inference-engine/tests/unit/engines/mkldnn/graph/layers/extensions/resample_tests.cpp

   1 // Copyright (C) 2018 Intel Corporation
   2 //
   3 // SPDX-License-Identifier: Apache-2.0
   4 //
   5
   6 #include <gtest/gtest.h>
   7 #include <gmock/gmock-spec-builders.h>
   8 #include "mkldnn_plugin/mkldnn_graph.h"
   9
  10 #include "test_graph.hpp"
  11
  12 #include "single_layer_common.hpp"
  13 #include <mkldnn_plugin/mkldnn_extension_utils.h>
  14 #include <extension/ext_list.hpp>
  15 #include "tests_common.hpp"
  16
  17 using namespace ::testing;
  18 using namespace std;
  19 using namespace mkldnn;
  20
  21 struct resample_test_params {
  22     struct {
  23         size_t n;
  24         size_t c;
  25         size_t h;
  26         size_t w;
  27     } in;
  28
  29     float factor;
  30     int antialias;
  31     std::string type;
  32
  33     size_t num_prim_desc;
  34     bool isBlockedFormat;
  35     int selectedType;
  36
  37     std::vector<std::function<void(MKLDNNPlugin::PrimitiveDescInfo)>> comp;
  38 };
  39
  40
  41 static inline float triangleCoeff(float x) {
  42     return std::max(0.0f, 1 - std::abs(x));
  43 }
  44
  45 template <typename data_t>
  46 void ref_resample(const InferenceEngine::TBlob<data_t> &src, InferenceEngine::TBlob<data_t> &dst, resample_test_params prm) {
  47     const data_t *src_data = src.readOnly();
  48     data_t *dst_data = dst.data();
  49
  50     size_t N = prm.in.n;
  51     size_t C = prm.in.c;
  52     size_t IH = prm.in.h;
  53     size_t IW = prm.in.w;
  54     size_t OH = prm.in.h / prm.factor;
  55     size_t OW = prm.in.w / prm.factor;
  56
  57     float fx = static_cast<float>(IW) / static_cast<float>(OW);
  58     float fy = static_cast<float>(IH) / static_cast<float>(OH);
  59
  60     if (prm.type == "caffe.ResampleParameter.NEAREST") {
  61         for (size_t b = 0; b < N; b++) {
  62             for (size_t c = 0; c < C; c++) {
  63                 const float *in_ptr = src_data + IW * IH * C * b + IW * IH * c;
  64                 float *out_ptr = dst_data + OW * OH * C * b + OW * OH * c;
  65
  66                 for (size_t oy = 0; oy < OH; oy++) {
  67                     for (size_t ox = 0; ox < OW; ox++) {
  68                         float ix = ox * fx + fy / 2.0f - 0.5f;
  69                         float iy = oy * fy + fx / 2.0f - 0.5f;
  70
  71                         size_t ix_r = static_cast<size_t>(round(ix));
  72                         size_t iy_r = static_cast<size_t>(round(iy));
  73
  74                         out_ptr[oy * OW + ox] = in_ptr[iy_r * IW + ix_r];
  75                     }
  76                 }
  77             }
  78         }
  79     } else if (prm.type == "caffe.ResampleParameter.LINEAR") {
  80         size_t kernel_width = 2;
  81         bool isDownsample = (fx > 1) || (fy > 1);
  82         bool antialias = isDownsample && prm.antialias;
  83
  84         for (size_t b = 0; b < N; b++) {
  85             for (size_t c = 0; c < C; c++) {
  86                 const float *in_ptr = src_data + IW * IH * C * b + IW * IH * c;
  87                 float *out_ptr = dst_data + OW * OH * C * b + OW * OH * c;
  88
  89                 for (size_t oy = 0; oy < OH; oy++) {
  90                     for (size_t ox = 0; ox < OW; ox++) {
  91                         float ix = ox * fx + fy / 2.0f - 0.5f;
  92                         float iy = oy * fy + fx / 2.0f - 0.5f;
  93
  94                         int ix_r = static_cast<int>(round(ix));
  95                         int iy_r = static_cast<int>(round(iy));
  96
  97                         float sum = 0;
