modules/dnn/test/test_tf_importer.cpp

   1 // This file is part of OpenCV project.
   2 // It is subject to the license terms in the LICENSE file found in the top-level directory
   3 // of this distribution and at http://opencv.org/license.html.
   4
   5 // Copyright (C) 2017, Intel Corporation, all rights reserved.
   6 // Third party copyrights are property of their respective owners.
   7
   8 /*
   9 Test for Tensorflow models loading
  10 */
  11
  12 #include "test_precomp.hpp"
  13 #include "npy_blob.hpp"
  14
  15 namespace opencv_test
  16 {
  17
  18 using namespace cv;
  19 using namespace cv::dnn;
  20
  21 template<typename TString>
  22 static std::string _tf(TString filename)
  23 {
  24     return (getOpenCVExtraDir() + "/dnn/") + filename;
  25 }
  26
  27 TEST(Test_TensorFlow, read_inception)
  28 {
  29     Net net;
  30     {
  31         const string model = findDataFile("dnn/tensorflow_inception_graph.pb", false);
  32         net = readNetFromTensorflow(model);
  33         ASSERT_FALSE(net.empty());
  34     }
  35
  36     Mat sample = imread(_tf("grace_hopper_227.png"));
  37     ASSERT_TRUE(!sample.empty());
  38     Mat input;
  39     resize(sample, input, Size(224, 224));
  40     input -= 128; // mean sub
  41
  42     Mat inputBlob = blobFromImage(input);
  43
  44     net.setInput(inputBlob, "input");
  45     Mat out = net.forward("softmax2");
  46
  47     std::cout << out.dims << std::endl;
  48 }
  49
  50 TEST(Test_TensorFlow, inception_accuracy)
  51 {
  52     Net net;
  53     {
  54         const string model = findDataFile("dnn/tensorflow_inception_graph.pb", false);
  55         net = readNetFromTensorflow(model);
  56         ASSERT_FALSE(net.empty());
  57     }
  58
  59     Mat sample = imread(_tf("grace_hopper_227.png"));
  60     ASSERT_TRUE(!sample.empty());
  61     resize(sample, sample, Size(224, 224));
  62     Mat inputBlob = blobFromImage(sample);
  63
  64     net.setInput(inputBlob, "input");
  65     Mat out = net.forward("softmax2");
  66
  67     Mat ref = blobFromNPY(_tf("tf_inception_prob.npy"));
  68
  69     normAssert(ref, out);
  70 }
  71
  72 static std::string path(const std::string& file)
  73 {
  74     return findDataFile("dnn/tensorflow/" + file, false);
  75 }
  76
  77 static void runTensorFlowNet(const std::string& prefix, int targetId = DNN_TARGET_CPU, bool hasText = false,
  78                              double l1 = 1e-5, double lInf = 1e-4,
  79                              bool memoryLoad = false)
  80 {
  81     std::string netPath = path(prefix + "_net.pb");
  82     std::string netConfig = (hasText ? path(prefix + "_net.pbtxt") : "");
  83     std::string inpPath = path(prefix + "_in.npy");
  84     std::string outPath = path(prefix + "_out.npy");
  85
  86     Net net;
  87     if (memoryLoad)
  88     {
  89         // Load files into a memory buffers
  90         string dataModel;
  91         ASSERT_TRUE(readFileInMemory(netPath, dataModel));
  92
  93         string dataConfig;
  94         if (hasText)
  95             ASSERT_TRUE(readFileInMemory(netConfig, dataConfig));
  96
  97         net = readNetFromTensorflow(dataModel.c_str(), dataModel.size(),
  98                                     dataConfig.c_str(), dataConfig.size());
  99     }
 100     else
 101         net = readNetFromTensorflow(netPath, netConfig);
 102
 103     ASSERT_FALSE(net.empty());
 104
 105     net.setPreferableBackend(DNN_BACKEND_DEFAULT);
 106     net.setPreferableTarget(targetId);
 107
