modules/dnn/test/test_torch_importer.cpp

   1 /*M///////////////////////////////////////////////////////////////////////////////////////
   2 //
   3 //  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
   4 //
   5 //  By downloading, copying, installing or using the software you agree to this license.
   6 //  If you do not agree to this license, do not download, install,
   7 //  copy or use the software.
   8 //
   9 //
  10 //                           License Agreement
  11 //                For Open Source Computer Vision Library
  12 //
  13 // Copyright (C) 2013, OpenCV Foundation, all rights reserved.
  14 // Third party copyrights are property of their respective owners.
  15 //
  16 // Redistribution and use in source and binary forms, with or without modification,
  17 // are permitted provided that the following conditions are met:
  18 //
  19 //   * Redistribution's of source code must retain the above copyright notice,
  20 //     this list of conditions and the following disclaimer.
  21 //
  22 //   * Redistribution's in binary form must reproduce the above copyright notice,
  23 //     this list of conditions and the following disclaimer in the documentation
  24 //     and/or other materials provided with the distribution.
  25 //
  26 //   * The name of the copyright holders may not be used to endorse or promote products
  27 //     derived from this software without specific prior written permission.
  28 //
  29 // This software is provided by the copyright holders and contributors "as is" and
  30 // any express or implied warranties, including, but not limited to, the implied
  31 // warranties of merchantability and fitness for a particular purpose are disclaimed.
  32 // In no event shall the Intel Corporation or contributors be liable for any direct,
  33 // indirect, incidental, special, exemplary, or consequential damages
  34 // (including, but not limited to, procurement of substitute goods or services;
  35 // loss of use, data, or profits; or business interruption) however caused
  36 // and on any theory of liability, whether in contract, strict liability,
  37 // or tort (including negligence or otherwise) arising in any way out of
  38 // the use of this software, even if advised of the possibility of such damage.
  39 //
  40 //M*/
  41
  42 #include "test_precomp.hpp"
  43 #include "npy_blob.hpp"
  44 #include <opencv2/dnn/shape_utils.hpp>
  45 #include <opencv2/dnn/layer.details.hpp>  // CV_DNN_REGISTER_LAYER_CLASS
  46
  47 namespace opencv_test
  48 {
  49
  50 using namespace std;
  51 using namespace testing;
  52 using namespace cv;
  53 using namespace cv::dnn;
  54
  55 template<typename TStr>
  56 static std::string _tf(TStr filename, bool inTorchDir = true)
  57 {
  58     String path = "dnn/";
  59     if (inTorchDir)
  60         path += "torch/";
  61     path += filename;
  62     return findDataFile(path, false);
  63 }
  64
  65 TEST(Torch_Importer, simple_read)
  66 {
  67     Net net;
  68     ASSERT_NO_THROW(net = readNetFromTorch(_tf("net_simple_net.txt"), false));
  69     ASSERT_FALSE(net.empty());
  70 }
  71
  72 class Test_Torch_layers : public DNNTestLayer
  73 {
  74 public:
  75     void runTorchNet(const String& prefix, String outLayerName = "",
  76                      bool check2ndBlob = false, bool isBinary = false, bool evaluate = true,
  77                      double l1 = 0.0, double lInf = 0.0)
  78     {
  79         String suffix = (isBinary) ? ".dat" : ".txt";
  80
  81         Mat inp, outRef;
  82         ASSERT_NO_THROW( inp = readTorchBlob(_tf(prefix + "_input" + suffix), isBinary) );
  83         ASSERT_NO_THROW( outRef = readTorchBlob(_tf(prefix + "_output" + suffix), isBinary) );
  84
  85         checkBackend(backend, target, &inp, &outRef);
  86
  87         Net net = readNetFromTorch(_tf(prefix + "_net" + suffix), isBinary, evaluate);
  88         ASSERT_FALSE(net.empty());
  89
  90         net.setPreferableBackend(backend);
  91         net.setPreferableTarget(target);
  92
  93         if (outLayerName.empty())
  94             outLayerName = net.getLayerNames().back();
  95
  96         net.setInput(inp);
  97         std::vector<Mat> outBlobs;
