tf_importer.cpp

// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

// Copyright (C) 2016, Intel Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.

/*
Implementation of Tensorflow models parser
*/

#include "../precomp.hpp"

#ifdef HAVE_PROTOBUF
#include "tf_io.hpp"

#include <iostream>
#include <fstream>
#include <algorithm>
#include <string>
#include <queue>
#include "tf_graph_simplifier.hpp"
#endif

namespace cv {
namespace dnn {
CV__DNN_EXPERIMENTAL_NS_BEGIN

#if HAVE_PROTOBUF

using ::google::protobuf::RepeatedField;
using ::google::protobuf::RepeatedPtrField;
using ::google::protobuf::Message;
using ::google::protobuf::Descriptor;
using ::google::protobuf::FieldDescriptor;
using ::google::protobuf::Reflection;

namespace
{

static int toNCHW(int idx)
{
    CV_Assert(-4 <= idx && idx < 4);
    if (idx == 0) return 0;
    else if (idx > 0) return idx % 3 + 1;
    else return (4 + idx) % 3 + 1;
}

// This values are used to indicate layer output's data layout where it's possible.
enum DataLayout
{
    DATA_LAYOUT_NHWC,
    DATA_LAYOUT_NCHW,
    DATA_LAYOUT_UNKNOWN
};

typedef std::vector<std::pair<String, int> > StrIntVector;

struct Pin
{
    Pin(const std::string &_name, int _blobIndex = 0) :
        name(_name), blobIndex(_blobIndex) {}

    Pin() :
        name(""), blobIndex(-1) {}

    std::string name;
    int blobIndex;
};

void blobShapeFromTensor(const tensorflow::TensorProto &tensor, MatShape& shape)
{
    shape.clear();
    if (tensor.has_tensor_shape())
    {
        const tensorflow::TensorShapeProto &_shape = tensor.tensor_shape();
        int i, n = _shape.dim_size();
        if (n)
        {
            shape.resize(n);

            for (i = 0; i < n; i++)
                shape[i] = (int)_shape.dim(i).size();
        }
        else
            shape.resize(1, 1);  // Scalar.
    }
    else
    {
        CV_Error(Error::StsError, "Unknown shape of input tensor");
    }
}

template <typename T>
void parseTensor(const tensorflow::TensorProto &tensor, Mat &dstBlob)
{
    MatShape shape;
    blobShapeFromTensor(tensor, shape);
    int dims = (int)shape.size();

    if (dims == 4)
    {
        // REORDER blob NHWC to NCHW
        swap(shape[2], shape[3]); // NHCW
        swap(shape[1], shape[2]); // NCHW
    }

    dstBlob.create(shape, CV_32F);

    Mat tensorContent = getTensorContent(tensor);
    int size = tensorContent.total();
    CV_Assert(size == (int)dstBlob.total());

    float *dstData = dstBlob.ptr<float>();
    const T *data = reinterpret_cast<const T*>(tensorContent.data);

    if (dims == 4)
    {
        int num = shape[0], channels = shape[1], height = shape[2], width = shape[3];
        int total = num*channels*height*width;
        for(int i_n = 0; i_n < shape[0]; i_n++) {
            for(int i_c = 0; i_c < shape[1]; i_c++) {
                for(int i_h = 0; i_h < shape[2]; i_h++) {
                    for(int i_w = 0; i_w < shape[3]; i_w++) {
                       int dst_i = channels*height*width*i_n + height*width*i_c + width*i_h + i_w;
                       int src_i = channels*height*width*i_n + i_c + channels*width*i_h + channels*i_w;

                       CV_Assert(dst_i < total);
                       CV_Assert(src_i < total);

                       dstData[dst_i] = data[src_i];
                    }
                }
            }
        }
    } else {
        for (int i = 0; i < size; i++)
            dstData[i] = data[i];
    }
}

void blobFromTensor(const tensorflow::TensorProto &tensor, Mat &dstBlob)
{
    switch (tensor.dtype()) {
        case tensorflow::DT_FLOAT:
        case tensorflow::DT_HALF:
            parseTensor<float>(tensor, dstBlob);
            break;
        case tensorflow::DT_DOUBLE:
            parseTensor<double>(tensor, dstBlob);
            break;
        default:
            CV_Error(Error::StsError, "Tensor's data type is not supported");
            break;
    }
}

void printList(const tensorflow::AttrValue::ListValue &val)
{
    std::cout << "(";
    for (int i = 0; i < val.i_size(); i++)
        std::cout << " " << val.i(i);
    std::cout << " )";
}

void printTensorShape(const tensorflow::TensorShapeProto &shape)
{
    std::cout << "[ ";
    for (int d = 0; d < shape.dim_size(); d++)
        std::cout << shape.dim(d).name() <<
                     ":" << shape.dim(d).size() << " ";
    std::cout << "]";
}

void printTensor(const tensorflow::TensorProto &tensor)
{
    printTensorShape(tensor.tensor_shape());

    if (tensor.tensor_content().empty())
        return;