  98                         float wsum = 0;
  99
 100                         float ax = 1.0f / (antialias ? fx : 1.0f);
 101                         float ay = 1.0f / (antialias ? fy : 1.0f);
 102
 103                         int rx = (fx < 1.0f) ? 2 : ceil(static_cast<float>(kernel_width) / ax);
 104                         int ry = (fy < 1.0f) ? 2 : ceil(static_cast<float>(kernel_width) / ay);
 105
 106                         for (int y = iy_r - ry; y <= iy_r + ry; y++) {
 107                             for (int x = ix_r - rx; x <= ix_r + rx; x++) {
 108                                 if (y < 0 || x < 0 || y >= static_cast<int>(IH) || x >= static_cast<int>(IW))
 109                                     continue;
 110
 111                                 float dx = ix - x;
 112                                 float dy = iy - y;
 113
 114                                 float w = ax * triangleCoeff(ax * dx) * ay * triangleCoeff(ay * dy);
 115
 116                                 sum += w * in_ptr[y * IW + x];
 117                                 wsum += w;
 118                             }
 119                         }
 120
 121                         out_ptr[oy * OW + ox] = (!wsum) ? 0 : (sum / wsum);
 122                     }
 123                 }
 124             }
 125         }
 126     } else {
 127         assert(!"Unsupported resample operation type");
 128     }
 129 }
 130
 131 class MKLDNNCPUExtResampleTests: public TestsCommon, public WithParamInterface<resample_test_params> {
 132     std::string model_t = R"V0G0N(
 133 <Net Name="MVN_net" version="2" precision="FP32" batch="1">
 134     <layers>
 135         <layer name="in1" type="Input" precision="FP32" id="0">
 136             <output>
 137                 <port id="0">
 138                     <dim>_IN_</dim>
 139                     <dim>_IC_</dim>
 140                     <dim>_IH_</dim>
 141                     <dim>_IW_</dim>
 142                 </port>
 143             </output>
 144         </layer>
 145         <layer name="fakeLayer" id="1" type="_FL_" precision="FP32">
 146             <input>
 147                 <port id="1">
 148                     <dim>_IN_</dim>
 149                     <dim>_IC_</dim>
 150                     <dim>_IH_</dim>
 151                     <dim>_IW_</dim>
 152                 </port>
 153             </input>
 154             <output>
 155                 <port id="2">
 156                     <dim>_IN_</dim>
 157                     <dim>_IC_</dim>
 158                     <dim>_IH_</dim>
 159                     <dim>_IW_</dim>
 160                 </port>
 161             </output>
 162         </layer>
 163         <layer name="resample" id="2" type="Resample" precision="FP32">
 164             <data antialias="_AN_" factor="_F_" type="_T_"/>
 165             <input>
 166                 <port id="3">
 167                     <dim>_IN_</dim>
 168                     <dim>_IC_</dim>
 169                     <dim>_IH_</dim>
 170                     <dim>_IW_</dim>
 171                 </port>
 172             </input>
 173             <output>
 174                 <port id="4">
 175                     <dim>_IN_</dim>
 176                     <dim>_IC_</dim>
 177                     <dim>_OH_</dim>
 178                     <dim>_OW_</dim>
 179                 </port>
 180             </output>
 181         </layer>
 182     </layers>
 183     <edges>
 184         <edge from-layer="0" from-port="0" to-layer="1" to-port="1"/>
 185         <edge from-layer="1" from-port="2" to-layer="2" to-port="3"/>
 186     </edges>
 187 </Net>
 188 )V0G0N";
 189
 190     std::string getModel(resample_test_params p) {
