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_INLINE_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 template<typename T1, typename T2>
  99 void convertInt64ToInt32(const T1& src, T2& dst, int size)
 100 {
 101     for (int i = 0; i < size; i++) {
 102         if (src[i] < std::numeric_limits<int32_t>::min() || src[i] > std::numeric_limits<int32_t>::max()) {
 103             CV_Error(Error::StsOutOfRange, "Input is out of OpenCV 32S range");
 104         }
 105         dst[i] = saturate_cast<int32_t>(src[i]);
 106     }
 107 }
 108
 109 Mat getMatFromTensor(opencv_onnx::TensorProto& tensor_proto)
 110 {
 111     CV_Assert(!tensor_proto.raw_data().empty() || !tensor_proto.float_data().empty()
 112                     || !tensor_proto.double_data().empty() || !tensor_proto.int64_data().empty());
 113
 114     opencv_onnx::TensorProto_DataType datatype = tensor_proto.data_type();
 115     Mat blob;
 116     std::vector<int> sizes;
 117     for (int i = 0; i < tensor_proto.dims_size(); i++) {
 118             sizes.push_back(tensor_proto.dims(i));
 119     }
 120     if (sizes.empty())
 121         sizes.assign(1, 1);
 122     if (datatype == opencv_onnx::TensorProto_DataType_FLOAT) {
 123
 124         if (!tensor_proto.float_data().empty()) {
 125             const ::google::protobuf::RepeatedField<float> field = tensor_proto.float_data();
 126             Mat(sizes, CV_32FC1, (void*)field.data()).copyTo(blob);
 127         }
 128         else {
 129             char* val = const_cast<char*>(tensor_proto.raw_data().c_str());
 130             Mat(sizes, CV_32FC1, val).copyTo(blob);
 131         }
 132     }
 133     else if (datatype == opencv_onnx::TensorProto_DataType_DOUBLE)
 134     {
 135         const ::google::protobuf::RepeatedField<double> field = tensor_proto.double_data();
 136         CV_Assert(!field.empty());
 137         Mat(sizes, CV_64FC1, (void*)field.data()).convertTo(blob, CV_32FC1);
 138     }
 139     else if (datatype == opencv_onnx::TensorProto_DataType_INT64)
 140     {
 141         blob.create(sizes, CV_32SC1);
 142         int32_t* dst = reinterpret_cast<int32_t*>(blob.data);
 143
 144         if (!tensor_proto.int64_data().empty()) {
 145             ::google::protobuf::RepeatedField< ::google::protobuf::int64> src = tensor_proto.int64_data();
 146             convertInt64ToInt32(src, dst, blob.total());
 147         }
 148         else
 149         {
 150             const char* val = tensor_proto.raw_data().c_str();
 151             // Aligned pointer is required: https://github.com/opencv/opencv/issues/16373
 152             // this doesn't work: typedef int64_t CV_DECL_ALIGNED(1) unaligned_int64_t;
 153             AutoBuffer<int64_t, 16> aligned_val;
 154             if (!isAligned<sizeof(int64_t)>(val))
 155             {
 156                 size_t sz = tensor_proto.raw_data().size();
 157                 aligned_val.allocate(divUp(sz, sizeof(int64_t)));
 158                 memcpy(aligned_val.data(), val, sz);
 159                 val = (const char*)aligned_val.data();
 160             }
 161             const int64_t* src = reinterpret_cast<const int64_t*>(val);
 162             convertInt64ToInt32(src, dst, blob.total());
 163         }
 164     }
 165     else
 166         CV_Error(Error::StsUnsupportedFormat, "Unsupported data type: " +
 167                         opencv_onnx::TensorProto_DataType_Name(datatype));
 168     if (tensor_proto.dims_size() == 0)
 169         blob.dims = 1;  // To force 1-dimensional cv::Mat for scalars.
