modules/dnn/src/onnx/onnx_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) 2018, Intel Corporation, all rights reserved.
   6 // Third party copyrights are property of their respective owners.
   7
   8 #include "../precomp.hpp"
   9 #include <opencv2/dnn/shape_utils.hpp>
  10
  11 #ifdef HAVE_PROTOBUF
  12
  13 #include <iostream>
  14 #include <fstream>
  15 #include <string>
  16 #include <limits>
  17 #include <algorithm>
  18
  19
  20 #if defined(__GNUC__) && __GNUC__ >= 5
  21 #pragma GCC diagnostic push
  22 #pragma GCC diagnostic ignored "-Wsuggest-override"
  23 #endif
  24 #include "opencv-onnx.pb.h"
  25 #if defined(__GNUC__) && __GNUC__ >= 5
  26 #pragma GCC diagnostic pop
  27 #endif
  28
  29 #include "onnx_graph_simplifier.hpp"
  30
  31 namespace cv {
  32 namespace dnn {
  33 CV__DNN_EXPERIMENTAL_NS_BEGIN
  34
  35
  36 class ONNXImporter
  37 {
  38     opencv_onnx::ModelProto model_proto;
  39     struct LayerInfo {
  40         int layerId;
  41         int outputId;
  42         LayerInfo(int _layerId, int _outputId) : layerId(_layerId), outputId(_outputId) {}
  43     };
  44
  45     std::map<std::string, Mat> getGraphTensors(
  46                                     const opencv_onnx::GraphProto& graph_proto);
  47     Mat getBlob(const opencv_onnx::NodeProto& node_proto, const std::map<std::string, Mat>& constBlobs, int index);
  48
  49     LayerParams getLayerParams(const opencv_onnx::NodeProto& node_proto);
  50     bool isCeilMode(const LayerParams& layerParams);
  51
  52 public:
  53
  54     ONNXImporter(const char *onnxFile)
  55     {
  56         std::fstream input(onnxFile, std::ios::in | std::ios::binary);
  57
  58         if (!model_proto.ParseFromIstream(&input))
  59             CV_Error(Error::StsUnsupportedFormat, "Failed to parse onnx model");
  60     }
  61
  62     ONNXImporter(const char* buffer, size_t sizeBuffer)
  63     {
  64         struct _Buf : public std::streambuf
  65         {
  66             _Buf(const char* buffer, size_t sizeBuffer)
  67             {
  68                 char* p = const_cast<char*>(buffer);
  69                 setg(p, p, p + sizeBuffer);
  70             }
  71         };
  72
  73         _Buf buf(buffer, sizeBuffer);
  74         std::istream input(&buf);
  75
  76         if (!model_proto.ParseFromIstream(&input))
  77             CV_Error(Error::StsUnsupportedFormat, "Failed to parse onnx model from in-memory byte array.");
  78     }
  79
  80     void populateNet(Net dstNet);
  81 };
  82
  83 inline void replaceLayerParam(LayerParams& layerParams, const String& oldKey, const String& newKey)
  84 {
  85     if (layerParams.has(oldKey)) {
  86         layerParams.set(newKey, layerParams.get(oldKey));
  87         layerParams.erase(oldKey);
  88     }
  89 }
  90
  91 void releaseONNXTensor(opencv_onnx::TensorProto& tensor_proto)
  92 {
  93     if (!tensor_proto.raw_data().empty()) {
  94         delete tensor_proto.release_raw_data();
  95     }
  96 }
  97
  98 void runLayer(LayerParams& params, const std::vector<Mat>& inputs,
  99               std::vector<Mat>& outputs)
 100 {
 101     Ptr<Layer> layer = LayerFactory::createLayerInstance(params.type, params);
 102     CV_Assert((bool)layer);
 103
 104     std::vector<MatShape> inpShapes(inputs.size());
 105     int ddepth = CV_32F;
 106     for (size_t i = 0; i < inputs.size(); ++i)
 107     {
 108         inpShapes[i] = shape(inputs[i]);
 109         if (i > 0 && ddepth != inputs[i].depth())
 110             CV_Error(Error::StsNotImplemented, "Mixed input data types.");
 111         ddepth = inputs[i].depth();
 112     }
 113
 114     std::vector<MatShape> outShapes, internalShapes;
 115     layer->getMemoryShapes(inpShapes, 0, outShapes, internalShapes);
 116
 117     std::vector<Mat> internals(internalShapes.size());
 118     outputs.resize(outShapes.size());
 119     for (size_t i = 0; i < outShapes.size(); ++i)
 120         outputs[i].create(outShapes[i], ddepth);
 121     for (size_t i = 0; i < internalShapes.size(); ++i)
 122         internals[i].create(internalShapes[i], ddepth);
 123
 124     layer->finalize(inputs, outputs);
 125     layer->forward(inputs, outputs, internals);
 126 }
 127
 128 std::map<std::string, Mat> ONNXImporter::getGraphTensors(
 129                                         const opencv_onnx::GraphProto& graph_proto)
 130 {
 131   opencv_onnx::TensorProto tensor_proto;
 132   std::map<std::string, Mat> layers_weights;
 133
 134   for (int i = 0; i < graph_proto.initializer_size(); i++)
 135   {
 136     tensor_proto = graph_proto.initializer(i);
 137     Mat mat = getMatFromTensor(tensor_proto);
 138     releaseONNXTensor(tensor_proto);
 139     layers_weights.insert(std::make_pair(tensor_proto.name(), mat));
 140   }
 141   return layers_weights;
 142 }
 143
 144 static DictValue parse(const ::google::protobuf::RepeatedField< ::google::protobuf::int64>& src) {
 145     std::vector<int32_t> dst(src.size());
 146     convertInt64ToInt32(src, dst, src.size());
 147     return DictValue::arrayInt(&dst[0], src.size());
 148 }
 149
 150 LayerParams ONNXImporter::getLayerParams(const opencv_onnx::NodeProto& node_proto)
 151 {
 152     LayerParams lp;
 153     for(int i = 0; i < node_proto.attribute_size(); i++)
 154     {
 155         opencv_onnx::AttributeProto attribute_proto = node_proto.attribute(i);
 156         std::string attribute_name = attribute_proto.name();
 157
 158         if(attribute_name == "kernel_shape")
 159         {
 160             CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
 161             lp.set("kernel_size", parse(attribute_proto.ints()));
 162         }
 163         else if(attribute_name == "strides")
 164         {
 165             CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
 166             lp.set("stride", parse(attribute_proto.ints()));
 167         }
 168         else if(attribute_name == "pads")
 169         {
 170             if (node_proto.op_type() == "Pad")
 171             {
 172                 // Padding layer.