  98         net.forward(outBlobs, outLayerName);
  99         l1 = l1 ? l1 : default_l1;
 100         lInf = lInf ? lInf : default_lInf;
 101         normAssert(outRef, outBlobs[0], "", l1, lInf);
 102
 103         if (check2ndBlob && backend != DNN_BACKEND_INFERENCE_ENGINE)
 104         {
 105             Mat out2 = outBlobs[1];
 106             Mat ref2 = readTorchBlob(_tf(prefix + "_output_2" + suffix), isBinary);
 107             normAssert(out2, ref2, "", l1, lInf);
 108         }
 109     }
 110 };
 111
 112 TEST_P(Test_Torch_layers, run_convolution)
 113 {
 114     // Output reference values are in range [23.4018, 72.0181]
 115     double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.08 : default_l1;
 116     double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.42 : default_lInf;
 117     runTorchNet("net_conv", "", false, true, true, l1, lInf);
 118 }
 119
 120 TEST_P(Test_Torch_layers, run_pool_max)
 121 {
 122     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
 123         throw SkipTestException("");
 124     runTorchNet("net_pool_max", "", true);
 125 }
 126
 127 TEST_P(Test_Torch_layers, run_pool_ave)
 128 {
 129     runTorchNet("net_pool_ave");
 130 }
 131
 132 TEST_P(Test_Torch_layers, run_reshape_change_batch_size)
 133 {
 134     runTorchNet("net_reshape");
 135 }
 136
 137 TEST_P(Test_Torch_layers, run_reshape)
 138 {
 139 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE >= 2018040000
 140     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 141         throw SkipTestException("Test is disabled for OpenVINO 2018R4");
 142 #endif
 143     runTorchNet("net_reshape_batch");
 144     runTorchNet("net_reshape_channels", "", false, true);
 145 }
 146
 147 TEST_P(Test_Torch_layers, run_reshape_single_sample)
 148 {
 149     // Reference output values in range [14.4586, 18.4492].
 150     runTorchNet("net_reshape_single_sample", "", false, false, true,
 151                 (target == DNN_TARGET_MYRIAD || target == DNN_TARGET_OPENCL_FP16) ? 0.033 : default_l1,
 152                 (target == DNN_TARGET_MYRIAD || target == DNN_TARGET_OPENCL_FP16) ? 0.05 : default_lInf);
 153 }
 154
 155 TEST_P(Test_Torch_layers, run_linear)
 156 {
 157     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
 158         throw SkipTestException("");
 159     runTorchNet("net_linear_2d");
 160 }
 161
 162 TEST_P(Test_Torch_layers, run_concat)
 163 {
 164     runTorchNet("net_concat", "l5_torchMerge");
 165 }
 166
 167 TEST_P(Test_Torch_layers, run_depth_concat)
 168 {
 169     runTorchNet("net_depth_concat", "", false, true, true, 0.0,
 170                 target == DNN_TARGET_OPENCL_FP16 ? 0.021 : 0.0);
 171 }
 172
 173 TEST_P(Test_Torch_layers, run_deconv)
 174 {
 175 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE >= 2018040000
 176     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 177         throw SkipTestException("Test is disabled for OpenVINO 2018R4");
 178 #endif
 179     runTorchNet("net_deconv");
 180 }
 181
 182 TEST_P(Test_Torch_layers, run_batch_norm)
 183 {
 184     runTorchNet("net_batch_norm", "", false, true);
 185     runTorchNet("net_batch_norm_train", "", false, true, false);
 186 }
 187
 188 TEST_P(Test_Torch_layers, net_prelu)
 189 {
 190     runTorchNet("net_prelu");
 191 }
 192
 193 TEST_P(Test_Torch_layers, net_cadd_table)
 194 {
 195     runTorchNet("net_cadd_table");
 196 }
 197
 198 TEST_P(Test_Torch_layers, net_softmax)
 199 {
 200     runTorchNet("net_softmax");
 201     runTorchNet("net_softmax_spatial");
 202 }
 203
 204 TEST_P(Test_Torch_layers, net_logsoftmax)
 205 {
 206     runTorchNet("net_logsoftmax");
 207     runTorchNet("net_logsoftmax_spatial");
 208 }
 209
 210 TEST_P(Test_Torch_layers, net_lp_pooling)
 211 {
 212     runTorchNet("net_lp_pooling_square", "", false, true);
 213     runTorchNet("net_lp_pooling_power", "", false, true);
 214 }
 215
 216 TEST_P(Test_Torch_layers, net_conv_gemm_lrn)
 217 {
 218     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 219         throw SkipTestException("");
 220     runTorchNet("net_conv_gemm_lrn", "", false, true, true,
 221                 target == DNN_TARGET_OPENCL_FP16 ? 0.046 : 0.0,