 191         std::string model = model_t;
 192         if (p.isBlockedFormat)
 193             REPLACE_WITH_STR(model, "_FL_", "FakeLayerBLK");
 194         else
 195             REPLACE_WITH_STR(model, "_FL_", "FakeLayerPLN");
 196
 197         REPLACE_WITH_NUM(model, "_IW_", p.in.w);
 198         REPLACE_WITH_NUM(model, "_IH_", p.in.h);
 199         REPLACE_WITH_NUM(model, "_IC_", p.in.c);
 200         REPLACE_WITH_NUM(model, "_IN_", p.in.n);
 201
 202         REPLACE_WITH_NUM(model, "_OW_", (int)(p.in.w / p.factor));
 203         REPLACE_WITH_NUM(model, "_OH_", (int)(p.in.h / p.factor));
 204
 205         REPLACE_WITH_NUM(model, "_AN_", p.antialias);
 206         REPLACE_WITH_NUM(model, "_F_", p.factor);
 207         REPLACE_WITH_STR(model, "_T_", p.type);
 208
 209         return model;
 210     }
 211
 212 protected:
 213     virtual void TearDown() {
 214     }
 215
 216     virtual void SetUp() {
 217         try {
 218             TestsCommon::SetUp();
 219             resample_test_params p = ::testing::WithParamInterface<resample_test_params>::GetParam();
 220             std::string model = getModel(p);
 221
 222             InferenceEngine::CNNNetReader net_reader;
 223             ASSERT_NO_THROW(net_reader.ReadNetwork(model.data(), model.length()));
 224
 225             std::shared_ptr<InferenceEngine::IExtension> cpuExt(new InferenceEngine::Extensions::Cpu::CpuExtensions());
 226             MKLDNNPlugin::MKLDNNExtensionManager::Ptr extMgr(new MKLDNNPlugin::MKLDNNExtensionManager());
 227             extMgr->AddExtension(cpuExt);
 228
 229             MKLDNNGraphTestClass graph;
 230             graph.CreateGraph(net_reader.getNetwork(), extMgr);
 231
 232             auto& nodes = graph.getNodes();
 233             nodes = graph.getNodes();
 234
 235             for (auto &node : nodes) {
 236                 if (node->getName() == "resample") {
 237                     ASSERT_EQ(p.num_prim_desc, node->getSupportedPrimitiveDescriptors().size());
 238                     for (size_t j = 0; j < p.num_prim_desc && j < p.comp.size(); j++) {
 239                         p.comp.at(j)(node->getSupportedPrimitiveDescriptors().at(j));
 240                     }
 241                     ASSERT_NE(nullptr, node->getSelectedPrimitiveDescriptor());
 242                     ASSERT_EQ(p.selectedType,
 243                               node->getSelectedPrimitiveDescriptor()->getImplementationType() & p.selectedType);
 244                 }
 245             }
 246
 247             if (p.isBlockedFormat)
 248                 ASSERT_EQ(6, nodes.size());
 249             else
 250                 ASSERT_EQ(4, nodes.size());
 251
 252             InferenceEngine::SizeVector dims_src = {p.in.w, p.in.h, p.in.c, p.in.n};
 253
 254             InferenceEngine::Blob::Ptr src = InferenceEngine::make_shared_blob<float, const InferenceEngine::SizeVector>(InferenceEngine::Precision::FP32, InferenceEngine::NHWC, dims_src);
 255             src->allocate();
 256             fill_data(src->buffer(), src->size());
 257
 258             auto * srcPtr = dynamic_cast<InferenceEngine::TBlob<float>*>(src.get());
 259
 260             if (srcPtr == nullptr)
 261                 FAIL() << "Cannot cast blob to TBlob<float>.";
 262
 263             InferenceEngine::BlobMap srcs;
 264             srcs.insert(std::pair<std::string, InferenceEngine::Blob::Ptr>("in1", src));
 265
 266             InferenceEngine::OutputsDataMap out;
 267             out = net_reader.getNetwork().getOutputsInfo();
 268             InferenceEngine::BlobMap outputBlobs;
 269
 270             std::pair<std::string, InferenceEngine::DataPtr> item = *out.begin();