 173                 // Paddings are in order begin0, begin1, .. beginN, end0, end1, ..., endN.
 174                 // We need to shuffle it to begin0, end0, begin1, end1, ...
 175                 CV_Assert(attribute_proto.ints_size() % 2 == 0);
 176                 const int dims = attribute_proto.ints_size() / 2;
 177                 std::vector<int32_t> paddings;
 178                 paddings.reserve(attribute_proto.ints_size());
 179                 for (int i = 0; i < dims; ++i)
 180                 {
 181                     paddings.push_back(attribute_proto.ints(i));
 182                     paddings.push_back(attribute_proto.ints(dims + i));
 183                 }
 184                 lp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
 185             }
 186             else
 187             {
 188                 // Convolution or pooling.
 189                 CV_Assert(attribute_proto.ints_size() == 4 || attribute_proto.ints_size() == 6);
 190                 lp.set("pad", parse(attribute_proto.ints()));
 191             }
 192         }
 193         else if(attribute_name == "auto_pad")
 194         {
 195             if (attribute_proto.s() == "SAME_UPPER" || attribute_proto.s() == "SAME_LOWER") {
 196                 lp.set("pad_mode",  "SAME");
 197             }
 198             else if (attribute_proto.s() == "VALID") {
 199                 lp.set("pad_mode", "VALID");
 200             }
 201         }
 202         else if(attribute_name == "dilations")
 203         {
 204             CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
 205             lp.set("dilation", parse(attribute_proto.ints()));
 206         }
 207         else if (attribute_proto.has_i())
 208         {
 209             ::google::protobuf::int64 src = attribute_proto.i();
 210             if (src < std::numeric_limits<int32_t>::min() || src > std::numeric_limits<int32_t>::max())
 211                 CV_Error(Error::StsOutOfRange, "Input is out of OpenCV 32S range");
 212             else
 213                 lp.set(attribute_name, saturate_cast<int32_t>(src));
 214         }
 215         else if (attribute_proto.has_f())
 216         {
 217             lp.set(attribute_name, attribute_proto.f());
 218         }
 219         else if (attribute_proto.has_s())
 220         {
 221             lp.set(attribute_name, attribute_proto.s());
 222         }
 223         else if (attribute_proto.floats_size() > 0)
 224         {
 225             lp.set(attribute_name, DictValue::arrayReal(
 226                 attribute_proto.floats().data(), attribute_proto.floats_size()));
 227         }
 228         else if (attribute_proto.ints_size() > 0)
 229         {
 230             lp.set(attribute_proto.name(), parse(attribute_proto.ints()));
 231         }
 232         else if (attribute_proto.has_t())
 233         {
 234             opencv_onnx::TensorProto tensor = attribute_proto.t();
 235             Mat blob = getMatFromTensor(tensor);
 236             lp.blobs.push_back(blob);
 237         }
 238         else if (attribute_proto.has_g() || attribute_proto.strings_size() > 0 ||
 239                     attribute_proto.tensors_size() > 0 || attribute_proto.graphs_size() > 0)
 240         {
 241                 CV_Error(Error::StsNotImplemented, "Unexpected attribute type");
 242         }
 243         else
 244             CV_Error(Error::StsNotImplemented, "Unsupported attribute type");
 245     }
 246     return lp;
 247 }
 248
 249 Mat ONNXImporter::getBlob(const opencv_onnx::NodeProto& node_proto,
 250                     const std::map<std::string, Mat>& constBlobs, int index)
 251 {
 252     CV_Assert(index < node_proto.input_size());
 253     std::map<std::string, Mat>::const_iterator constBlob;
 254     constBlob = constBlobs.find(node_proto.input(index));
 255     if (constBlob == constBlobs.end()) {
 256         CV_Error(Error::StsObjectNotFound,
 257              "Blob " + node_proto.input(index) + " not found in const blobs");
 258     }
 259     return constBlob->second;
 260 }
 261
 262 void ONNXImporter::populateNet(Net dstNet)
 263 {
 264     CV_Assert(model_proto.has_graph());
 265     opencv_onnx::GraphProto graph_proto = model_proto.graph();
 266
 267     simplifySubgraphs(graph_proto);
 268
 269     std::map<std::string, Mat> constBlobs = getGraphTensors(graph_proto);
 270     // List of internal blobs shapes.
