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     runTorchNet("net_deconv");
 176 }
 177
 178 TEST_P(Test_Torch_layers, run_batch_norm)
 179 {
 180     runTorchNet("net_batch_norm", "", false, true);
 181     runTorchNet("net_batch_norm_train", "", false, true, false);
 182 }
 183
 184 TEST_P(Test_Torch_layers, net_prelu)
 185 {
 186     runTorchNet("net_prelu");
 187 }
 188
 189 TEST_P(Test_Torch_layers, net_cadd_table)
 190 {
 191     runTorchNet("net_cadd_table");
 192 }
 193
 194 TEST_P(Test_Torch_layers, net_softmax)
 195 {
 196     runTorchNet("net_softmax");
 197     runTorchNet("net_softmax_spatial");
 198 }
 199
 200 TEST_P(Test_Torch_layers, net_logsoftmax)
 201 {
 202     runTorchNet("net_logsoftmax");
 203     runTorchNet("net_logsoftmax_spatial");
 204 }
 205
 206 TEST_P(Test_Torch_layers, net_lp_pooling)
 207 {
 208     runTorchNet("net_lp_pooling_square", "", false, true);
 209     runTorchNet("net_lp_pooling_power", "", false, true);
 210 }
 211
 212 TEST_P(Test_Torch_layers, net_conv_gemm_lrn)
 213 {
 214     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 215         throw SkipTestException("");
 216     runTorchNet("net_conv_gemm_lrn", "", false, true, true,
 217                 target == DNN_TARGET_OPENCL_FP16 ? 0.046 : 0.0,
 218                 target == DNN_TARGET_OPENCL_FP16 ? 0.023 : 0.0);
 219 }
 220
 221 TEST_P(Test_Torch_layers, net_inception_block)
 222 {
 223 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE == 2018030000
 224     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 225         throw SkipTestException("");
 226 #endif
 227     runTorchNet("net_inception_block", "", false, true);
 228 }
 229
 230 TEST_P(Test_Torch_layers, net_normalize)
 231 {
 232     runTorchNet("net_normalize", "", false, true);
 233 }
 234
 235 TEST_P(Test_Torch_layers, net_padding)
 236 {
 237     runTorchNet("net_padding", "", false, true);
 238     runTorchNet("net_spatial_zero_padding", "", false, true);
 239     runTorchNet("net_spatial_reflection_padding", "", false, true);
 240 }
 241
 242 TEST_P(Test_Torch_layers, net_non_spatial)
 243 {
 244     if (backend == DNN_BACKEND_INFERENCE_ENGINE &&
 245         (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
 246         throw SkipTestException("");
 247     runTorchNet("net_non_spatial", "", false, true);
 248 }
 249
 250 TEST_P(Test_Torch_layers, run_paralel)
 251 {
 252     if (backend != DNN_BACKEND_OPENCV || target != DNN_TARGET_CPU)
 253         throw SkipTestException("");
 254     runTorchNet("net_parallel", "l5_torchMerge");
 255 }
 256
 257 TEST_P(Test_Torch_layers, net_residual)
 258 {
 259 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE == 2018050000
 260     if (backend == DNN_BACKEND_INFERENCE_ENGINE && (target == DNN_TARGET_OPENCL ||
 261                                                     target == DNN_TARGET_OPENCL_FP16))
 262         throw SkipTestException("Test is disabled for OpenVINO 2018R5");
 263 #endif
 264     runTorchNet("net_residual", "", false, true);
 265 }
 266
 267 class Test_Torch_nets : public DNNTestLayer {};
 268
 269 TEST_P(Test_Torch_nets, OpenFace_accuracy)
 270 {
 271 #if defined(INF_ENGINE_RELEASE)
 272     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
 273         throw SkipTestException("Test is disabled for Myriad targets");
 274 #endif
 275     checkBackend();
 276
 277     const string model = findDataFile("dnn/openface_nn4.small2.v1.t7", false);
 278     Net net = readNetFromTorch(model);
 279
 280     net.setPreferableBackend(backend);
 281     net.setPreferableTarget(target);
 282
 283     Mat sample = imread(findDataFile("cv/shared/lena.png", false));
 284     Mat sampleF32(sample.size(), CV_32FC3);
 285     sample.convertTo(sampleF32, sampleF32.type());
 286     sampleF32 /= 255;
 287     resize(sampleF32, sampleF32, Size(96, 96), 0, 0, INTER_NEAREST);
 288
 289     Mat inputBlob = blobFromImage(sampleF32, 1.0, Size(), Scalar(), /*swapRB*/true);
 290
 291     net.setInput(inputBlob);
 292     Mat out = net.forward();
 293
 294     // Reference output values are in range [-0.17212, 0.263492]
 295     // on Myriad problem layer: l4_Pooling - does not use pads_begin
 296     float l1 = (target == DNN_TARGET_OPENCL_FP16) ? 4e-4 : 1e-5;
 297     float lInf = (target == DNN_TARGET_OPENCL_FP16) ? 1.5e-3 : 1e-3;
 298     Mat outRef = readTorchBlob(_tf("net_openface_output.dat"), true);
 299     normAssert(out, outRef, "", l1, lInf);
 300 }
 301
 302 static Mat getSegmMask(const Mat& scores)
 303 {
 304     const int rows = scores.size[2];
 305     const int cols = scores.size[3];
 306     const int numClasses = scores.size[1];
 307
 308     Mat maxCl = Mat::zeros(rows, cols, CV_8UC1);
 309     Mat maxVal(rows, cols, CV_32FC1, Scalar(0));