 271
 272             InferenceEngine::TBlob<float>::Ptr output;
 273             output = InferenceEngine::make_shared_blob<float>(item.second->getTensorDesc());
 274             output->allocate();
 275             outputBlobs[item.first] = output;
 276
 277             graph.Infer(srcs, outputBlobs);
 278
 279             InferenceEngine::TBlob<float> dst_ref(item.second->getTensorDesc());
 280             dst_ref.allocate();
 281             ref_resample(*srcPtr, dst_ref, p);
 282             compare(*output, dst_ref);
 283         } catch (const InferenceEngine::details::InferenceEngineException &e) {
 284             FAIL() << e.what();
 285         }
 286     }
 287 };
 288
 289 TEST_P(MKLDNNCPUExtResampleTests, TestsResample) {}
 290
 291 INSTANTIATE_TEST_CASE_P(
 292         TestsResample, MKLDNNCPUExtResampleTests,
 293         ::testing::Values(
 294                 resample_test_params{{2, 64, 15, 25}, 1.f, 0, "caffe.ResampleParameter.NEAREST", 2, false, MKLDNNPlugin::impl_desc_type::unknown },
 295                 resample_test_params{{2, 64, 15, 25}, 1.f, 0, "caffe.ResampleParameter.NEAREST", 2, true, MKLDNNPlugin::impl_desc_type::unknown },
 296                 resample_test_params{{2, 64, 15, 25}, 1.f, 1, "caffe.ResampleParameter.LINEAR", 1, false, MKLDNNPlugin::impl_desc_type::unknown },
 297                 resample_test_params{{2, 64, 10, 20}, 0.25f, 0, "caffe.ResampleParameter.NEAREST", 2, false, MKLDNNPlugin::impl_desc_type::unknown },
 298                 resample_test_params{{2, 64, 10, 20}, 0.25f, 0, "caffe.ResampleParameter.NEAREST", 2, true, MKLDNNPlugin::impl_desc_type::unknown },
 299                 resample_test_params{{2, 64, 10, 20}, 0.25f, 1, "caffe.ResampleParameter.LINEAR", 1, false, MKLDNNPlugin::impl_desc_type::unknown },
 300                 resample_test_params{{2, 64, 10, 20}, 4.f, 0, "caffe.ResampleParameter.NEAREST", 2, false, MKLDNNPlugin::impl_desc_type::unknown },
 301                 resample_test_params{{2, 64, 10, 20}, 4.f, 0, "caffe.ResampleParameter.NEAREST", 2, true, MKLDNNPlugin::impl_desc_type::unknown },
 302                 resample_test_params{{2, 64, 10, 20}, 4.f, 1, "caffe.ResampleParameter.LINEAR", 1, false, MKLDNNPlugin::impl_desc_type::unknown },
 303                 resample_test_params{{2, 3, 15, 25}, 1.f, 0, "caffe.ResampleParameter.NEAREST", 2, false, MKLDNNPlugin::impl_desc_type::unknown },
 304                 resample_test_params{{2, 3, 15, 25}, 1.f, 0, "caffe.ResampleParameter.NEAREST", 2, true, MKLDNNPlugin::impl_desc_type::unknown },
 305                 resample_test_params{{2, 3, 15, 25}, 1.f, 1, "caffe.ResampleParameter.LINEAR", 1, false, MKLDNNPlugin::impl_desc_type::unknown },
 306                 resample_test_params{{2, 3, 10, 20}, 0.25f, 0, "caffe.ResampleParameter.NEAREST", 2, false, MKLDNNPlugin::impl_desc_type::unknown },
 307                 resample_test_params{{2, 3, 10, 20}, 0.25f, 0, "caffe.ResampleParameter.NEAREST", 2, true, MKLDNNPlugin::impl_desc_type::unknown },
 308                 resample_test_params{{2, 3, 10, 20}, 0.25f, 1, "caffe.ResampleParameter.LINEAR", 1, false, MKLDNNPlugin::impl_desc_type::unknown },
 309                 resample_test_params{{2, 3, 10, 20}, 4.f, 0, "caffe.ResampleParameter.NEAREST", 2, false, MKLDNNPlugin::impl_desc_type::unknown },
 310                 resample_test_params{{2, 3, 10, 20}, 4.f, 0, "caffe.ResampleParameter.NEAREST", 2, true, MKLDNNPlugin::impl_desc_type::unknown },
 311                 resample_test_params{{2, 3, 10, 20}, 4.f, 1, "caffe.ResampleParameter.LINEAR", 1, false, MKLDNNPlugin::impl_desc_type::unknown }));