 170     return blob;
 171 }
 172
 173 void runLayer(LayerParams& params, const std::vector<Mat>& inputs,
 174               std::vector<Mat>& outputs)
 175 {
 176     Ptr<Layer> layer = LayerFactory::createLayerInstance(params.type, params);
 177     CV_Assert((bool)layer);
 178
 179     std::vector<MatShape> inpShapes(inputs.size());
 180     int ddepth = CV_32F;
 181     for (size_t i = 0; i < inputs.size(); ++i)
 182     {
 183         inpShapes[i] = shape(inputs[i]);
 184         if (i > 0 && ddepth != inputs[i].depth())
 185             CV_Error(Error::StsNotImplemented, "Mixed input data types.");
 186         ddepth = inputs[i].depth();
 187     }
 188
 189     std::vector<MatShape> outShapes, internalShapes;
 190     layer->getMemoryShapes(inpShapes, 0, outShapes, internalShapes);
 191
 192     std::vector<Mat> internals(internalShapes.size());
 193     outputs.resize(outShapes.size());
 194     for (size_t i = 0; i < outShapes.size(); ++i)
 195         outputs[i].create(outShapes[i], ddepth);
 196     for (size_t i = 0; i < internalShapes.size(); ++i)
 197         internals[i].create(internalShapes[i], ddepth);
 198
 199     layer->finalize(inputs, outputs);
 200     layer->forward(inputs, outputs, internals);
 201 }
 202
 203 std::map<std::string, Mat> ONNXImporter::getGraphTensors(
 204                                         const opencv_onnx::GraphProto& graph_proto)
 205 {
 206   opencv_onnx::TensorProto tensor_proto;
 207   std::map<std::string, Mat> layers_weights;
 208
 209   for (int i = 0; i < graph_proto.initializer_size(); i++)
 210   {
 211     tensor_proto = graph_proto.initializer(i);
 212     Mat mat = getMatFromTensor(tensor_proto);
 213     releaseONNXTensor(tensor_proto);
 214     layers_weights.insert(std::make_pair(tensor_proto.name(), mat));
 215   }
 216   return layers_weights;
 217 }
 218
 219 static DictValue parse(const ::google::protobuf::RepeatedField< ::google::protobuf::int64>& src) {
 220     std::vector<int32_t> dst(src.size());
 221     convertInt64ToInt32(src, dst, src.size());
 222     return DictValue::arrayInt(&dst[0], src.size());
 223 }
 224
 225 LayerParams ONNXImporter::getLayerParams(const opencv_onnx::NodeProto& node_proto)
 226 {
 227     LayerParams lp;
 228     for(int i = 0; i < node_proto.attribute_size(); i++)
 229     {
 230         opencv_onnx::AttributeProto attribute_proto = node_proto.attribute(i);
 231         std::string attribute_name = attribute_proto.name();
 232
 233         if(attribute_name == "kernel_shape")
 234         {
 235             CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
 236             lp.set("kernel_size", parse(attribute_proto.ints()));
 237         }
 238         else if(attribute_name == "strides")
 239         {
 240             CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
 241             lp.set("stride", parse(attribute_proto.ints()));
 242         }
 243         else if(attribute_name == "pads")
 244         {
 245             if (node_proto.op_type() == "Pad")
 246             {
 247                 // Padding layer.
 248                 // Paddings are in order begin0, begin1, .. beginN, end0, end1, ..., endN.
 249                 // We need to shuffle it to begin0, end0, begin1, end1, ...
 250                 CV_Assert(attribute_proto.ints_size() % 2 == 0);
 251                 const int dims = attribute_proto.ints_size() / 2;
 252                 std::vector<int32_t> paddings;
 253                 paddings.reserve(attribute_proto.ints_size());
 254                 for (int i = 0; i < dims; ++i)
 255                 {
 256                     paddings.push_back(attribute_proto.ints(i));
 257                     paddings.push_back(attribute_proto.ints(dims + i));
 258                 }
 259                 lp.set("paddings", DictValue::arrayInt(&paddings[0], paddings.size()));
 260             }
 261             else
 262             {
 263                 // Convolution or pooling.
 264                 CV_Assert(attribute_proto.ints_size() == 4 || attribute_proto.ints_size() == 6);
 265                 lp.set("pad", parse(attribute_proto.ints()));
 266             }
 267         }
 268         else if(attribute_name == "auto_pad")
 269         {
 270             if (attribute_proto.s() == "SAME_UPPER" || attribute_proto.s() == "SAME_LOWER") {
 271                 lp.set("pad_mode",  "SAME");
 272             }
 273             else if (attribute_proto.s() == "VALID") {
 274                 lp.set("pad_mode", "VALID");
 275             }
 276         }
 277         else if(attribute_name == "dilations")
 278         {
 279             CV_Assert(attribute_proto.ints_size() == 2 || attribute_proto.ints_size() == 3);
 280             lp.set("dilation", parse(attribute_proto.ints()));
 281         }
 282         else if (attribute_proto.has_i())
 283         {
 284             ::google::protobuf::int64 src = attribute_proto.i();
 285             if (src < std::numeric_limits<int32_t>::min() || src > std::numeric_limits<int32_t>::max())
 286                 CV_Error(Error::StsOutOfRange, "Input is out of OpenCV 32S range");
 287             else
 288                 lp.set(attribute_name, saturate_cast<int32_t>(src));
 289         }
 290         else if (attribute_proto.has_f())
 291         {
 292             lp.set(attribute_name, attribute_proto.f());
 293         }
 294         else if (attribute_proto.has_s())
 295         {
 296             lp.set(attribute_name, attribute_proto.s());
 297         }
 298         else if (attribute_proto.floats_size() > 0)
 299         {
 300             lp.set(attribute_name, DictValue::arrayReal(
 301                 attribute_proto.floats().data(), attribute_proto.floats_size()));
 302         }
 303         else if (attribute_proto.ints_size() > 0)
 304         {
 305             lp.set(attribute_proto.name(), parse(attribute_proto.ints()));
 306         }
 307         else if (attribute_proto.has_t())
 308         {
 309             opencv_onnx::TensorProto tensor = attribute_proto.t();
 310             Mat blob = getMatFromTensor(tensor);
 311             lp.blobs.push_back(blob);
 312         }
 313         else if (attribute_proto.has_g() || attribute_proto.strings_size() > 0 ||
 314                     attribute_proto.tensors_size() > 0 || attribute_proto.graphs_size() > 0)
 315         {
 316                 CV_Error(Error::StsNotImplemented, "Unexpected attribute type");
 317         }
 318         else
 319             CV_Error(Error::StsNotImplemented, "Unsupported attribute type");
 320     }
 321     return lp;
 322 }
 323
 324 Mat ONNXImporter::getBlob(const opencv_onnx::NodeProto& node_proto,
 325                     const std::map<std::string, Mat>& constBlobs, int index)
 326 {
 327     CV_Assert(index < node_proto.input_size());
 328     std::map<std::string, Mat>::const_iterator constBlob;
 329     constBlob = constBlobs.find(node_proto.input(index));
 330     if (constBlob == constBlobs.end()) {
 331         CV_Error(Error::StsObjectNotFound,
 332              "Blob " + node_proto.input(index) + " not found in const blobs");
 333     }
 334     return constBlob->second;
 335 }
 336
 337 void ONNXImporter::populateNet(Net dstNet)
 338 {
 339     CV_Assert(model_proto.has_graph());
 340     opencv_onnx::GraphProto graph_proto = model_proto.graph();
 341
 342     simplifySubgraphs(graph_proto);
 343
 344     std::map<std::string, Mat> constBlobs = getGraphTensors(graph_proto);
 345     // List of internal blobs shapes.