 271     std::map<std::string, MatShape> outShapes;
 272     // Add all the inputs shapes. It includes as constant blobs as network's inputs shapes.
 273     for (int i = 0; i < graph_proto.input_size(); ++i)
 274     {
 275         opencv_onnx::ValueInfoProto valueInfoProto = graph_proto.input(i);
 276         CV_Assert(valueInfoProto.has_type());
 277         opencv_onnx::TypeProto typeProto = valueInfoProto.type();
 278         CV_Assert(typeProto.has_tensor_type());
 279         opencv_onnx::TypeProto::Tensor tensor = typeProto.tensor_type();
 280         CV_Assert(tensor.has_shape());
 281         opencv_onnx::TensorShapeProto tensorShape = tensor.shape();
 282
 283         MatShape inpShape(tensorShape.dim_size());
 284         for (int j = 0; j < inpShape.size(); ++j)
 285         {
 286             inpShape[j] = tensorShape.dim(j).dim_value();
 287         }
 288         outShapes[valueInfoProto.name()] = inpShape;
 289     }
 290
 291     std::string framework_name;
 292     if (model_proto.has_producer_name()) {
 293         framework_name = model_proto.producer_name();
 294     }
 295
 296     // create map with network inputs (without const blobs)
 297     std::map<std::string, LayerInfo> layer_id;
 298     std::map<std::string, LayerInfo>::iterator layerId;
 299     std::map<std::string, MatShape>::iterator shapeIt;
 300     // fill map: push layer name, layer id and output id
 301     std::vector<String> netInputs;
 302     for (int j = 0; j < graph_proto.input_size(); j++)
 303     {
 304         const std::string& name = graph_proto.input(j).name();
 305         if (constBlobs.find(name) == constBlobs.end()) {
 306             netInputs.push_back(name);
 307             layer_id.insert(std::make_pair(name, LayerInfo(0, netInputs.size() - 1)));
 308         }
 309     }
 310     dstNet.setInputsNames(netInputs);
 311
 312     int layersSize = graph_proto.node_size();
 313     LayerParams layerParams;
 314     opencv_onnx::NodeProto node_proto;
 315
 316     for(int li = 0; li < layersSize; li++)
 317     {
 318         node_proto = graph_proto.node(li);
 319         layerParams = getLayerParams(node_proto);
 320         CV_Assert(node_proto.output_size() >= 1);
 321         layerParams.name = node_proto.output(0);
 322
 323         std::string layer_type = node_proto.op_type();
 324         layerParams.type = layer_type;
 325
 326
 327         if (layer_type == "MaxPool")
 328         {
 329             layerParams.type = "Pooling";
 330             layerParams.set("pool", "MAX");
 331             layerParams.set("ceil_mode", layerParams.has("pad_mode"));
 332         }
 333         else if (layer_type == "AveragePool")
 334         {
 335             layerParams.type = "Pooling";
 336             layerParams.set("pool", "AVE");
 337             layerParams.set("ceil_mode", layerParams.has("pad_mode"));
 338             layerParams.set("ave_pool_padded_area", framework_name == "pytorch");
 339         }
 340         else if (layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool" || layer_type == "ReduceMean")
 341         {
 342             CV_Assert(node_proto.input_size() == 1);
 343             layerParams.type = "Pooling";
 344             layerParams.set("pool", layer_type == "GlobalMaxPool"? "MAX" : "AVE");
 345             layerParams.set("global_pooling", layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool");
 346
 347             if (layer_type == "ReduceMean")
 348             {
 349                 if (layerParams.get<int>("keepdims") == 0 || !layerParams.has("axes"))
 350                     CV_Error(Error::StsNotImplemented, "Unsupported mode of ReduceMean operation.");
 351
 352                 MatShape inpShape = outShapes[node_proto.input(0)];
 353                 if (inpShape.size() != 4 && inpShape.size() != 5)
 354                     CV_Error(Error::StsNotImplemented, "Unsupported input shape of reduce_mean operation.");
 355
 356                 DictValue axes = layerParams.get("axes");
 357                 CV_Assert(axes.size() <= inpShape.size() - 2);
 358                 std::vector<int> kernel_size(inpShape.size() - 2, 1);
 359                 for (int i = 0; i < axes.size(); i++) {
 360                     int axis = axes.get<int>(i);
 361                     CV_Assert_N(axis >= 2 + i, axis < inpShape.size());
 362                     kernel_size[axis - 2] = inpShape[axis];
 363                 }
 364
 365                 layerParams.set("kernel_size", DictValue::arrayInt(&kernel_size[0], kernel_size.size()));
 366             }
 367         }
 368         else if (layer_type == "Slice")
 369         {
 370             if (layerParams.has("steps")) {
 371                 DictValue steps = layerParams.get("steps");
 372                 for (int i = 0; i < steps.size(); ++i) {
 373                     if (steps.get<int>(i) != 1)
 374                         CV_Error(Error::StsNotImplemented,
 375                                  "Slice layer only supports steps = 1");
 376                 }
 377             }
 378
 379             int axis = 0;
 380             if (layerParams.has("axes")) {
 381                 DictValue axes = layerParams.get("axes");
 382                 for (int i = 1; i < axes.size(); ++i) {
 383                     CV_Assert(axes.get<int>(i - 1) == axes.get<int>(i) - 1);
 384                 }
 385                 axis = axes.get<int>(0);
 386             }
 387             layerParams.set("axis", axis);
 388
 389             DictValue starts = layerParams.get("starts");
 390             DictValue ends = layerParams.get("ends");