 108     cv::Mat input = blobFromNPY(inpPath);
 109     cv::Mat target = blobFromNPY(outPath);
 110
 111     net.setInput(input);
 112     cv::Mat output = net.forward();
 113     normAssert(target, output, "", l1, lInf);
 114 }
 115
 116 typedef testing::TestWithParam<DNNTarget> Test_TensorFlow_layers;
 117
 118 TEST_P(Test_TensorFlow_layers, conv)
 119 {
 120     int targetId = GetParam();
 121     runTensorFlowNet("single_conv", targetId);
 122     runTensorFlowNet("atrous_conv2d_valid", targetId);
 123     runTensorFlowNet("atrous_conv2d_same", targetId);
 124     runTensorFlowNet("depthwise_conv2d", targetId);
 125 }
 126
 127 TEST_P(Test_TensorFlow_layers, padding)
 128 {
 129     int targetId = GetParam();
 130     runTensorFlowNet("padding_same", targetId);
 131     runTensorFlowNet("padding_valid", targetId);
 132     runTensorFlowNet("spatial_padding", targetId);
 133 }
 134
 135 TEST_P(Test_TensorFlow_layers, eltwise_add_mul)
 136 {
 137     runTensorFlowNet("eltwise_add_mul", GetParam());
 138 }
 139
 140 TEST_P(Test_TensorFlow_layers, pad_and_concat)
 141 {
 142     runTensorFlowNet("pad_and_concat", GetParam());
 143 }
 144
 145 TEST_P(Test_TensorFlow_layers, batch_norm)
 146 {
 147     int targetId = GetParam();
 148     runTensorFlowNet("batch_norm", targetId);
 149     runTensorFlowNet("fused_batch_norm", targetId);
 150     runTensorFlowNet("batch_norm_text", targetId, true);
 151     runTensorFlowNet("mvn_batch_norm", targetId);
 152     runTensorFlowNet("mvn_batch_norm_1x1", targetId);
 153     runTensorFlowNet("unfused_batch_norm", targetId);
 154     runTensorFlowNet("fused_batch_norm_no_gamma", targetId);
 155     runTensorFlowNet("unfused_batch_norm_no_gamma", targetId);
 156 }
 157
 158 TEST_P(Test_TensorFlow_layers, pooling)
 159 {
 160     int targetId = GetParam();
 161     runTensorFlowNet("max_pool_even", targetId);
 162     runTensorFlowNet("max_pool_odd_valid", targetId);
 163     runTensorFlowNet("ave_pool_same", targetId);
 164     runTensorFlowNet("max_pool_odd_same", targetId);
 165     runTensorFlowNet("reduce_mean", targetId);  // an average pooling over all spatial dimensions.
 166 }
 167
 168 TEST_P(Test_TensorFlow_layers, deconvolution)
 169 {
 170     int targetId = GetParam();
 171     runTensorFlowNet("deconvolution", targetId);
 172     runTensorFlowNet("deconvolution_same", targetId);
 173     runTensorFlowNet("deconvolution_stride_2_same", targetId);
 174     runTensorFlowNet("deconvolution_adj_pad_valid", targetId);
 175     runTensorFlowNet("deconvolution_adj_pad_same", targetId);
 176 }
 177
 178 TEST_P(Test_TensorFlow_layers, matmul)
 179 {
 180     int targetId = GetParam();
 181     runTensorFlowNet("matmul", targetId);
 182     runTensorFlowNet("nhwc_reshape_matmul", targetId);
 183     runTensorFlowNet("nhwc_transpose_reshape_matmul", targetId);
 184 }
 185
 186 TEST_P(Test_TensorFlow_layers, reshape)
 187 {
 188     int targetId = GetParam();
 189     runTensorFlowNet("shift_reshape_no_reorder", targetId);
 190     runTensorFlowNet("reshape_reduce", targetId);
 191     runTensorFlowNet("flatten", targetId, true);
 192     runTensorFlowNet("unfused_flatten", targetId);
 193     runTensorFlowNet("unfused_flatten_unknown_batch", targetId);
 194 }
 195
 196 INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_layers, availableDnnTargets());
 197
 198 typedef testing::TestWithParam<DNNTarget> Test_TensorFlow_nets;
 199
 200 TEST_P(Test_TensorFlow_nets, MobileNet_SSD)
 201 {
 202     std::string netPath = findDataFile("dnn/ssd_mobilenet_v1_coco.pb", false);