 346     std::map<std::string, MatShape> outShapes;
 347     // Add all the inputs shapes. It includes as constant blobs as network's inputs shapes.
 348     for (int i = 0; i < graph_proto.input_size(); ++i)
 349     {
 350         opencv_onnx::ValueInfoProto valueInfoProto = graph_proto.input(i);
 351         CV_Assert(valueInfoProto.has_type());
 352         opencv_onnx::TypeProto typeProto = valueInfoProto.type();
 353         CV_Assert(typeProto.has_tensor_type());
 354         opencv_onnx::TypeProto::Tensor tensor = typeProto.tensor_type();
 355         CV_Assert(tensor.has_shape());
 356         opencv_onnx::TensorShapeProto tensorShape = tensor.shape();
 357
 358         MatShape inpShape(tensorShape.dim_size());
 359         for (int j = 0; j < inpShape.size(); ++j)
 360         {
 361             inpShape[j] = tensorShape.dim(j).dim_value();
 362         }
 363         outShapes[valueInfoProto.name()] = inpShape;
 364     }
 365
 366     std::string framework_name;
 367     if (model_proto.has_producer_name()) {
 368         framework_name = model_proto.producer_name();
 369     }
 370
 371     // create map with network inputs (without const blobs)
 372     std::map<std::string, LayerInfo> layer_id;
 373     std::map<std::string, LayerInfo>::iterator layerId;
 374     std::map<std::string, MatShape>::iterator shapeIt;
 375     // fill map: push layer name, layer id and output id
 376     std::vector<String> netInputs;
 377     for (int j = 0; j < graph_proto.input_size(); j++)
 378     {
 379         const std::string& name = graph_proto.input(j).name();
 380         if (constBlobs.find(name) == constBlobs.end()) {
 381             netInputs.push_back(name);
 382             layer_id.insert(std::make_pair(name, LayerInfo(0, netInputs.size() - 1)));
 383         }
 384     }
 385     dstNet.setInputsNames(netInputs);
 386
 387     int layersSize = graph_proto.node_size();
 388     LayerParams layerParams;
 389     opencv_onnx::NodeProto node_proto;
 390
 391     for(int li = 0; li < layersSize; li++)
 392     {
 393         node_proto = graph_proto.node(li);
 394         layerParams = getLayerParams(node_proto);
 395         CV_Assert(node_proto.output_size() >= 1);
 396         layerParams.name = node_proto.output(0);
 397
 398         std::string layer_type = node_proto.op_type();
 399         layerParams.type = layer_type;
 400
 401
 402         if (layer_type == "MaxPool")
 403         {
 404             layerParams.type = "Pooling";
 405             layerParams.set("pool", "MAX");
 406             layerParams.set("ceil_mode", layerParams.has("pad_mode"));
 407         }
 408         else if (layer_type == "AveragePool")
 409         {
 410             layerParams.type = "Pooling";
 411             layerParams.set("pool", "AVE");
 412             layerParams.set("ceil_mode", layerParams.has("pad_mode"));
 413             layerParams.set("ave_pool_padded_area", framework_name == "pytorch");
 414         }
 415         else if (layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool" || layer_type == "ReduceMean")
 416         {
 417             CV_Assert(node_proto.input_size() == 1);
 418             layerParams.type = "Pooling";
 419             layerParams.set("pool", layer_type == "GlobalMaxPool"? "MAX" : "AVE");
 420             layerParams.set("global_pooling", layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool");
 421
 422             if (layer_type == "ReduceMean")
 423             {
 424                 if (layerParams.get<int>("keepdims") == 0 || !layerParams.has("axes"))
 425                     CV_Error(Error::StsNotImplemented, "Unsupported mode of ReduceMean operation.");
 426
 427                 MatShape inpShape = outShapes[node_proto.input(0)];
 428                 if (inpShape.size() != 4 && inpShape.size() != 5)
 429                     CV_Error(Error::StsNotImplemented, "Unsupported input shape of reduce_mean operation.");
 430
 431                 DictValue axes = layerParams.get("axes");
 432                 CV_Assert(axes.size() <= inpShape.size() - 2);
 433                 std::vector<int> kernel_size(inpShape.size() - 2, 1);
 434                 for (int i = 0; i < axes.size(); i++) {
 435                     int axis = axes.get<int>(i);
 436                     CV_Assert_N(axis >= 2 + i, axis < inpShape.size());
 437                     kernel_size[axis - 2] = inpShape[axis];
 438                 }
 439
 440                 layerParams.set("kernel_size", DictValue::arrayInt(&kernel_size[0], kernel_size.size()));
 441             }
 442         }
 443         else if (layer_type == "Slice")
 444         {
 445             if (layerParams.has("steps")) {
 446                 DictValue steps = layerParams.get("steps");
 447                 for (int i = 0; i < steps.size(); ++i) {
 448                     if (steps.get<int>(i) != 1)
 449                         CV_Error(Error::StsNotImplemented,
 450                                  "Slice layer only supports steps = 1");
 451                 }
 452             }
 453
 454             int axis = 0;
 455             if (layerParams.has("axes")) {
 456                 DictValue axes = layerParams.get("axes");
 457                 for (int i = 1; i < axes.size(); ++i) {
 458                     CV_Assert(axes.get<int>(i - 1) == axes.get<int>(i) - 1);
 459                 }
 460                 axis = axes.get<int>(0);
 461             }
 462             layerParams.set("axis", axis);
 463
 464             DictValue starts = layerParams.get("starts");