 391             CV_Assert(starts.size() == ends.size());
 392
 393             std::vector<int> begin;
 394             std::vector<int> end;
 395             if (axis > 0) {
 396                 begin.resize(axis, 0);
 397                 end.resize(axis, -1);
 398             }
 399
 400             for (int i = 0; i < starts.size(); ++i)
 401             {
 402                 begin.push_back(starts.get<int>(i));
 403                 int finish = ends.get<int>(i);
 404                 end.push_back((finish < 0) ? --finish : finish); // numpy doesn't include last dim
 405             }
 406             layerParams.set("begin", DictValue::arrayInt(&begin[0], begin.size()));
 407             layerParams.set("end", DictValue::arrayInt(&end[0], end.size()));
 408          }
 409         else if (layer_type == "Split")
 410         {
 411             if (layerParams.has("split"))
 412             {
 413                 DictValue splits = layerParams.get("split");
 414                 const int numSplits = splits.size();
 415                 CV_Assert(numSplits > 1);
 416
 417                 std::vector<int> slicePoints(numSplits - 1, splits.get<int>(0));
 418                 for (int i = 1; i < splits.size() - 1; ++i)
 419                 {
 420                     slicePoints[i] = slicePoints[i - 1] + splits.get<int>(i - 1);
 421                 }
 422                 layerParams.set("slice_point", DictValue::arrayInt(&slicePoints[0], slicePoints.size()));
 423             }
 424             else
 425             {
 426                 layerParams.set("num_split", node_proto.output_size());
 427             }
 428             layerParams.type = "Slice";
 429         }
 430         else if (layer_type == "Add" || layer_type == "Sum")
 431         {
 432             if (layer_id.find(node_proto.input(1)) == layer_id.end())
 433             {
 434                 Mat blob = getBlob(node_proto, constBlobs, 1);
 435                 blob = blob.reshape(1, 1);
 436                 if (blob.total() == 1) {
 437                     layerParams.type = "Power";
 438                     layerParams.set("shift", blob.at<float>(0));
 439                 }
 440                 else {
 441                     layerParams.type = "Scale";
 442                     layerParams.set("bias_term", true);
 443                     layerParams.blobs.push_back(blob);
 444                 }
 445             }
 446             else {
 447                 layerParams.type = "Eltwise";
 448             }
 449         }
 450         else if (layer_type == "Max")
 451         {
 452             layerParams.type = "Eltwise";
 453             layerParams.set("operation", "max");
 454         }
 455         else if (layer_type == "Sub")
 456         {
 457             Mat blob = getBlob(node_proto, constBlobs, 1);
 458             if (blob.total() == 1) {
 459                 layerParams.type = "Power";
 460                 layerParams.set("shift", -blob.at<float>(0));
 461             }
 462             else {
 463                 layerParams.type = "Scale";
 464                 layerParams.set("has_bias", true);
 465                 layerParams.blobs.push_back(-1.0f * blob.reshape(1, 1));
 466             }
 467         }
 468         else if (layer_type == "Div")
 469         {
 470             if (constBlobs.find(node_proto.input(1)) == constBlobs.end())
 471             {
 472                 layerParams.type = "Eltwise";
 473                 layerParams.set("operation", "div");
 474             }
 475             else
 476             {
 477                 Mat blob = getBlob(node_proto, constBlobs, 1);
 478                 CV_Assert_N(blob.type() == CV_32F, blob.total());
 479                 if (blob.total() == 1)
 480                 {
 481                     layerParams.set("scale", 1.0f / blob.at<float>(0));
 482                     layerParams.type = "Power";
 483                 }
 484                 else
 485                 {
 486                     layerParams.type = "Scale";
 487                     divide(1.0, blob, blob);
 488                     layerParams.blobs.push_back(blob);
 489                     layerParams.set("bias_term", false);
 490                 }
 491             }
 492         }
 493         else if (layer_type == "Neg")
 494         {
 495             layerParams.type = "Power";
 496             layerParams.set("scale", -1);
 497         }
 498         else if (layer_type == "Constant")
 499         {
 500             CV_Assert(node_proto.input_size() == 0);
 501             CV_Assert(layerParams.blobs.size() == 1);
 502             constBlobs.insert(std::make_pair(layerParams.name, layerParams.blobs[0]));
 503             continue;
 504         }
 505         else if (layer_type == "ImageScaler")
 506         {
 507             const float scale = layerParams.has("scale") ? layerParams.get<float>("scale") : 1.0f;
 508             layerParams.erase("scale");
 509
 510             if (layerParams.has("bias"))
 511             {
 512                 layerParams.type = "Scale";
 513                 layerParams.blobs.push_back(
 514                     Mat(Size(1,  layerParams.get("bias").size()), CV_32FC1, scale));
 515
 516                 layerParams.set("bias_term", true);
 517                 Mat bias(1, layerParams.get("bias").size(), CV_32FC1);
 518                 for (int j = 0; j < bias.total(); j++) {
 519                     bias.at<float>(0, j) = layerParams.get("bias").getRealValue(j);
 520                 }
 521                 layerParams.blobs.push_back(bias);
 522                 layerParams.erase("bias");