 222                 target == DNN_TARGET_OPENCL_FP16 ? 0.023 : 0.0);
 223 }
 224
 225 TEST_P(Test_Torch_layers, net_inception_block)
 226 {
 227 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE == 2018030000
 228     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 229         throw SkipTestException("");
 230 #endif
 231     runTorchNet("net_inception_block", "", false, true);
 232 }
 233
 234 TEST_P(Test_Torch_layers, net_normalize)
 235 {
 236     runTorchNet("net_normalize", "", false, true);
 237 }
 238
 239 TEST_P(Test_Torch_layers, net_padding)
 240 {
 241     runTorchNet("net_padding", "", false, true);
 242     runTorchNet("net_spatial_zero_padding", "", false, true);
 243     runTorchNet("net_spatial_reflection_padding", "", false, true);
 244 }
 245
 246 TEST_P(Test_Torch_layers, net_non_spatial)
 247 {
 248     if (backend == DNN_BACKEND_INFERENCE_ENGINE &&
 249         (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
 250         throw SkipTestException("");
 251     runTorchNet("net_non_spatial", "", false, true);
 252 }
 253
 254 TEST_P(Test_Torch_layers, run_paralel)
 255 {
 256     if (backend != DNN_BACKEND_OPENCV || target != DNN_TARGET_CPU)
 257         throw SkipTestException("");
 258     runTorchNet("net_parallel", "l5_torchMerge");
 259 }
 260
 261 TEST_P(Test_Torch_layers, net_residual)
 262 {
 263 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE == 2018050000
 264     if (backend == DNN_BACKEND_INFERENCE_ENGINE && (target == DNN_TARGET_OPENCL ||
 265                                                     target == DNN_TARGET_OPENCL_FP16))
 266         throw SkipTestException("Test is disabled for OpenVINO 2018R5");
 267 #endif
 268     runTorchNet("net_residual", "", false, true);
 269 }
 270
 271 class Test_Torch_nets : public DNNTestLayer {};
 272
 273 TEST_P(Test_Torch_nets, OpenFace_accuracy)
 274 {
 275 #if defined(INF_ENGINE_RELEASE)
 276     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 277         throw SkipTestException("Test is disabled for Myriad targets");
 278 #endif
 279     checkBackend();
 280
 281     const string model = findDataFile("dnn/openface_nn4.small2.v1.t7", false);
 282     Net net = readNetFromTorch(model);
 283
 284     net.setPreferableBackend(backend);
 285     net.setPreferableTarget(target);
 286
 287     Mat sample = imread(findDataFile("cv/shared/lena.png", false));
 288     Mat sampleF32(sample.size(), CV_32FC3);
 289     sample.convertTo(sampleF32, sampleF32.type());
 290     sampleF32 /= 255;
 291     resize(sampleF32, sampleF32, Size(96, 96), 0, 0, INTER_NEAREST);
 292
 293     Mat inputBlob = blobFromImage(sampleF32, 1.0, Size(), Scalar(), /*swapRB*/true);
 294
 295     net.setInput(inputBlob);
 296     Mat out = net.forward();
 297
 298     // Reference output values are in range [-0.17212, 0.263492]
 299     // on Myriad problem layer: l4_Pooling - does not use pads_begin
 300     float l1 = (target == DNN_TARGET_OPENCL_FP16) ? 4e-4 : 1e-5;
 301     float lInf = (target == DNN_TARGET_OPENCL_FP16) ? 1.5e-3 : 1e-3;
 302     Mat outRef = readTorchBlob(_tf("net_openface_output.dat"), true);
 303     normAssert(out, outRef, "", l1, lInf);
 304 }
 305
 306 static Mat getSegmMask(const Mat& scores)
 307 {
 308     const int rows = scores.size[2];
 309     const int cols = scores.size[3];
 310     const int numClasses = scores.size[1];
 311
 312     Mat maxCl = Mat::zeros(rows, cols, CV_8UC1);
 313     Mat maxVal(rows, cols, CV_32FC1, Scalar(0));
 314     for (int ch = 0; ch < numClasses; ch++)
 315     {
 316         for (int row = 0; row < rows; row++)
 317         {
 318             const float *ptrScore = scores.ptr<float>(0, ch, row);
 319             uint8_t *ptrMaxCl = maxCl.ptr<uint8_t>(row);
 320             float *ptrMaxVal = maxVal.ptr<float>(row);
 321             for (int col = 0; col < cols; col++)
 322             {
 323                 if (ptrScore[col] > ptrMaxVal[col])
 324                 {
 325                     ptrMaxVal[col] = ptrScore[col];
 326                     ptrMaxCl[col] = (uchar)ch;
 327                 }
 328             }
 329         }
 330     }
 331     return maxCl;
 332 }
 333
 334 // Computer per-class intersection over union metric.