 465             DictValue ends = layerParams.get("ends");
 466             CV_Assert(starts.size() == ends.size());
 467
 468             std::vector<int> begin;
 469             std::vector<int> end;
 470             if (axis > 0) {
 471                 begin.resize(axis, 0);
 472                 end.resize(axis, -1);
 473             }
 474
 475             for (int i = 0; i < starts.size(); ++i)
 476             {
 477                 begin.push_back(starts.get<int>(i));
 478                 int finish = ends.get<int>(i);
 479                 end.push_back((finish < 0) ? --finish : finish); // numpy doesn't include last dim
 480             }
 481             layerParams.set("begin", DictValue::arrayInt(&begin[0], begin.size()));
 482             layerParams.set("end", DictValue::arrayInt(&end[0], end.size()));
 483          }
 484         else if (layer_type == "Split")
 485         {
 486             DictValue splits = layerParams.get("split");
 487             const int numSplits = splits.size();
 488             CV_Assert(numSplits > 1);
 489
 490             std::vector<int> slicePoints(numSplits - 1, splits.get<int>(0));
 491             for (int i = 1; i < splits.size() - 1; ++i)
 492             {
 493                 slicePoints[i] = slicePoints[i - 1] + splits.get<int>(i - 1);
 494             }
 495             layerParams.set("slice_point", DictValue::arrayInt(&slicePoints[0], slicePoints.size()));
 496             layerParams.type = "Slice";
 497         }
 498         else if (layer_type == "Add" || layer_type == "Sum")
 499         {
 500             if (layer_id.find(node_proto.input(1)) == layer_id.end())
 501             {
 502                 Mat blob = getBlob(node_proto, constBlobs, 1);
 503                 blob = blob.reshape(1, 1);
 504                 if (blob.total() == 1) {
 505                     layerParams.type = "Power";
 506                     layerParams.set("shift", blob.at<float>(0));
 507                 }
 508                 else {
 509                     layerParams.type = "Scale";
 510                     layerParams.set("bias_term", true);
 511                     layerParams.blobs.push_back(blob);
 512                 }
 513             }
 514             else {
 515                 layerParams.type = "Eltwise";
 516             }
 517         }
 518         else if (layer_type == "Max")
 519         {
 520             layerParams.type = "Eltwise";
 521             layerParams.set("operation", "max");
 522         }
 523         else if (layer_type == "Sub")
 524         {
 525             Mat blob = getBlob(node_proto, constBlobs, 1);
 526             if (blob.total() == 1) {
 527                 layerParams.type = "Power";
 528                 layerParams.set("shift", -blob.at<float>(0));
 529             }
 530             else {
 531                 layerParams.type = "Scale";
 532                 layerParams.set("has_bias", true);
 533                 layerParams.blobs.push_back(-1.0f * blob.reshape(1, 1));
 534             }
 535         }
 536         else if (layer_type == "Div")
 537         {
 538             if (constBlobs.find(node_proto.input(1)) == constBlobs.end())
 539             {
 540                 layerParams.type = "Eltwise";
 541                 layerParams.set("operation", "div");
 542             }
 543             else
 544             {
 545                 Mat blob = getBlob(node_proto, constBlobs, 1);
 546                 CV_Assert_N(blob.type() == CV_32F, blob.total());
 547                 if (blob.total() == 1)
 548                 {
 549                     layerParams.set("scale", 1.0f / blob.at<float>(0));
 550                     layerParams.type = "Power";
 551                 }
 552                 else
 553                 {
 554                     layerParams.type = "Scale";
 555                     divide(1.0, blob, blob);
 556                     layerParams.blobs.push_back(blob);
 557                     layerParams.set("bias_term", false);
 558                 }
 559             }
 560         }
 561         else if (layer_type == "Neg")
 562         {
 563             layerParams.type = "Power";
 564             layerParams.set("scale", -1);
 565         }
 566         else if (layer_type == "Constant")
 567         {
 568             CV_Assert(node_proto.input_size() == 0);
 569             CV_Assert(layerParams.blobs.size() == 1);
 570             constBlobs.insert(std::make_pair(layerParams.name, layerParams.blobs[0]));
 571             continue;
 572         }
 573         else if (layer_type == "ImageScaler")
 574         {
 575             const float scale = layerParams.has("scale") ? layerParams.get<float>("scale") : 1.0f;
 576             layerParams.erase("scale");
 577
 578             if (layerParams.has("bias"))
 579             {
 580                 layerParams.type = "Scale";
 581                 layerParams.blobs.push_back(
 582                     Mat(Size(1,  layerParams.get("bias").size()), CV_32FC1, scale));
 583
 584                 layerParams.set("bias_term", true);
 585                 Mat bias(1, layerParams.get("bias").size(), CV_32FC1);
 586                 for (int j = 0; j < bias.total(); j++) {
 587                     bias.at<float>(0, j) = layerParams.get("bias").getRealValue(j);
 588                 }
 589                 layerParams.blobs.push_back(bias);
 590                 layerParams.erase("bias");
 591             }
 592             else {
 593                 layerParams.set("scale", scale);
 594                 layerParams.type = "Power";
 595             }