 523             }
 524             else {
 525                 layerParams.set("scale", scale);
 526                 layerParams.type = "Power";
 527             }
 528         }
 529         else if (layer_type == "Clip")
 530         {
 531             layerParams.type = "ReLU6";
 532             replaceLayerParam(layerParams, "min", "min_value");
 533             replaceLayerParam(layerParams, "max", "max_value");
 534
 535         }
 536         else if (layer_type == "LeakyRelu")
 537         {
 538             layerParams.type = "ReLU";
 539             replaceLayerParam(layerParams, "alpha", "negative_slope");
 540         }
 541         else if (layer_type == "LRN")
 542         {
 543             replaceLayerParam(layerParams, "size", "local_size");
 544         }
 545         else if (layer_type == "InstanceNormalization")
 546         {
 547             if (node_proto.input_size() != 3)
 548                 CV_Error(Error::StsNotImplemented,
 549                          "Expected input, scale, bias");
 550
 551             layerParams.blobs.resize(4);
 552             layerParams.blobs[2] = getBlob(node_proto, constBlobs, 1);  // weightData
 553             layerParams.blobs[3] = getBlob(node_proto, constBlobs, 2);  // biasData
 554             layerParams.set("has_bias", true);
 555             layerParams.set("has_weight", true);
 556
 557             // Get number of channels in input
 558             int size = layerParams.blobs[2].total();
 559             layerParams.blobs[0] = Mat::zeros(size, 1, CV_32F); // mean
 560             layerParams.blobs[1] = Mat::ones(size, 1, CV_32F); // std
 561
 562             LayerParams mvnParams;
 563             mvnParams.name = layerParams.name + "/MVN";
 564             mvnParams.type = "MVN";
 565             mvnParams.set("eps", layerParams.get<float>("epsilon"));
 566             layerParams.erase("epsilon");
 567
 568             //Create MVN layer
 569             int id = dstNet.addLayer(mvnParams.name, mvnParams.type, mvnParams);
 570             //Connect to input
 571             layerId = layer_id.find(node_proto.input(0));
 572             CV_Assert(layerId != layer_id.end());
 573             dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
 574             //Add shape
 575             layer_id.insert(std::make_pair(mvnParams.name, LayerInfo(id, 0)));
 576             outShapes[mvnParams.name] = outShapes[node_proto.input(0)];
 577
 578             //Replace Batch Norm's input to MVN
 579             node_proto.set_input(0, mvnParams.name);
 580             layerParams.type = "BatchNorm";
 581         }
 582         else if (layer_type == "BatchNormalization")
 583         {
 584             if (node_proto.input_size() != 5)
 585                 CV_Error(Error::StsNotImplemented,
 586                          "Expected input, scale, bias, mean and var");
 587
 588             layerParams.type = "BatchNorm";
 589             replaceLayerParam(layerParams, "epsilon", "eps");
 590             replaceLayerParam(layerParams, "spatial", "use_global_stats");
 591
 592             Mat meanData = getBlob(node_proto, constBlobs, 3);
 593             Mat stdData =  getBlob(node_proto, constBlobs, 4);
 594
 595             layerParams.blobs.push_back(meanData);
 596             layerParams.blobs.push_back(stdData);
 597
 598             if (!node_proto.input(1).empty()) {
 599                 layerParams.set("has_weight", true);
 600                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, 1));  // weightData
 601             } else {
 602                 layerParams.set("has_weight", false);
 603             }
 604
 605             if (!node_proto.input(2).empty()) {
 606                 layerParams.set("has_bias", true);
 607                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, 2)); // biasData
 608             } else {
 609                 layerParams.set("has_bias", false);
 610             }
 611         }
 612         else if (layer_type == "Gemm")
 613         {
 614             CV_Assert(node_proto.input_size() >= 2);
 615             layerParams.type = "InnerProduct";
 616             Mat weights = getBlob(node_proto, constBlobs, 1);
 617             int ind_num_out = 0;
 618             if (layerParams.has("transB") && !layerParams.get<int>("transB")) {
 619                 transpose(weights, weights);
 620                 ind_num_out = 1;
 621             }
 622             layerParams.blobs.push_back(weights);
 623
 624             if (node_proto.input_size() == 3) {
 625                 Mat bias = getBlob(node_proto, constBlobs, 2);
 626                 layerParams.blobs.push_back(bias);
 627             }
 628
 629             layerParams.set("num_output", layerParams.blobs[0].size[ind_num_out]);
 630             layerParams.set("bias_term", node_proto.input_size() == 3);
 631         }
 632         else if (layer_type == "MatMul")
 633         {
 634             CV_Assert(node_proto.input_size() == 2);
 635             layerParams.type = "InnerProduct";
 636             Mat blob = getBlob(node_proto, constBlobs, 1);
 637             layerParams.blobs.push_back(blob.t());
 638             layerParams.set("bias_term", false);
 639             layerParams.set("num_output", layerParams.blobs[0].size[0]);
 640         }
 641         else if (layer_type == "Mul")
 642         {
 643             CV_Assert(node_proto.input_size() == 2);
 644             if (layer_id.find(node_proto.input(1)) == layer_id.end()) {
 645                 Mat blob = getBlob(node_proto, constBlobs, 1);
 646                 blob = blob.reshape(1, 1);