 203     std::string netConfig = findDataFile("dnn/ssd_mobilenet_v1_coco.pbtxt", false);
 204     std::string imgPath = findDataFile("dnn/street.png", false);
 205
 206     Mat inp;
 207     resize(imread(imgPath), inp, Size(300, 300));
 208     inp = blobFromImage(inp, 1.0f / 127.5, Size(), Scalar(127.5, 127.5, 127.5), true);
 209
 210     std::vector<String> outNames(3);
 211     outNames[0] = "concat";
 212     outNames[1] = "concat_1";
 213     outNames[2] = "detection_out";
 214
 215     std::vector<Mat> target(outNames.size());
 216     for (int i = 0; i < outNames.size(); ++i)
 217     {
 218         std::string path = findDataFile("dnn/tensorflow/ssd_mobilenet_v1_coco." + outNames[i] + ".npy", false);
 219         target[i] = blobFromNPY(path);
 220     }
 221
 222     Net net = readNetFromTensorflow(netPath, netConfig);
 223
 224     net.setPreferableTarget(GetParam());
 225
 226     net.setInput(inp);
 227
 228     std::vector<Mat> output;
 229     net.forward(output, outNames);
 230
 231     normAssert(target[0].reshape(1, 1), output[0].reshape(1, 1), "", 1e-5, 1.5e-4);
 232     normAssert(target[1].reshape(1, 1), output[1].reshape(1, 1), "", 1e-5, 3e-4);
 233     normAssert(target[2].reshape(1, 1), output[2].reshape(1, 1), "", 4e-5, 1e-2);
 234 }
 235
 236 TEST_P(Test_TensorFlow_nets, Inception_v2_SSD)
 237 {
 238     std::string proto = findDataFile("dnn/ssd_inception_v2_coco_2017_11_17.pbtxt", false);
 239     std::string model = findDataFile("dnn/ssd_inception_v2_coco_2017_11_17.pb", false);
 240
 241     Net net = readNetFromTensorflow(model, proto);
 242     Mat img = imread(findDataFile("dnn/street.png", false));
 243     Mat blob = blobFromImage(img, 1.0f / 127.5, Size(300, 300), Scalar(127.5, 127.5, 127.5), true, false);
 244
 245     net.setPreferableTarget(GetParam());
 246
 247     net.setInput(blob);
 248     // Output has shape 1x1xNx7 where N - number of detections.
 249     // An every detection is a vector of values [id, classId, confidence, left, top, right, bottom]
 250     Mat out = net.forward();
 251     out = out.reshape(1, out.total() / 7);
 252
 253     Mat detections;
 254     for (int i = 0; i < out.rows; ++i)
 255     {
 256         if (out.at<float>(i, 2) > 0.5)
 257           detections.push_back(out.row(i).colRange(1, 7));
 258     }
 259
 260     Mat ref = (Mat_<float>(5, 6) << 1, 0.90176028, 0.19872092, 0.36311883, 0.26461923, 0.63498729,
 261                                     3, 0.93569964, 0.64865261, 0.45906419, 0.80675775, 0.65708131,
 262                                     3, 0.75838411, 0.44668293, 0.45907149, 0.49459291, 0.52197015,
 263                                     10, 0.95932811, 0.38349164, 0.32528657, 0.40387636, 0.39165527,
 264                                     10, 0.93973452, 0.66561931, 0.37841269, 0.68074018, 0.42907384);
 265     normAssert(detections, ref);
 266 }
 267
 268 TEST_P(Test_TensorFlow_nets, opencv_face_detector_uint8)
 269 {
 270     std::string proto = findDataFile("dnn/opencv_face_detector.pbtxt", false);
 271     std::string model = findDataFile("dnn/opencv_face_detector_uint8.pb", false);
 272
 273     Net net = readNetFromTensorflow(model, proto);
 274     Mat img = imread(findDataFile("gpu/lbpcascade/er.png", false));
 275     Mat blob = blobFromImage(img, 1.0, Size(), Scalar(104.0, 177.0, 123.0), false, false);
 276
 277     net.setPreferableTarget(GetParam());
 278
 279     net.setInput(blob);
 280     // Output has shape 1x1xNx7 where N - number of detections.