 596         }
 597         else if (layer_type == "Clip")
 598         {
 599             layerParams.type = "ReLU6";
 600             replaceLayerParam(layerParams, "min", "min_value");
 601             replaceLayerParam(layerParams, "max", "max_value");
 602
 603         }
 604         else if (layer_type == "LeakyRelu")
 605         {
 606             layerParams.type = "ReLU";
 607             replaceLayerParam(layerParams, "alpha", "negative_slope");
 608         }
 609         else if (layer_type == "LRN")
 610         {
 611             replaceLayerParam(layerParams, "size", "local_size");
 612         }
 613         else if (layer_type == "InstanceNormalization")
 614         {
 615             if (node_proto.input_size() != 3)
 616                 CV_Error(Error::StsNotImplemented,
 617                          "Expected input, scale, bias");
 618
 619             layerParams.blobs.resize(4);
 620             layerParams.blobs[2] = getBlob(node_proto, constBlobs, 1);  // weightData
 621             layerParams.blobs[3] = getBlob(node_proto, constBlobs, 2);  // biasData
 622             layerParams.set("has_bias", true);
 623             layerParams.set("has_weight", true);
 624
 625             // Get number of channels in input
 626             int size = layerParams.blobs[2].total();
 627             layerParams.blobs[0] = Mat::zeros(size, 1, CV_32F); // mean
 628             layerParams.blobs[1] = Mat::ones(size, 1, CV_32F); // std
 629
 630             LayerParams mvnParams;
 631             mvnParams.name = layerParams.name + "/MVN";
 632             mvnParams.type = "MVN";
 633             mvnParams.set("eps", layerParams.get<float>("epsilon"));
 634             layerParams.erase("epsilon");
 635
 636             //Create MVN layer
 637             int id = dstNet.addLayer(mvnParams.name, mvnParams.type, mvnParams);
 638             //Connect to input
 639             layerId = layer_id.find(node_proto.input(0));
 640             CV_Assert(layerId != layer_id.end());
 641             dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, 0);
 642             //Add shape
 643             layer_id.insert(std::make_pair(mvnParams.name, LayerInfo(id, 0)));
 644             outShapes[mvnParams.name] = outShapes[node_proto.input(0)];
 645
 646             //Replace Batch Norm's input to MVN
 647             node_proto.set_input(0, mvnParams.name);
 648             layerParams.type = "BatchNorm";
 649         }
 650         else if (layer_type == "BatchNormalization")
 651         {
 652             if (node_proto.input_size() != 5)
 653                 CV_Error(Error::StsNotImplemented,
 654                          "Expected input, scale, bias, mean and var");
 655
 656             layerParams.type = "BatchNorm";
 657             replaceLayerParam(layerParams, "epsilon", "eps");
 658             replaceLayerParam(layerParams, "spatial", "use_global_stats");
 659
 660             Mat meanData = getBlob(node_proto, constBlobs, 3);
 661             Mat stdData =  getBlob(node_proto, constBlobs, 4);
 662
 663             layerParams.blobs.push_back(meanData);
 664             layerParams.blobs.push_back(stdData);
 665
 666             if (!node_proto.input(1).empty()) {
 667                 layerParams.set("has_weight", true);
 668                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, 1));  // weightData
 669             } else {
 670                 layerParams.set("has_weight", false);
 671             }
 672
 673             if (!node_proto.input(2).empty()) {
 674                 layerParams.set("has_bias", true);
 675                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, 2)); // biasData
 676             } else {
 677                 layerParams.set("has_bias", false);
 678             }
 679         }
 680         else if (layer_type == "Gemm")
 681         {
 682             CV_Assert(node_proto.input_size() >= 2);
 683             layerParams.type = "InnerProduct";
 684             Mat weights = getBlob(node_proto, constBlobs, 1);
 685             int ind_num_out = 0;
 686             if (layerParams.has("transB") && !layerParams.get<int>("transB")) {
 687                 transpose(weights, weights);
 688                 ind_num_out = 1;
 689             }
 690             layerParams.blobs.push_back(weights);
 691
 692             if (node_proto.input_size() == 3) {
 693                 Mat bias = getBlob(node_proto, constBlobs, 2);
 694                 layerParams.blobs.push_back(bias);
 695             }
 696
 697             layerParams.set("num_output", layerParams.blobs[0].size[ind_num_out]);
 698             layerParams.set("bias_term", node_proto.input_size() == 3);
 699         }
 700         else if (layer_type == "MatMul")
 701         {
 702             CV_Assert(node_proto.input_size() == 2);
 703             layerParams.type = "InnerProduct";
 704             Mat blob = getBlob(node_proto, constBlobs, 1);
 705             layerParams.blobs.push_back(blob.t());
 706             layerParams.set("bias_term", false);
 707             layerParams.set("num_output", layerParams.blobs[0].size[0]);
 708         }
 709         else if (layer_type == "Mul")
 710         {
 711             CV_Assert(node_proto.input_size() == 2);
 712             if (layer_id.find(node_proto.input(1)) == layer_id.end()) {
 713                 Mat blob = getBlob(node_proto, constBlobs, 1);
 714                 blob = blob.reshape(1, 1);
 715                 if (blob.total() == 1) {
 716                     layerParams.set("scale", blob.at<float>(0));