 647                 if (blob.total() == 1) {
 648                     layerParams.set("scale", blob.at<float>(0));
 649                     layerParams.type = "Power";
 650                 }
 651                 else {
 652                     layerParams.blobs.push_back(blob);
 653                     layerParams.type = "Scale";
 654                 }
 655             }
 656             else {
 657                 layerParams.type = "Eltwise";
 658                 layerParams.set("operation", "prod");
 659             }
 660         }
 661         else if (layer_type == "Conv")
 662         {
 663             CV_Assert(node_proto.input_size() >= 2);
 664             layerParams.type = "Convolution";
 665             for (int j = 1; j < node_proto.input_size(); j++) {
 666                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
 667             }
 668             layerParams.set("num_output", layerParams.blobs[0].size[0]);
 669             layerParams.set("bias_term", node_proto.input_size() == 3);
 670         }
 671         else if (layer_type == "ConvTranspose")
 672         {
 673             CV_Assert(node_proto.input_size() >= 2);
 674             layerParams.type = "Deconvolution";
 675             for (int j = 1; j < node_proto.input_size(); j++) {
 676                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
 677             }
 678             layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1));
 679             layerParams.set("bias_term", node_proto.input_size() == 3);
 680
 681             if (!layerParams.has("kernel_size"))
 682                 CV_Error(Error::StsNotImplemented,
 683                          "Required attribute 'kernel_size' is not present.");
 684
 685             if (layerParams.has("output_shape"))
 686             {
 687                 const DictValue& outShape = layerParams.get("output_shape");
 688                 DictValue strides = layerParams.get("stride");
 689                 DictValue kernel = layerParams.get("kernel_size");
 690
 691                 String padMode;
 692                 std::vector<int> adjust_pads;
 693                 if (layerParams.has("pad_mode"))
 694                 {
 695                     padMode = toUpperCase(layerParams.get<String>("pad_mode"));
 696                     if (padMode != "SAME" && padMode != "VALID")
 697                         CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
 698
 699                     for (int i = 0; i < strides.size(); i++)
 700                     {
 701                         int sz = outShape.get<int>(2 + i);
 702                         int stride = strides.get<int>(i);
 703                         adjust_pads.push_back(padMode == "SAME"? (sz - 1) % stride :
 704                                                                  (sz - kernel.get<int>(i)) % stride);
 705                     }
 706                     layerParams.set("adj", DictValue::arrayInt(&adjust_pads[0], adjust_pads.size()));
 707                 }
 708             }
 709             else if (layerParams.has("output_padding"))
 710             {
 711                 replaceLayerParam(layerParams, "output_padding", "adj");
 712             }
 713         }
 714         else if (layer_type == "Transpose")
 715         {
 716             layerParams.type = "Permute";
 717             replaceLayerParam(layerParams, "perm", "order");
 718
 719             CV_Assert(node_proto.input_size() == 1);
 720             if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
 721             {
 722                 std::vector<Mat> inputs(1, getBlob(node_proto, constBlobs, 0)), transposed;
 723                 runLayer(layerParams, inputs, transposed);
 724                 CV_Assert(transposed.size() == 1);
 725                 constBlobs.insert(std::make_pair(layerParams.name, transposed[0]));
 726                 continue;
 727             }
 728         }
 729         else if (layer_type == "ReduceL2")
 730         {
 731             CV_Assert_N(node_proto.input_size() == 1, layerParams.has("axes"));
 732             CV_Assert(graph_proto.node_size() > li + 1 && graph_proto.node(li + 1).op_type() == "Div");
 733             ++li;
 734             node_proto = graph_proto.node(li);
 735             layerParams.name = node_proto.output(0);
 736             layerParams.type = "Normalize";
 737
 738             DictValue axes_dict = layerParams.get("axes");
 739             if (axes_dict.size() != 1)
 740                 CV_Error(Error::StsNotImplemented, "Multidimensional reduceL2");
 741             int axis = axes_dict.getIntValue(0);
 742             layerParams.set("axis",axis);
 743             layerParams.set("end_axis", axis);
 744         }
 745         else if (layer_type == "Squeeze")
 746         {
 747             CV_Assert_N(node_proto.input_size() == 1, layerParams.has("axes"));
 748             DictValue axes_dict = layerParams.get("axes");
 749             if (axes_dict.size() != 1)
 750                 CV_Error(Error::StsNotImplemented, "Multidimensional squeeze");
 751
 752             int axis = axes_dict.getIntValue(0);
 753             layerParams.set("axis", axis - 1);
 754             layerParams.set("end_axis", axis);
 755             layerParams.type = "Flatten";
 756         }
 757         else if (layer_type == "Unsqueeze")
 758         {
 759             CV_Assert(node_proto.input_size() == 1);
 760             DictValue axes = layerParams.get("axes");
 761             if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
 762             {
 763                 // Constant input.