 281     // An every detection is a vector of values [id, classId, confidence, left, top, right, bottom]
 282     Mat out = net.forward();
 283
 284     // References are from test for Caffe model.
 285     Mat ref = (Mat_<float>(6, 5) << 0.99520785, 0.80997437, 0.16379407, 0.87996572, 0.26685631,
 286                                     0.9934696, 0.2831718, 0.50738752, 0.345781, 0.5985168,
 287                                     0.99096733, 0.13629119, 0.24892329, 0.19756334, 0.3310290,
 288                                     0.98977017, 0.23901358, 0.09084064, 0.29902688, 0.1769477,
 289                                     0.97203469, 0.67965847, 0.06876482, 0.73999709, 0.1513494,
 290                                     0.95097077, 0.51901293, 0.45863652, 0.5777427, 0.5347801);
 291     normAssert(out.reshape(1, out.total() / 7).rowRange(0, 6).colRange(2, 7), ref, "", 2.8e-4, 3.4e-3);
 292 }
 293
 294 INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_nets, availableDnnTargets());
 295
 296 TEST(Test_TensorFlow, defun)
 297 {
 298     runTensorFlowNet("defun_dropout");
 299 }
 300
 301 TEST(Test_TensorFlow, fp16)
 302 {
 303     const float l1 = 1e-3;
 304     const float lInf = 1e-2;
 305     runTensorFlowNet("fp16_single_conv", DNN_TARGET_CPU, false, l1, lInf);
 306     runTensorFlowNet("fp16_deconvolution", DNN_TARGET_CPU, false, l1, lInf);
 307     runTensorFlowNet("fp16_max_pool_odd_same", DNN_TARGET_CPU, false, l1, lInf);
 308     runTensorFlowNet("fp16_padding_valid", DNN_TARGET_CPU, false, l1, lInf);
 309     runTensorFlowNet("fp16_eltwise_add_mul", DNN_TARGET_CPU, false, l1, lInf);
 310     runTensorFlowNet("fp16_max_pool_odd_valid", DNN_TARGET_CPU, false, l1, lInf);
 311     runTensorFlowNet("fp16_pad_and_concat", DNN_TARGET_CPU, false, l1, lInf);
 312     runTensorFlowNet("fp16_max_pool_even", DNN_TARGET_CPU, false, l1, lInf);
 313     runTensorFlowNet("fp16_padding_same", DNN_TARGET_CPU, false, l1, lInf);
 314 }
 315
 316 TEST(Test_TensorFlow, quantized)
 317 {
 318     runTensorFlowNet("uint8_single_conv");
 319 }
 320
 321 TEST(Test_TensorFlow, lstm)
 322 {
 323     runTensorFlowNet("lstm", DNN_TARGET_CPU, true);
 324 }
 325
 326 TEST(Test_TensorFlow, split)
 327 {
 328     runTensorFlowNet("split_equals");
 329 }
 330
 331 TEST(Test_TensorFlow, resize_nearest_neighbor)
 332 {
 333     runTensorFlowNet("resize_nearest_neighbor");
 334 }
 335
 336 TEST(Test_TensorFlow, slice)
 337 {
 338     runTensorFlowNet("slice_4d");
 339 }
 340
 341 TEST(Test_TensorFlow, softmax)
 342 {
 343     runTensorFlowNet("keras_softmax");
 344 }
 345
 346 TEST(Test_TensorFlow, relu6)
 347 {
 348     runTensorFlowNet("keras_relu6");
 349 }
 350
 351 TEST(Test_TensorFlow, keras_mobilenet_head)
 352 {
 353     runTensorFlowNet("keras_mobilenet_head");
 354 }
 355
 356 TEST(Test_TensorFlow, memory_read)
 357 {
 358     double l1 = 1e-5;
 359     double lInf = 1e-4;
 360     runTensorFlowNet("lstm", DNN_TARGET_CPU, true, l1, lInf, true);
 361
 362     runTensorFlowNet("batch_norm", DNN_TARGET_CPU, false, l1, lInf, true);
 363     runTensorFlowNet("fused_batch_norm", DNN_TARGET_CPU, false, l1, lInf, true);
 364     runTensorFlowNet("batch_norm_text", DNN_TARGET_CPU, true, l1, lInf, true);
 365 }
 366
 367 }