 717                     layerParams.type = "Power";
 718                 }
 719                 else {
 720                     layerParams.blobs.push_back(blob);
 721                     layerParams.type = "Scale";
 722                 }
 723             }
 724             else {
 725                 layerParams.type = "Eltwise";
 726                 layerParams.set("operation", "prod");
 727             }
 728         }
 729         else if (layer_type == "Conv")
 730         {
 731             CV_Assert(node_proto.input_size() >= 2);
 732             layerParams.type = "Convolution";
 733             for (int j = 1; j < node_proto.input_size(); j++) {
 734                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
 735             }
 736             layerParams.set("num_output", layerParams.blobs[0].size[0]);
 737             layerParams.set("bias_term", node_proto.input_size() == 3);
 738         }
 739         else if (layer_type == "ConvTranspose")
 740         {
 741             CV_Assert(node_proto.input_size() >= 2);
 742             layerParams.type = "Deconvolution";
 743             for (int j = 1; j < node_proto.input_size(); j++) {
 744                 layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
 745             }
 746             layerParams.set("num_output", layerParams.blobs[0].size[1] * layerParams.get<int>("group", 1));
 747             layerParams.set("bias_term", node_proto.input_size() == 3);
 748
 749             if (!layerParams.has("kernel_size"))
 750                 CV_Error(Error::StsNotImplemented,
 751                          "Required attribute 'kernel_size' is not present.");
 752
 753             if (layerParams.has("output_shape"))
 754             {
 755                 const DictValue& outShape = layerParams.get("output_shape");
 756                 DictValue strides = layerParams.get("stride");
 757                 DictValue kernel = layerParams.get("kernel_size");
 758
 759                 String padMode;
 760                 std::vector<int> adjust_pads;
 761                 if (layerParams.has("pad_mode"))
 762                 {
 763                     padMode = toUpperCase(layerParams.get<String>("pad_mode"));
 764                     if (padMode != "SAME" && padMode != "VALID")
 765                         CV_Error(Error::StsError, "Unsupported padding mode " + padMode);
 766
 767                     for (int i = 0; i < strides.size(); i++)
 768                     {
 769                         int sz = outShape.get<int>(2 + i);
 770                         int stride = strides.get<int>(i);
 771                         adjust_pads.push_back(padMode == "SAME"? (sz - 1) % stride :
 772                                                                  (sz - kernel.get<int>(i)) % stride);
 773                     }
 774                     layerParams.set("adj", DictValue::arrayInt(&adjust_pads[0], adjust_pads.size()));
 775                 }
 776             }
 777             else if (layerParams.has("output_padding"))
 778             {
 779                 replaceLayerParam(layerParams, "output_padding", "adj");
 780             }
 781         }
 782         else if (layer_type == "Transpose")
 783         {
 784             layerParams.type = "Permute";
 785             replaceLayerParam(layerParams, "perm", "order");
 786
 787             CV_Assert(node_proto.input_size() == 1);
 788             if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
 789             {
 790                 std::vector<Mat> inputs(1, getBlob(node_proto, constBlobs, 0)), transposed;
 791                 runLayer(layerParams, inputs, transposed);
 792                 CV_Assert(transposed.size() == 1);
 793                 constBlobs.insert(std::make_pair(layerParams.name, transposed[0]));
 794                 continue;
 795             }
 796         }
 797         else if (layer_type == "ReduceL2")
 798         {
 799             CV_Assert_N(node_proto.input_size() == 1, layerParams.has("axes"));
 800             CV_Assert(graph_proto.node_size() > li + 1 && graph_proto.node(li + 1).op_type() == "Div");
 801             ++li;
 802             node_proto = graph_proto.node(li);
 803             layerParams.name = node_proto.output(0);
 804             layerParams.type = "Normalize";
 805
 806             DictValue axes_dict = layerParams.get("axes");
 807             if (axes_dict.size() != 1)
 808                 CV_Error(Error::StsNotImplemented, "Multidimensional reduceL2");
 809             int axis = axes_dict.getIntValue(0);
 810             layerParams.set("axis",axis);
 811             layerParams.set("end_axis", axis);
 812         }
 813         else if (layer_type == "Squeeze")
 814         {
 815             CV_Assert_N(node_proto.input_size() == 1, layerParams.has("axes"));
 816             DictValue axes_dict = layerParams.get("axes");
 817             if (axes_dict.size() != 1)
 818                 CV_Error(Error::StsNotImplemented, "Multidimensional squeeze");
 819
 820             int axis = axes_dict.getIntValue(0);
 821             layerParams.set("axis", axis - 1);
 822             layerParams.set("end_axis", axis);
 823             layerParams.type = "Flatten";
 824         }
 825         else if (layer_type == "Unsqueeze")
 826         {
 827             CV_Assert(node_proto.input_size() == 1);
 828             DictValue axes = layerParams.get("axes");
 829             if (constBlobs.find(node_proto.input(0)) != constBlobs.end())
 830             {
 831                 // Constant input.