 764                 Mat input = getBlob(node_proto, constBlobs, 0);
 765
 766                 std::vector<int> dims;
 767                 for (int j = 0; j < input.dims; j++) {
 768                     dims.push_back(input.size[j]);
 769                 }
 770                 CV_Assert(axes.getIntValue(axes.size()-1) <= dims.size());
 771                 for (int j = 0; j < axes.size(); j++) {
 772                     dims.insert(dims.begin() + axes.getIntValue(j), 1);
 773                 }
 774
 775                 Mat out = input.reshape(0, dims);
 776                 constBlobs.insert(std::make_pair(layerParams.name, out));
 777                 continue;
 778             }
 779
 780             // Variable input.
 781             if (axes.size() != 1)
 782                 CV_Error(Error::StsNotImplemented, "Multidimensional unsqueeze");
 783
 784             MatShape inpShape = outShapes[node_proto.input(0)];
 785             int axis = axes.getIntValue(0);
 786             CV_Assert(0 <= axis && axis <= inpShape.size());
 787             std::vector<int> outShape = inpShape;
 788             outShape.insert(outShape.begin() + axis, 1);
 789             layerParams.type = "Reshape";
 790             layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
 791         }
 792         else if (layer_type == "Reshape")
 793         {
 794             CV_Assert(node_proto.input_size() == 2 || layerParams.has("shape"));
 795
 796             if (node_proto.input_size() == 2) {
 797                 Mat blob = getBlob(node_proto, constBlobs, 1);
 798                 CV_Assert(blob.type() == CV_32SC1);
 799
 800                 layerParams.set("dim", DictValue::arrayInt<int*>(
 801                             blob.ptr<int>(), blob.total() ));
 802
 803                 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
 804                     std::vector<Mat> inputs(1, getBlob(node_proto, constBlobs, 0)), outputs;
 805                     runLayer(layerParams, inputs, outputs);
 806                     constBlobs.insert(std::make_pair(layerParams.name, outputs[0]));
 807                     continue;
 808                 }
 809             }
 810             else {
 811                 DictValue shape = layerParams.get("shape");
 812                 std::vector<int> dim;
 813                 for (int j = 0; j < shape.size(); j++) {
 814                     dim.push_back(shape.getIntValue(j));
 815                 }
 816
 817                 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
 818                     Mat input = getBlob(node_proto, constBlobs, 0);
 819                     Mat out = input.reshape(0, dim);
 820                     constBlobs.insert(std::make_pair(layerParams.name, out));
 821                     continue;
 822                 }
 823                 replaceLayerParam(layerParams, "shape", "dim");
 824             }
 825         }
 826         else if (layer_type == "Pad")
 827         {
 828             layerParams.type = "Padding";
 829         }
 830         else if (layer_type == "Shape")
 831         {
 832             CV_Assert(node_proto.input_size() == 1);
 833             shapeIt = outShapes.find(node_proto.input(0));
 834             CV_Assert(shapeIt != outShapes.end());
 835             MatShape inpShape = shapeIt->second;
 836
 837             Mat shapeMat(inpShape.size(), 1, CV_32S);
 838             for (int j = 0; j < inpShape.size(); ++j)
 839                 shapeMat.at<int>(j) = inpShape[j];
 840             shapeMat.dims = 1;
 841
 842             constBlobs.insert(std::make_pair(layerParams.name, shapeMat));
 843             continue;
 844         }
 845         else if (layer_type == "Gather")
 846         {
 847             CV_Assert(node_proto.input_size() == 2);
 848             CV_Assert(layerParams.has("axis"));
 849             Mat input = getBlob(node_proto, constBlobs, 0);
 850             Mat indexMat = getBlob(node_proto, constBlobs, 1);
 851             CV_Assert_N(indexMat.type() == CV_32S, indexMat.total() == 1);
 852             int index = indexMat.at<int>(0);
 853             int axis = layerParams.get<int>("axis");
 854
 855             std::vector<cv::Range> ranges(input.dims, Range::all());
 856             ranges[axis] = Range(index, index + 1);
 857
 858             Mat out = input(ranges);
 859             constBlobs.insert(std::make_pair(layerParams.name, out));
 860             continue;
 861         }
 862         else if (layer_type == "Concat")
 863         {
 864             bool hasVariableInps = false;
 865             for (int i = 0; i < node_proto.input_size(); ++i)
 866             {
 867                 if (layer_id.find(node_proto.input(i)) != layer_id.end())
 868                 {
 869                     hasVariableInps = true;
 870                     break;
 871                 }
 872             }
 873
 874             if (!hasVariableInps)
