1# Copyright 2019 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 20import mindspore.nn as nn 21from mindspore import Tensor 22from mindspore.ops import operations as P 23from mindspore.ops.operations import _inner_ops as inner 24from mindspore.common.parameter import Parameter 25from mindspore.common.initializer import initializer 26 27 28class NetConv2d(nn.Cell): 29 def __init__(self): 30 super(NetConv2d, self).__init__() 31 out_channel = 2 32 kernel_size = 1 33 self.conv = P.Conv2D(out_channel, 34 kernel_size, 35 mode=1, 36 pad_mode="valid", 37 pad=0, 38 stride=1, 39 dilation=1, 40 group=1) 41 42 def construct(self, x, w): 43 return self.conv(x, w) 44 45 46@pytest.mark.level0 47@pytest.mark.platform_x86_gpu_training 48@pytest.mark.env_onecard 49def test_conv2d(): 50 x = Tensor(np.arange(1 * 3 * 3 * 3).reshape(1, 3, 3, 3).astype(np.float32)) 51 w = Tensor(np.arange(2 * 3 * 1 * 1).reshape(2, 3, 1, 1).astype(np.float32)) 52 expect = np.array([[[[45, 48, 51], 53 [54, 57, 60], 54 [63, 66, 69]], 55 [[126, 138, 150], 56 [162, 174, 186], 57 [198, 210, 222]]]]).astype(np.float32) 58 59 context.set_context(mode=context.PYNATIVE_MODE, device_target="GPU", max_device_memory="0.2GB") 60 conv2d = NetConv2d() 61 output = conv2d(x, w) 62 assert (output.asnumpy() == expect).all() 63 context.set_context(mode=context.GRAPH_MODE, device_target="GPU") 64 conv2d = NetConv2d() 65 output = conv2d(x, w) 66 assert (output.asnumpy() == expect).all() 67 68 69class NetConv(nn.Cell): 70 def __init__(self, weight, x): 71 super(NetConv, self).__init__() 72 self.conv = nn.Conv2d(in_channels=3, 73 out_channels=3, 74 kernel_size=(5, 3), 75 stride=2, 76 pad_mode='same', 77 padding=(0, 0, 0, 0), 78 dilation=(1, 1), 79 group=1, 80 has_bias=False, 81 weight_init=Tensor(weight) 82 ) 83 self.x = Parameter(initializer(Tensor(x), [1, 3, 4, 2]), name="x") 84 85 def construct(self): 86 return self.conv(self.x) 87 88 89@pytest.mark.level0 90@pytest.mark.platform_x86_gpu_training 91@pytest.mark.env_onecard 92def test_conv(): 93 context.set_context(mode=context.GRAPH_MODE, device_target="GPU") 94 weight = np.array([[[[0.38968208, 0.14398979, 0.7962463], 95 [-2.1836321, -0.63823014, -0.50588065], 96 [0.6660469, 0.64673275, -0.13160042], 97 [1.3683757, 1.4005762, -0.37235805], 98 [-0.22638111, 0.45427424, -0.10293389]], 99 [[1.4985064, -0.29318333, -0.92694616], 100 [1.539068, 0.8937254, -1.2598171], 101 [0.9658142, -0.63945454, -0.23185322], 102 [1.363089, -0.41694695, -2.2750475], 103 [-0.4865508, -1.6938025, 0.609849]], 104 [[1.1844803, 0.99874926, -1.9475793], 105 [0.4987858, 0.5307887, -0.04226681], 106 [0.4529779, -1.1960793, 0.9456575], 107 [3.133675, 0.2309789, -0.29201075], 108 [-0.59632736, -0.0789804, -0.69486314]]], 109 [[[-0.5606142, 0.6420862, 0.2478745], 110 [0.02717604, 1.5483379, -0.9373383], 111 [-1.1017276, -0.259478, 1.0311872], 112 [1.8387799, 0.16468556, 0.33392152], 113 [-1.8781787, 1.0158662, 1.6527579]], 114 115 [[0.45696944, -0.5652523, -1.5618048], 116 [-0.30304828, 0.1331878, -0.36955845], 117 [0.91655576, 0.66612357, 0.3068175], 118 [-0.45732066, 0.8923335, 1.0542952], 119 [-0.73519516, 1.0518405, -1.0273266]], 120 121 [[-0.79712886, -0.26814285, 0.12779616], 122 [1.0367643, -1.6180774, 0.42999932], 123 [-0.81818223, -0.81502074, 0.882194], 124 [0.53640485, 0.4178927, 1.6037121], 125 [0.9256354, -1.1006796, 0.16614541]]], 126 127 [[[-1.5216796, -1.2473261, 0.6549515], 128 [0.63627815, 0.7221449, 0.02977821], 129 [-0.61331123, -0.49451825, 0.33852202], 130 [1.4510741, -1.3818305, -0.791747], 131 [0.6989747, 0.49558765, 1.0813237]], 132 133 [[-0.03969796, 0.71586496, 0.8326594], 134 [-0.15443641, 1.0389746, -0.59301984], 135 [0.7197836, 0.03257621, 1.8398637], 136 [0.6111736, -0.16166899, -2.4869773], 137 [1.3066711, -1.8003578, 0.17412892]], 138 139 [[-0.31470737, -0.5938182, -1.1311078], 140 [-0.99081016, 0.4005125, 0.44154453], 141 [1.0876914, -2.5958562, -0.5914863], 142 [1.3759689, -0.7741513, 0.19928917], 143 [1.6792973, 2.2744863, -0.04308867]]]]).astype(np.float32) 144 x = np.array([[[[-1.4311737, 1.015344], 145 [0.04431088, -2.2886624], 146 [1.4832113, 1.240908], 147 [0.67040104, 0.15266363]], 148 149 [[0.44226435, 1.1461105], 150 [1.194218, 1.5547837], 151 [0.23152256, 1.5911953], 152 [0.11206784, 0.17978816]], 153 154 [[-0.57803905, 0.8039611], 155 [0.0823025, -0.6134477], 156 [-1.4171146, 1.6269946], 157 [0.48878875, 0.9117505]]]]).astype(np.float32) 158 conv2d = NetConv(weight, x) 159 output = conv2d() 160 expected = np.array([[[[2.3498724], 161 [-1.9199573]], 162 [[5.376562], 163 [-5.425745]], 164 [[5.9105043], 165 [7.469034]]]]).astype(np.float32) 166 loss = np.abs(expected - output.asnumpy()) 167 error = 1e-4 * np.ones(loss.shape) 168 assert (loss < error).all() 169 170 171class NetConv2dDynamic(nn.Cell): 172 def __init__(self, axis=0, out_nums=1): 173 super(NetConv2dDynamic, self).__init__() 174 self.dynshape = inner.GpuConvertToDynamicShape() 175 out_channel = 2 176 kernel_size = 1 177 self.conv = P.Conv2D(out_channel, 178 kernel_size, 179 mode=1, 180 pad_mode="valid", 181 pad=0, 182 stride=1, 183 dilation=1, 184 group=1) 185 186 def construct(self, x, w): 187 x_dyn = self.dynshape(x) 188 w_dyn = self.dynshape(w) 189 x_conv = self.conv(x_dyn, w_dyn) 190 return x_conv 191 192 193@pytest.mark.level0 194@pytest.mark.platform_x86_gpu_training 195@pytest.mark.env_onecard 196def test_conv2d_dynamic(): 197 x1 = Tensor(np.arange(1 * 3 * 3 * 3).reshape(1, 3, 3, 3).astype(np.float32)) 198 w1 = Tensor(np.arange(2 * 3 * 1 * 1).reshape(2, 3, 1, 1).astype(np.float32)) 199 expect1 = np.array([[[[45, 48, 51], 200 [54, 57, 60], 201 [63, 66, 69]], 202 [[126, 138, 150], 203 [162, 174, 186], 204 [198, 210, 222]]]]).astype(np.float32) 205 206 x2 = Tensor(np.arange(5 * 1 * 2 * 2).reshape(5, 1, 2, 2).astype(np.float32)) 207 w2 = Tensor(np.arange(2 * 1 * 1 * 1).reshape(2, 1, 1, 1).astype(np.float32)) 208 expect2 = np.array([[[[0., 0.], 209 [0., 0.]], 210 [[0., 1.], 211 [2., 3.]]], 212 [[[0., 0.], 213 [0., 0.]], 214 [[4., 5.], 215 [6., 7.]]], 216 [[[0., 0.], 217 [0., 0.]], 218 [[8., 9.], 219 [10., 11.]]], 220 [[[0., 0.], 221 [0., 0.]], 222 [[12., 13.], 223 [14., 15.]]], 224 [[[0., 0.], 225 [0., 0.]], 226 [[16., 17.], 227 [18., 19.]]]]).astype(np.float32) 228 229 context.set_context(mode=context.GRAPH_MODE, device_target="GPU") 230 conv2d = NetConv2dDynamic() 231 output1 = conv2d(x1, w1) 232 assert (output1.asnumpy() == expect1).all() 233 output2 = conv2d(x2, w2) 234 assert (output2.asnumpy() == expect2).all() 235 236 237class NetConvNHWC(nn.Cell): 238 def __init__(self, weight, x): 239 super(NetConvNHWC, self).__init__() 240 self.conv = nn.Conv2d(in_channels=1, 241 out_channels=3, 242 kernel_size=2, 243 stride=2, 244 pad_mode="valid", 245 weight_init=Tensor(weight), 246 data_format='NHWC' 247 ) 248 self.x = Parameter(initializer(Tensor(x), [1, 4, 4, 1]), name="x") 249 250 def construct(self): 251 return self.conv(self.x) 252 253 254@pytest.mark.level0 255@pytest.mark.platform_x86_gpu_training 256@pytest.mark.env_onecard 257def test_conv_NHWC(): 258 context.set_context(mode=context.GRAPH_MODE, device_target="GPU") 259 x1 = Tensor(np.arange(1 * 4 * 4 * 1).reshape(1, 4, 4, 1).astype(np.float32)) 260 w1 = Tensor(np.arange(3 * 2 * 2 * 1).reshape(3, 2, 2, 1).astype(np.float32)) 261 expected = np.array([[[[24., 64., 104.], 262 [36., 108., 180.]], 263 [[72., 240., 408.], 264 [84., 284., 484.]]]]).astype(np.float32) 265 conv2d = NetConvNHWC(w1, x1) 266 output = conv2d() 267 assert (output.asnumpy() == expected).all() 268