 832                 Mat input = getBlob(node_proto, constBlobs, 0);
 833
 834                 std::vector<int> dims;
 835                 for (int j = 0; j < input.dims; j++) {
 836                     dims.push_back(input.size[j]);
 837                 }
 838                 CV_Assert(axes.getIntValue(axes.size()-1) <= dims.size());
 839                 for (int j = 0; j < axes.size(); j++) {
 840                     dims.insert(dims.begin() + axes.getIntValue(j), 1);
 841                 }
 842
 843                 Mat out = input.reshape(0, dims);
 844                 constBlobs.insert(std::make_pair(layerParams.name, out));
 845                 continue;
 846             }
 847
 848             // Variable input.
 849             if (axes.size() != 1)
 850                 CV_Error(Error::StsNotImplemented, "Multidimensional unsqueeze");
 851
 852             MatShape inpShape = outShapes[node_proto.input(0)];
 853             int axis = axes.getIntValue(0);
 854             CV_Assert(0 <= axis && axis <= inpShape.size());
 855             std::vector<int> outShape = inpShape;
 856             outShape.insert(outShape.begin() + axis, 1);
 857             layerParams.type = "Reshape";
 858             layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
 859         }
 860         else if (layer_type == "Reshape")
 861         {
 862             CV_Assert(node_proto.input_size() == 2 || layerParams.has("shape"));
 863
 864             if (node_proto.input_size() == 2) {
 865                 Mat blob = getBlob(node_proto, constBlobs, 1);
 866                 CV_Assert(blob.type() == CV_32SC1);
 867
 868                 layerParams.set("dim", DictValue::arrayInt<int*>(
 869                             blob.ptr<int>(), blob.total() ));
 870
 871                 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
 872                     std::vector<Mat> inputs(1, getBlob(node_proto, constBlobs, 0)), outputs;
 873                     runLayer(layerParams, inputs, outputs);
 874                     constBlobs.insert(std::make_pair(layerParams.name, outputs[0]));
 875                     continue;
 876                 }
 877             }
 878             else {
 879                 DictValue shape = layerParams.get("shape");
 880                 std::vector<int> dim;
 881                 for (int j = 0; j < shape.size(); j++) {
 882                     dim.push_back(shape.getIntValue(j));
 883                 }
 884
 885                 if (layer_id.find(node_proto.input(0)) == layer_id.end()) {
 886                     Mat input = getBlob(node_proto, constBlobs, 0);
 887                     Mat out = input.reshape(0, dim);
 888                     constBlobs.insert(std::make_pair(layerParams.name, out));
 889                     continue;
 890                 }
 891                 replaceLayerParam(layerParams, "shape", "dim");
 892             }
 893         }
 894         else if (layer_type == "Pad")
 895         {
 896             layerParams.type = "Padding";
 897         }
 898         else if (layer_type == "Shape")
 899         {
 900             CV_Assert(node_proto.input_size() == 1);
 901             shapeIt = outShapes.find(node_proto.input(0));
 902             CV_Assert(shapeIt != outShapes.end());
 903             MatShape inpShape = shapeIt->second;
 904
 905             Mat shapeMat(inpShape.size(), 1, CV_32S);
 906             for (int j = 0; j < inpShape.size(); ++j)
 907                 shapeMat.at<int>(j) = inpShape[j];
 908             shapeMat.dims = 1;
 909
 910             constBlobs.insert(std::make_pair(layerParams.name, shapeMat));
 911             continue;
 912         }
 913         else if (layer_type == "Gather")
 914         {
 915             CV_Assert(node_proto.input_size() == 2);
 916             CV_Assert(layerParams.has("axis"));
 917             Mat input = getBlob(node_proto, constBlobs, 0);
 918             Mat indexMat = getBlob(node_proto, constBlobs, 1);
 919             CV_Assert_N(indexMat.type() == CV_32S, indexMat.total() == 1);
 920             int index = indexMat.at<int>(0);
 921             int axis = layerParams.get<int>("axis");
 922
 923             std::vector<cv::Range> ranges(input.dims, Range::all());
 924             ranges[axis] = Range(index, index + 1);
 925
 926             Mat out = input(ranges);
 927             constBlobs.insert(std::make_pair(layerParams.name, out));
 928             continue;
 929         }
 930         else if (layer_type == "Concat")
 931         {
 932             bool hasVariableInps = false;
 933             for (int i = 0; i < node_proto.input_size(); ++i)
 934             {