 875             {
 876                 std::vector<Mat> inputs(node_proto.input_size()), concatenated;
 877                 for (size_t i = 0; i < inputs.size(); ++i)
 878                 {
 879                     inputs[i] = getBlob(node_proto, constBlobs, i);
 880                 }
 881                 runLayer(layerParams, inputs, concatenated);
 882
 883                 CV_Assert(concatenated.size() == 1);
 884                 constBlobs.insert(std::make_pair(layerParams.name, concatenated[0]));
 885                 continue;
 886             }
 887         }
 888         else if (layer_type == "Upsample")
 889         {
 890             layerParams.type = "Resize";
 891             if (layerParams.has("scales"))
 892             {
 893                 // Pytorch layer
 894                 DictValue scales = layerParams.get("scales");
 895                 CV_Assert(scales.size() == 4);
 896                 layerParams.set("zoom_factor_y", scales.getIntValue(2));
 897                 layerParams.set("zoom_factor_x", scales.getIntValue(3));
 898             }
 899             else
 900             {
 901                 // Caffe2 layer
 902                 replaceLayerParam(layerParams, "height_scale", "zoom_factor_y");
 903                 replaceLayerParam(layerParams, "width_scale", "zoom_factor_x");
 904             }
 905             replaceLayerParam(layerParams, "mode", "interpolation");
 906
 907             if (layerParams.get<String>("interpolation") == "linear" && framework_name == "pytorch") {
 908                 layerParams.type = "Resize";
 909                 Mat scales = getBlob(node_proto, constBlobs, 1);
 910                 CV_Assert(scales.total() == 4);
 911                 layerParams.set("interpolation", "opencv_linear");
 912                 layerParams.set("zoom_factor_y", scales.at<float>(2));
 913                 layerParams.set("zoom_factor_x", scales.at<float>(3));
 914             }
 915         }
 916         else if (layer_type == "LogSoftmax")
 917         {
 918             layerParams.type = "Softmax";
 919             layerParams.set("log_softmax", true);
 920         }
 921         else
 922         {
 923             for (int j = 0; j < node_proto.input_size(); j++) {
 924                 if (layer_id.find(node_proto.input(j)) == layer_id.end())
 925                     layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
 926             }
 927         }
 928
 929         int id = dstNet.addLayer(layerParams.name, layerParams.type, layerParams);
 930         for (int i = 0; i < node_proto.output_size(); ++i)
 931         {
 932             layer_id.insert(std::make_pair(node_proto.output(i), LayerInfo(id, i)));
 933         }
 934
 935         std::vector<MatShape> layerInpShapes, layerOutShapes, layerInternalShapes;
 936         for (int j = 0; j < node_proto.input_size(); j++) {
 937             layerId = layer_id.find(node_proto.input(j));
 938             if (layerId != layer_id.end()) {
 939                 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, j);
 940                 // Collect input shapes.
 941                 shapeIt = outShapes.find(node_proto.input(j));
 942                 CV_Assert(shapeIt != outShapes.end());
 943                 layerInpShapes.push_back(shapeIt->second);
 944             }
 945         }
 946
 947         // Compute shape of output blob for this layer.
 948         Ptr<Layer> layer = dstNet.getLayer(id);
 949         layer->getMemoryShapes(layerInpShapes, 0, layerOutShapes, layerInternalShapes);
 950         for (int i = 0; i < node_proto.output_size() && i < (int)layerOutShapes.size(); ++i)
 951         {
 952             outShapes[node_proto.output(i)] = layerOutShapes[i];
 953         }
 954     }
 955 }
 956
 957 Net readNetFromONNX(const String& onnxFile)
 958 {
 959     ONNXImporter onnxImporter(onnxFile.c_str());
 960     Net net;
 961     onnxImporter.populateNet(net);
 962     return net;
 963 }
 964
 965 Net readNetFromONNX(const char* buffer, size_t sizeBuffer)
 966 {
 967     ONNXImporter onnxImporter(buffer, sizeBuffer);
 968     Net net;
 969     onnxImporter.populateNet(net);
 970     return net;
 971 }
 972
 973 Net readNetFromONNX(const std::vector<uchar>& buffer)
 974 {
 975     return readNetFromONNX(reinterpret_cast<const char*>(buffer.data()), buffer.size());
 976 }
 977
 978 Mat readTensorFromONNX(const String& path)
 979 {
 980     opencv_onnx::TensorProto tensor_proto = opencv_onnx::TensorProto();
 981     std::fstream input(path.c_str(), std::ios::in | std::ios::binary);
 982     if (!tensor_proto.ParseFromIstream(&input)) {
 983         CV_Error(Error::StsUnsupportedFormat, "Failed to parse data");
 984     }
 985     Mat mat = getMatFromTensor(tensor_proto);
 986     releaseONNXTensor(tensor_proto);
 987     return mat;
 988 }
 989
 990 CV__DNN_EXPERIMENTAL_NS_END
 991 }} // namespace
 992
 993 #endif