 310     for (int ch = 0; ch < numClasses; ch++)
 311     {
 312         for (int row = 0; row < rows; row++)
 313         {
 314             const float *ptrScore = scores.ptr<float>(0, ch, row);
 315             uint8_t *ptrMaxCl = maxCl.ptr<uint8_t>(row);
 316             float *ptrMaxVal = maxVal.ptr<float>(row);
 317             for (int col = 0; col < cols; col++)
 318             {
 319                 if (ptrScore[col] > ptrMaxVal[col])
 320                 {
 321                     ptrMaxVal[col] = ptrScore[col];
 322                     ptrMaxCl[col] = (uchar)ch;
 323                 }
 324             }
 325         }
 326     }
 327     return maxCl;
 328 }
 329
 330 // Computer per-class intersection over union metric.
 331 static void normAssertSegmentation(const Mat& ref, const Mat& test)
 332 {
 333     CV_Assert_N(ref.dims == 4, test.dims == 4);
 334     const int numClasses = ref.size[1];
 335     CV_Assert(numClasses == test.size[1]);
 336
 337     Mat refMask = getSegmMask(ref);
 338     Mat testMask = getSegmMask(test);
 339     EXPECT_EQ(countNonZero(refMask != testMask), 0);
 340 }
 341
 342 TEST_P(Test_Torch_nets, ENet_accuracy)
 343 {
 344     applyTestTag(target == DNN_TARGET_CPU ? "" : CV_TEST_TAG_MEMORY_512MB);
 345     checkBackend();
 346     if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
 347         (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
 348         throw SkipTestException("");
 349
 350     Net net;
 351     {
 352         const string model = findDataFile("dnn/Enet-model-best.net", false);
 353         net = readNetFromTorch(model, true);
 354         ASSERT_TRUE(!net.empty());
 355     }
 356
 357     net.setPreferableBackend(backend);
 358     net.setPreferableTarget(target);
 359
 360     Mat sample = imread(_tf("street.png", false));
 361     Mat inputBlob = blobFromImage(sample, 1./255, Size(), Scalar(), /*swapRB*/true);
 362
 363     net.setInput(inputBlob, "");
 364     Mat out = net.forward();
 365     Mat ref = blobFromNPY(_tf("torch_enet_prob.npy", false));
 366     // Due to numerical instability in Pooling-Unpooling layers (indexes jittering)
 367     // thresholds for ENet must be changed. Accuracy of results was checked on
 368     // Cityscapes dataset and difference in mIOU with Torch is 10E-4%
 369     normAssert(ref, out, "", 0.00044, /*target == DNN_TARGET_CPU ? 0.453 : */0.552);
 370     normAssertSegmentation(ref, out);
 371
 372     const int N = 3;
 373     for (int i = 0; i < N; i++)
 374     {
 375         net.setInput(inputBlob, "");
 376         Mat out = net.forward();
 377         normAssert(ref, out, "", 0.00044, /*target == DNN_TARGET_CPU ? 0.453 : */0.552);
 378         normAssertSegmentation(ref, out);
 379     }
 380 }
 381
 382 // Check accuracy of style transfer models from https://github.com/jcjohnson/fast-neural-style
 383 // th fast_neural_style.lua \
 384 //   -input_image ~/opencv_extra/testdata/dnn/googlenet_1.png \
 385 //   -output_image lena.png \
 386 //   -median_filter 0 \
 387 //   -image_size 0 \
 388 //   -model models/eccv16/starry_night.t7
 389 // th fast_neural_style.lua \
 390 //   -input_image ~/opencv_extra/testdata/dnn/googlenet_1.png \
 391 //   -output_image lena.png \
 392 //   -median_filter 0 \
 393 //   -image_size 0 \
 394 //   -model models/instance_norm/feathers.t7
 395 TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
 396 {
 397 #if defined INF_ENGINE_RELEASE
 398     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD
 399             && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
 400         throw SkipTestException("Test is disabled for MyriadX target");
 401 #endif
 402
 403     checkBackend();
 404
 405 #if defined(INF_ENGINE_RELEASE)
 406 #if INF_ENGINE_RELEASE <= 2018050000
 407     if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL)
 408         throw SkipTestException("");
 409 #endif
 410 #endif
 411
 412     std::string models[] = {"dnn/fast_neural_style_eccv16_starry_night.t7",
 413                             "dnn/fast_neural_style_instance_norm_feathers.t7"};
 414     std::string targets[] = {"dnn/lena_starry_night.png", "dnn/lena_feathers.png"};
 415
 416     for (int i = 0; i < 2; ++i)
 417     {
 418         const string model = findDataFile(models[i], false);
 419         Net net = readNetFromTorch(model);
 420
 421         net.setPreferableBackend(backend);
 422         net.setPreferableTarget(target);
 423
 424         Mat img = imread(findDataFile("dnn/googlenet_1.png", false));
 425         Mat inputBlob = blobFromImage(img, 1.0, Size(), Scalar(103.939, 116.779, 123.68), false);
 426
 427         net.setInput(inputBlob);
 428         Mat out = net.forward();
 429
 430         // Deprocessing.