 935                 if (layer_id.find(node_proto.input(i)) != layer_id.end())
 936                 {
 937                     hasVariableInps = true;
 938                     break;
 939                 }
 940             }
 941
 942             if (!hasVariableInps)
 943             {
 944                 std::vector<Mat> inputs(node_proto.input_size()), concatenated;
 945                 for (size_t i = 0; i < inputs.size(); ++i)
 946                 {
 947                     inputs[i] = getBlob(node_proto, constBlobs, i);
 948                 }
 949                 runLayer(layerParams, inputs, concatenated);
 950
 951                 CV_Assert(concatenated.size() == 1);
 952                 constBlobs.insert(std::make_pair(layerParams.name, concatenated[0]));
 953                 continue;
 954             }
 955         }
 956         else if (layer_type == "Upsample")
 957         {
 958             layerParams.type = "Resize";
 959             if (layerParams.has("scales"))
 960             {
 961                 // Pytorch layer
 962                 DictValue scales = layerParams.get("scales");
 963                 CV_Assert(scales.size() == 4);
 964                 layerParams.set("zoom_factor_y", scales.getIntValue(2));
 965                 layerParams.set("zoom_factor_x", scales.getIntValue(3));
 966             }
 967             else
 968             {
 969                 // Caffe2 layer
 970                 replaceLayerParam(layerParams, "height_scale", "zoom_factor_y");
 971                 replaceLayerParam(layerParams, "width_scale", "zoom_factor_x");
 972             }
 973             replaceLayerParam(layerParams, "mode", "interpolation");
 974         }
 975         else if (layer_type == "LogSoftmax")
 976         {
 977             layerParams.type = "Softmax";
 978             layerParams.set("log_softmax", true);
 979         }
 980         else
 981         {
 982             for (int j = 0; j < node_proto.input_size(); j++) {
 983                 if (layer_id.find(node_proto.input(j)) == layer_id.end())
 984                     layerParams.blobs.push_back(getBlob(node_proto, constBlobs, j));
 985             }
 986         }
 987
 988         int id = dstNet.addLayer(layerParams.name, layerParams.type, layerParams);
 989         for (int i = 0; i < node_proto.output_size(); ++i)
 990         {
 991             layer_id.insert(std::make_pair(node_proto.output(i), LayerInfo(id, i)));
 992         }
 993
 994         std::vector<MatShape> layerInpShapes, layerOutShapes, layerInternalShapes;
 995         for (int j = 0; j < node_proto.input_size(); j++) {
 996             layerId = layer_id.find(node_proto.input(j));
 997             if (layerId != layer_id.end()) {
 998                 dstNet.connect(layerId->second.layerId, layerId->second.outputId, id, j);
 999                 // Collect input shapes.
1000                 shapeIt = outShapes.find(node_proto.input(j));
1001                 CV_Assert(shapeIt != outShapes.end());
1002                 layerInpShapes.push_back(shapeIt->second);
1003             }
1004         }
1005
1006         // Compute shape of output blob for this layer.
1007         Ptr<Layer> layer = dstNet.getLayer(id);
1008         layer->getMemoryShapes(layerInpShapes, 0, layerOutShapes, layerInternalShapes);
1009         for (int i = 0; i < node_proto.output_size() && i < (int)layerOutShapes.size(); ++i)
1010         {
1011             outShapes[node_proto.output(i)] = layerOutShapes[i];
1012         }
1013     }
1014 }
1015
1016 Net readNetFromONNX(const String& onnxFile)
1017 {
1018     ONNXImporter onnxImporter(onnxFile.c_str());
1019     Net net;
1020     onnxImporter.populateNet(net);
1021     return net;
1022 }
1023
1024 Net readNetFromONNX(const char* buffer, size_t sizeBuffer)
1025 {
1026     ONNXImporter onnxImporter(buffer, sizeBuffer);
1027     Net net;
1028     onnxImporter.populateNet(net);
1029     return net;
1030 }
1031
1032 Net readNetFromONNX(const std::vector<uchar>& buffer)
1033 {
1034     return readNetFromONNX(reinterpret_cast<const char*>(buffer.data()), buffer.size());
1035 }
1036
1037 Mat readTensorFromONNX(const String& path)
1038 {
1039     opencv_onnx::TensorProto tensor_proto = opencv_onnx::TensorProto();
1040     std::fstream input(path.c_str(), std::ios::in | std::ios::binary);
1041     if (!tensor_proto.ParseFromIstream(&input)) {
1042         CV_Error(Error::StsUnsupportedFormat, "Failed to parse data");
1043     }
1044     Mat mat = getMatFromTensor(tensor_proto);
1045     releaseONNXTensor(tensor_proto);
1046     return mat;
1047 }
1048
1049 CV__DNN_INLINE_NS_END
1050 }} // namespace
1051
1052 #endif