gpu_training 67@pytest.mark.env_onecard 68def test_broadcast(): 69 context.set_context(mode=context.GRAPH_MODE, device_target='GPU') 70 71 x1_np = np.array([[[[0.659578], 72 [0.49113268], 73 [0.75909054], 74 [0.71681815], 75 [0.30421826]]], 76 [[[0.30322495], 77 [0.02858258], 78 [0.06398096], 79 [0.09519596], 80 [0.12498625]]], 81 [[[0.7347768], 82 [0.166469], 83 [0.328553], 84 [0.54908437], 85 [0.23673844]]]]).astype(np.float32) 86 x2_np = np.array([[[[0.9154968, 0.29014662, 0.6492294, 0.39918253, 0.1648203, 0.00861965]], 87 [[0.996885, 0.24152198, 0.3601213, 0.51664376, 0.7933056, 0.84706444]], 88 [[0.75606346, 0.974512, 0.3939527, 0.69697475, 0.83400667, 0.6348955]], 89 [[0.68492866, 0.24609096, 0.4924665, 0.22500521, 0.38474053, 0.5586104]]]]).astype(np.float32) 90 dy_np = np.array([[[[0.42891738, 0.03434946, 0.06192983, 0.21216309, 0.37450036, 0.6619524], 91 [0.8583447, 0.5765161, 0.1468952, 0.9975385, 0.6908136, 0.4903796], 92 [0.68952006, 0.39336833, 0.9049695, 0.66886294, 0.2338471, 0.913618], 93 [0.0428149, 0.6243054, 0.8519898, 0.12088962, 0.9735885, 0.45661286], 94 [0.41563734, 0.41607043, 0.4754915, 0.32207987, 0.33823156, 0.47422352]], 95 96 [[0.64478457, 0.22430937, 0.7682554, 0.46082005, 0.8938723, 0.20490853], 97 [0.44393885, 0.08278944, 0.4734108, 0.5543551, 0.39428464, 0.44424313], 98 [0.12612297, 0.76566416, 0.71133816, 0.81280327, 0.20583127, 0.54058075], 99 [0.41341263, 0.48118508, 0.00401995, 0.37259838, 0.05435474, 0.5240658], 100 [0.4081956, 0.48718935, 0.9132831, 0.67969185, 0.0119757, 0.8328054]], 101 102 [[0.91695577, 0.95370644, 0.263782, 0.7477626, 0.6448147, 0.8080634], 103 [0.15576603, 0.9104615, 0.3778708, 0.6912833, 0.2092224, 0.67462957], 104 [0.7087075, 0.7888326, 0.4672294, 0.98221505, 0.25210258, 0.98920417], 105 [0.7466197, 0.22702982, 0.01991269, 0.6846591, 0.7515228, 0.5890395], 106 [0.04531088, 0.21740614, 0.8406235, 0.36480767, 0.37733936, 0.02914464]], 107 108 [[0.33069974, 0.5497569, 0.9896345, 0.4167176, 0.78057563, 0.04659131], 109 [0.7747768, 0.21427679, 0.29893255, 0.7706969, 0.9755185, 0.42388415], 110 [0.3910244, 0.39381978, 0.37065396, 0.15558061, 0.05012341, 0.15870963], 111 [0.17791101, 0.47219893, 0.13899496, 0.32323205, 0.3628809, 0.02580585], 112 [0.30274773, 0.62890774, 0.11024303, 0.6980051, 0.35346958, 0.062852]]], 113 114 [[[0.6925081, 0.74668753, 0.80145043, 0.06598313, 0.665123, 0.15073007], 115 [0.11784806, 0.6385372, 0.5228278, 0.5349848, 0.84671104, 0.8096436], 116 [0.09516156, 0.63298017, 0.52382874, 0.36734378, 0.66497755, 0.6019127], 117 [0.46438488, 0.0194377, 0.9388292, 0.7286089, 0.29178405, 0.11872514], 118 [0.22101837, 0.6164887, 0.6139798, 0.11711904, 0.6227745, 0.09701069]], 119 120 [[0.80480653, 