 335 static void normAssertSegmentation(const Mat& ref, const Mat& test)
 336 {
 337     CV_Assert_N(ref.dims == 4, test.dims == 4);
 338     const int numClasses = ref.size[1];
 339     CV_Assert(numClasses == test.size[1]);
 340
 341     Mat refMask = getSegmMask(ref);
 342     Mat testMask = getSegmMask(test);
 343     EXPECT_EQ(countNonZero(refMask != testMask), 0);
 344 }
 345
 346 TEST_P(Test_Torch_nets, ENet_accuracy)
 347 {
 348     applyTestTag(target == DNN_TARGET_CPU ? "" : CV_TEST_TAG_MEMORY_512MB);
 349     checkBackend();
 350     if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
 351         (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
 352         throw SkipTestException("");
 353
 354     Net net;
 355     {
 356         const string model = findDataFile("dnn/Enet-model-best.net", false);
 357         net = readNetFromTorch(model, true);
 358         ASSERT_TRUE(!net.empty());
 359     }
 360
 361     net.setPreferableBackend(backend);
 362     net.setPreferableTarget(target);
 363
 364     Mat sample = imread(_tf("street.png", false));
 365     Mat inputBlob = blobFromImage(sample, 1./255, Size(), Scalar(), /*swapRB*/true);
 366
 367     net.setInput(inputBlob, "");
 368     Mat out = net.forward();
 369     Mat ref = blobFromNPY(_tf("torch_enet_prob.npy", false));
 370     // Due to numerical instability in Pooling-Unpooling layers (indexes jittering)
 371     // thresholds for ENet must be changed. Accuracy of results was checked on
 372     // Cityscapes dataset and difference in mIOU with Torch is 10E-4%
 373     normAssert(ref, out, "", 0.00044, /*target == DNN_TARGET_CPU ? 0.453 : */0.552);
 374     normAssertSegmentation(ref, out);
 375
 376     const int N = 3;
 377     for (int i = 0; i < N; i++)
 378     {
 379         net.setInput(inputBlob, "");
 380         Mat out = net.forward();
 381         normAssert(ref, out, "", 0.00044, /*target == DNN_TARGET_CPU ? 0.453 : */0.552);
 382         normAssertSegmentation(ref, out);
 383     }
 384 }
 385
 386 // Check accuracy of style transfer models from https://github.com/jcjohnson/fast-neural-style
 387 // th fast_neural_style.lua \
 388 //   -input_image ~/opencv_extra/testdata/dnn/googlenet_1.png \
 389 //   -output_image lena.png \
 390 //   -median_filter 0 \
 391 //   -image_size 0 \
 392 //   -model models/eccv16/starry_night.t7
 393 // th fast_neural_style.lua \
 394 //   -input_image ~/opencv_extra/testdata/dnn/googlenet_1.png \
 395 //   -output_image lena.png \
 396 //   -median_filter 0 \
 397 //   -image_size 0 \
 398 //   -model models/instance_norm/feathers.t7
 399 TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
 400 {
 401 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LE(2018050000)
 402     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD
 403             && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
 404         throw SkipTestException("Test is disabled for OpenVINO <= 2018R5 + MyriadX target");
 405 #endif
 406
 407     checkBackend();
 408
 409 #if defined(INF_ENGINE_RELEASE)
 410 #if INF_ENGINE_RELEASE <= 2018050000
 411     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL)
 412         throw SkipTestException("");
 413 #endif
 414 #endif
 415
 416     std::string models[] = {"dnn/fast_neural_style_eccv16_starry_night.t7",
 417                             "dnn/fast_neural_style_instance_norm_feathers.t7"};
 418     std::string targets[] = {"dnn/lena_starry_night.png", "dnn/lena_feathers.png"};
 419
 420     for (int i = 0; i < 2; ++i)
 421     {
 422         const string model = findDataFile(models[i], false);
 423         Net net = readNetFromTorch(model);
 424
 425         net.setPreferableBackend(backend);
 426         net.setPreferableTarget(target);
 427
 428         Mat img = imread(findDataFile("dnn/googlenet_1.png", false));
 429         Mat inputBlob = blobFromImage(img, 1.0, Size(), Scalar(103.939, 116.779, 123.68), false);
 430
 431         net.setInput(inputBlob);
 432         Mat out = net.forward();
 433
 434         // Deprocessing.