 431         getPlane(out, 0, 0) += 103.939;
 432         getPlane(out, 0, 1) += 116.779;
 433         getPlane(out, 0, 2) += 123.68;
 434         out = cv::min(cv::max(0, out), 255);
 435
 436         Mat ref = imread(findDataFile(targets[i]));
 437         Mat refBlob = blobFromImage(ref, 1.0, Size(), Scalar(), false);
 438
 439         if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
 440         {
 441             double normL1 = cvtest::norm(refBlob, out, cv::NORM_L1) / refBlob.total();
 442             if (target == DNN_TARGET_MYRIAD)
 443                 EXPECT_LE(normL1, 4.0f);
 444             else
 445                 EXPECT_LE(normL1, 0.6f);
 446         }
 447         else
 448             normAssert(out, refBlob, "", 0.5, 1.1);
 449     }
 450 }
 451
 452 INSTANTIATE_TEST_CASE_P(/**/, Test_Torch_nets, dnnBackendsAndTargets());
 453
 454 // Test a custom layer
 455 // https://github.com/torch/nn/blob/master/doc/convolution.md#nn.SpatialUpSamplingNearest
 456 class SpatialUpSamplingNearestLayer CV_FINAL : public Layer
 457 {
 458 public:
 459     SpatialUpSamplingNearestLayer(const LayerParams &params) : Layer(params)
 460     {
 461         scale = params.get<int>("scale_factor");
 462     }
 463
 464     static Ptr<Layer> create(LayerParams& params)
 465     {
 466         return Ptr<Layer>(new SpatialUpSamplingNearestLayer(params));
 467     }
 468
 469     virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
 470                                  const int requiredOutputs,
 471                                  std::vector<std::vector<int> > &outputs,
 472                                  std::vector<std::vector<int> > &internals) const CV_OVERRIDE
 473     {
 474         std::vector<int> outShape(4);
 475         outShape[0] = inputs[0][0];  // batch size
 476         outShape[1] = inputs[0][1];  // number of channels
 477         outShape[2] = scale * inputs[0][2];
 478         outShape[3] = scale * inputs[0][3];
 479         outputs.assign(1, outShape);
 480         return false;
 481     }
 482
 483     void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays) CV_OVERRIDE
 484     {
 485         CV_TRACE_FUNCTION();
 486         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 487
 488         std::vector<Mat> inputs, outputs;
 489         inputs_arr.getMatVector(inputs);
 490         outputs_arr.getMatVector(outputs);
 491
 492         Mat& inp = inputs[0];
 493         Mat& out = outputs[0];
 494         const int outHeight = out.size[2];
 495         const int outWidth = out.size[3];
 496         for (size_t n = 0; n < inp.size[0]; ++n)
 497         {
 498             for (size_t ch = 0; ch < inp.size[1]; ++ch)
 499             {
 500                 resize(getPlane(inp, n, ch), getPlane(out, n, ch),
 501                        Size(outWidth, outHeight), 0, 0, INTER_NEAREST);
 502             }
 503         }
 504     }
 505
 506 private:
 507     int scale;
 508 };
 509
 510 TEST_P(Test_Torch_layers, upsampling_nearest)
 511 {
 512     // Test a custom layer.
 513     CV_DNN_REGISTER_LAYER_CLASS(SpatialUpSamplingNearest, SpatialUpSamplingNearestLayer);
 514     try
 515     {
 516         runTorchNet("net_spatial_upsampling_nearest", "", false, true);
 517     }
 518     catch (...)
 519     {
 520         LayerFactory::unregisterLayer("SpatialUpSamplingNearest");
 521         throw;
 522     }
 523     LayerFactory::unregisterLayer("SpatialUpSamplingNearest");
 524
 525     // Test an implemented layer.
 526     runTorchNet("net_spatial_upsampling_nearest", "", false, true);
 527 }
 528
 529 INSTANTIATE_TEST_CASE_P(/**/, Test_Torch_layers, dnnBackendsAndTargets());
 530
 531 }