0.90034056, 0.8633447, 0.97415197, 0.08309154, 0.8446033], 121 [0.9473769, 0.791024, 0.26339203, 0.01155075, 0.2673186, 0.7116369], 122 [0.9687511, 0.24281934, 0.37777108, 0.09802654, 0.2421312, 0.87095344], 123 [0.6311381, 0.23368953, 0.0998995, 0.4364419, 0.9187446, 0.5043872], 124 [0.35226053, 0.09357589, 0.41317305, 0.85930043, 0.16249318, 0.5478765]], 125 126 [[0.14338651, 0.24859418, 0.4246941, 0.73034066, 0.47172204, 0.8717199], 127 [0.05415315, 0.78556925, 0.99214983, 0.7415298, 0.673708, 0.87817156], 128 [0.616975, 0.42843062, 0.05179814, 0.1566958, 0.04536059, 0.70166487], 129 [0.15493333, 0.776598, 0.4361967, 0.40253627, 0.89210516, 0.8144414], 130 [0.04816005, 0.29696834, 0.4586605, 0.3419852, 0.5595613, 0.74093205]], 131 132 [[0.1388035, 0.9168704, 0.64287645, 0.83864623, 0.48026922, 0.78323376], 133 [0.12724937, 0.83034366, 0.42557436, 0.50578654, 0.25630295, 0.15349793], 134 [0.27256685, 0.04547984, 0.5385756, 0.39270344, 0.7661698, 0.23722854], 135 [0.24620503, 0.25431684, 0.71564585, 0.01161419, 0.846467, 0.7043044], 136 [0.63272387, 0.11857849, 0.3772076, 0.16758402, 0.46743023, 0.05919575]]], 137 138 [[[0.18827082, 0.8912264, 0.6841404, 0.74436826, 0.9582085, 0.1083683], 139 [0.60695344, 0.09742349, 0.25074378, 0.87940735, 0.21116392, 0.39418384], 140 [0.744686, 0.35679692, 0.01308284, 0.45166633, 0.68166, 0.8634658], 141 [0.7331758, 0.21113694, 0.3935488, 0.87934476, 0.70728546, 0.09309767], 142 [0.12128611, 0.93696386, 0.81177396, 0.85402405, 0.5827289, 0.9776509]], 143 144 [[0.54069614, 0.66651285, 0.10646132, 0.17342485, 0.88795924, 0.03551182], 145 [0.25531697, 0.87946486, 0.74267226, 0.89230734, 0.95171434, 0.94697934], 146 [0.3708397, 0.507355, 0.97099817, 0.4918163, 0.17212386, 0.5008048], 147 [0.62530744, 0.25210327, 0.73966664, 0.71555346, 0.82484317, 0.6094874], 148 [0.4589691, 0.1386695, 0.27448782, 0.20373994, 0.27805242, 0.23292768]], 149 150 [[0.7414099, 0.2270226, 0.90431255, 0.47035843, 0.9581062, 0.5359226], 151 [0.79603523, 0.45549425, 0.80858237, 0.7705133, 0.017761, 0.98001194], 152 [0.06013146, 0.99240226, 0.33515573, 0.04110833, 0.41470334, 0.7130743], 153 [0.5687417, 0.5788611, 0.00722461, 0.6603336, 0.3420471, 0.75181854], 154 [0.4699261, 0.51390815, 0.343182, 0.81498754, 0.8942413, 0.46532857]], 155 156 [[0.4589523, 0.5534698, 0.2825786, 0.8205943, 0.78258514, 0.43154418], 157 [0.27020997, 0.01667354, 0.60871965, 0.90670526, 0.3208025, 0.96995634], 158 [0.85337156, 0.9711295, 0.1381724, 0.53670496, 0.7347996, 0.73380876], 159 [0.6137464, 0.54751194, 0.9037335, 0.23134394, 0.61411524, 0.26583543], 160 [0.70770144, 0.01813207, 0.24718016, 0.70329237, 0.7062925, 0.14399007]]]]).astype(np.float32) 161 162 expect_dx1 = np.array([[[[5.7664223], 163 [6.981018], 164 [2.6029902], 165 [2.7598202], 166 [6.763105]]], 167 [[[10.06558], 168 [12.077246], 169 [9.338394], 170 [11.52271], 171 [8.889048]]], 172 [[[3.5789769], 173 [13.424448], 174 [8.732746], 175 [6.9677467], 176 [9.635765]]]]).astype(np.float32) 177 178 expect_dx2 = np.array([[[[0., 4.250458, 2.5030296, 3.623167, 6.4171505, 7.2115746]], 179 [[0., 4.367449, 2.803152, 2.5352, 0., 0.]], 180 [[0.7087075, 0., 2.040332, 2.1372325, 0., 2.9222295]], 181 [[1.0278877, 5.247942, 2.6855955, 5.494814, 3.5657988, 182 0.66265094]]]]).astype(np.float32) 183 184 net = Grad(MinimumNet()) 185 output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np)) 186 assert np.allclose(output_ms[0].asnumpy(), expect_dx1) 187 assert np.allclose(output_ms[1].asnumpy(), expect_dx2) 188 189 190@pytest.mark.level0 191@pytest.mark.platform_x86_gpu_training 192@pytest.mark.env_onecard 193def test_broadcast_diff_dims(): 194 context.set_context(mode=context.GRAPH_MODE, device_target='GPU') 195 196 x1_np = np.array([[[0.275478, 0.48933202, 0.71846116], 197 [0.9803821, 0.57205725, 0.28511533]], 198 [[0.61111903, 0.9671023, 0.70624334], 199 [0.53730786, 0.90413177, 0.94349676]]]).astype(np.float32) 200 201 x2_np = np.array([[0.01045662, 0.82126397, 0.6365063], 202 [0.9900942, 0.6584232, 0.98537433]]).astype(np.float32) 203 204 dy_np = np.array([[[0.3897645, 0.61152864, 0.33675498], 205 [0.5303635, 0.84893036, 0.4959739]], 206 [[0.5391046, 0.8443047, 0.4174708], 207 [0.57513475, 0.9225578, 0.46760973]]]).astype(np.float32) 208 209 expect_dx1 = np.array([[[0., 0.61152864, 0.], 210 [0.5303635, 0.84893036, 0.4959739]], 211 [[0., 0., 0.], 212 [0.57513475, 0., 0.46760973]]]).astype(np.float32) 213 214 expect_dx2 = np.array([[0.92886907, 0.8443047, 0.7542258], 215 [0., 0.9225578, 0.]]).astype(np.float32) 216 217 net = Grad(MinimumNet()) 218 output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np)) 219 assert np.allclose(output_ms[0].asnumpy(), expect_dx1) 220 assert np.allclose(output_ms[1].asnumpy(), expect_dx2) 221 222 223@pytest.mark.level0 224@pytest.mark.platform_x86_gpu_training 225@pytest.mark.env_onecard 226def test_broadcast_int32(): 227 context.set_context(mode=context.GRAPH_MODE, device_target='GPU') 228 229 x1_np = np.random.rand(3, 4).astype(np.int32) 230 x2_np = np.random.rand(3, 4).astype(np.int32) 231 dy_np = np.random.rand(3, 4).astype(np.int32) 232 233 net = Grad(MinimumNet()) 234 output_ms = net(Tensor(x1_np), Tensor(x2_np), Tensor(dy_np)) 235 output0_np = np.where(x1_np < x2_np, dy_np, 0) 236 output1_np = np.where(x1_np < x2_np, 0, dy_np) 237 assert np.allclose(output_ms[0].asnumpy(), output0_np) 238 assert np.allclose(output_ms[1].asnumpy(), output1_np) 239