 435         getPlane(out, 0, 0) += 103.939;
 436         getPlane(out, 0, 1) += 116.779;
 437         getPlane(out, 0, 2) += 123.68;
 438         out = cv::min(cv::max(0, out), 255);
 439
 440         Mat ref = imread(findDataFile(targets[i]));
 441         Mat refBlob = blobFromImage(ref, 1.0, Size(), Scalar(), false);
 442
 443         if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
 444         {
 445             double normL1 = cvtest::norm(refBlob, out, cv::NORM_L1) / refBlob.total();
 446             if (target == DNN_TARGET_MYRIAD)
 447                 EXPECT_LE(normL1, 4.0f);
 448             else
 449                 EXPECT_LE(normL1, 0.6f);
 450         }
 451         else
 452             normAssert(out, refBlob, "", 0.5, 1.1);
 453     }
 454 }
 455
 456 INSTANTIATE_TEST_CASE_P(/**/, Test_Torch_nets, dnnBackendsAndTargets());
 457
 458 // Test a custom layer
 459 // https://github.com/torch/nn/blob/master/doc/convolution.md#nn.SpatialUpSamplingNearest
 460 class SpatialUpSamplingNearestLayer CV_FINAL : public Layer
 461 {
 462 public:
 463     SpatialUpSamplingNearestLayer(const LayerParams &params) : Layer(params)
 464     {
 465         scale = params.get<int>("scale_factor");
 466     }
 467
 468     static Ptr<Layer> create(LayerParams& params)
 469     {
 470         return Ptr<Layer>(new SpatialUpSamplingNearestLayer(params));
 471     }
 472
 473     virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
 474                                  const int requiredOutputs,
 475                                  std::vector<std::vector<int> > &outputs,
 476                                  std::vector<std::vector<int> > &internals) const CV_OVERRIDE
 477     {
 478         std::vector<int> outShape(4);
 479         outShape[0] = inputs[0][0];  // batch size
 480         outShape[1] = inputs[0][1];  // number of channels
 481         outShape[2] = scale * inputs[0][2];
 482         outShape[3] = scale * inputs[0][3];
 483         outputs.assign(1, outShape);
 484         return false;
 485     }
 486
 487     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays) CV_OVERRIDE
 488     {
 489         CV_TRACE_FUNCTION();
 490         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 491
 492         std::vector<Mat> inputs, outputs;
 493         inputs_arr.getMatVector(inputs);
 494         outputs_arr.getMatVector(outputs);
 495
 496         Mat& inp = inputs[0];
 497         Mat& out = outputs[0];
 498         const int outHeight = out.size[2];
 499         const int outWidth = out.size[3];
 500         for (size_t n = 0; n < inp.size[0]; ++n)
 501         {
 502             for (size_t ch = 0; ch < inp.size[1]; ++ch)
 503             {
 504                 resize(getPlane(inp, n, ch), getPlane(out, n, ch),
 505                        Size(outWidth, outHeight), 0, 0, INTER_NEAREST);
 506             }
 507         }
 508     }
 509
 510 private:
 511     int scale;
 512 };
 513
 514 TEST_P(Test_Torch_layers, upsampling_nearest)
 515 {
 516     // Test a custom layer.
 517     CV_DNN_REGISTER_LAYER_CLASS(SpatialUpSamplingNearest, SpatialUpSamplingNearestLayer);
 518     try
 519     {
 520         runTorchNet("net_spatial_upsampling_nearest", "", false, true);
 521     }
 522     catch (...)
 523     {
 524         LayerFactory::unregisterLayer("SpatialUpSamplingNearest");
 525         throw;
 526     }
 527     LayerFactory::unregisterLayer("SpatialUpSamplingNearest");
 528
 529     // Test an implemented layer.
 530     runTorchNet("net_spatial_upsampling_nearest", "", false, true);
 531 }
 532
 533 INSTANTIATE_TEST_CASE_P(/**/, Test_Torch_layers, dnnBackendsAndTargets());
 534
 535 }