    switch (tensor.dtype())
    {
    case tensorflow::DT_FLOAT:
        {
            const float *data = reinterpret_cast<const float*>(tensor.tensor_content().c_str());
            int size = tensor.tensor_content().size() / sizeof(float);
            for (int i = 0; i < std::min(10, size); i++)
                std::cout << " " << data[i];
            if (size > 10)
                std::cout << " ... " << size - 10 << " more";
            break;
        }
    case tensorflow::DT_INT32:
        {
            const int *data = reinterpret_cast<const int*>(tensor.tensor_content().c_str());
            int size = tensor.tensor_content().size() / sizeof(int);
            for (int i = 0; i < std::min(10, size); i++)
                std::cout << " " << data[i];
            if (size > 10)
                std::cout << " ... " << size - 10 << " more";
            break;
        }
    default:
        CV_Error(Error::StsError, "Tensor type is not supported");
        break;
    }
}

void printLayerAttr(const tensorflow::NodeDef &layer)
{
    std::cout << std::endl << layer.name() << ":" << layer.op();
    for (int ii = 0; ii < layer.input_size(); ii++)
        std::cout << "(" << layer.input(ii) << ")";
    std::cout << std::endl;
    google::protobuf::Map<std::string, tensorflow::AttrValue> attr
            = layer.attr();
    for (google::protobuf::Map<std::string, tensorflow::AttrValue>::const_iterator ai = attr.begin();
         ai != attr.end(); ++ai)
    {
        std::cout << ai->first << ":";
        if (ai->first == "dtype" || ai->first == "T")
            std::cout << ai->second.i();
        else if (ai->first == "padding")
            std::cout << ai->second.s();
        else if (ai->first == "transpose_a" || ai->first == "transpose_b")
            std::cout << ai->second.b();
        //            else if (ai->first == "shape")
        //              printTensorShape(ai->second.shape());
        else if (ai->first == "strides" || ai->first == "ksize")
            printList(ai->second.list());
        else
            printTensor(ai->second.tensor());
        std::cout << std::endl;
    }
}

bool hasLayerAttr(const tensorflow::NodeDef &layer, const std::string &name)
{
    google::protobuf::Map<std::string, tensorflow::AttrValue> attr = layer.attr();
    return attr.find(name) != attr.end();
}

const tensorflow::AttrValue& getLayerAttr(const tensorflow::NodeDef &layer, const std::string &name)
{
    return layer.attr().at(name);
}

void setStrides(LayerParams &layerParams, const tensorflow::NodeDef &layer)
{
    if (hasLayerAttr(layer, "strides"))
    {
        const tensorflow::AttrValue& val = getLayerAttr(layer, "strides");
        if (val.list().i_size() != 4 ||
            val.list().i(0) != 1 || val.list().i(3) != 1)
            CV_Error(Error::StsError, "Unsupported strides");
        layerParams.set("stride_h", static_cast<int>(val.list().i(1)));
        layerParams.set("stride_w", static_cast<int>(val.list().i(2)));
    }
}

DictValue parseDims(const tensorflow::TensorProto &tensor) {
    MatShape shape;
    blobShapeFromTensor(tensor, shape);
    int dims = (int)shape.size();

    CV_Assert(tensor.dtype() == tensorflow::DT_INT32);
    CV_Assert(dims == 1);

    Mat values = getTensorContent(tensor);
    CV_Assert(values.type() == CV_32SC1);
    // TODO: add reordering shape if dims == 4
    return DictValue::arrayInt((int*)values.data, values.total());
}

void setKSize(LayerParams &layerParams, const tensorflow::NodeDef &layer)
{
    if (hasLayerAttr(layer, "ksize"))
    {
        const tensorflow::AttrValue& val = getLayerAttr(layer, "ksize");
        if (val.list().i_size() != 4 ||
            val.list().i(0) != 1 || val.list().i(3) != 1)
            CV_Error(Error::StsError, "Unsupported ksize");
        layerParams.set("kernel_h", static_cast<int>(val.list().i(1)));
        layerParams.set("kernel_w", static_cast<int>(val.list().i(2)));
    }
    else
    {
        layerParams.set("kernel_h", 1);
        layerParams.set("kernel_w", 1);
    }
}

void setPadding(LayerParams &layerParams, const tensorflow::NodeDef &layer)
{
    if (hasLayerAttr(layer, "padding"))
        layerParams.set("pad_mode", getLayerAttr(layer, "padding").s());
}

Pin parsePin(const std::string &name)
{
    Pin pin(name);

    size_t delimiter_pos = name.find_first_of(":");
    if (delimiter_pos != std::string::npos)
    {
        pin.name = name.substr(0, delimiter_pos);
        std::istringstream(name.substr(delimiter_pos + 1)) >> pin.blobIndex;
    }

    return pin;
}

StrIntVector getNextLayers(const tensorflow::GraphDef& net, const String& layer_name, const String& type = "")
{
   StrIntVector layers;

   for (int li = 0; li < net.node_size(); li++)
   {
       const tensorflow::NodeDef& layer = net.node(li);
       for (int input_id = 0; input_id < layer.input_size(); input_id++) {
           String input_op_name = parsePin(layer.input(input_id)).name;
           bool type_ok = type.empty() ? true : type == layer.op();
           if (input_op_name == layer_name && type_ok)
               layers.push_back(std::make_pair(layer.name(), li));
       }
   }

   return layers;
}

void ExcludeLayer(tensorflow::GraphDef& net, const int layer_index, const int input_blob_index, bool remove_from_net = true) {
    String layer_name = net.node(layer_index).name();
    StrIntVector layers = getNextLayers(net, layer_name);

    String removed_layer_input = net.node(layer_index).input(input_blob_index);

    for (size_t i = 0; i < layers.size(); i++)
    {
        tensorflow::NodeDef* layer = net.mutable_node(layers[i].second);
        for (int input_id = 0; input_id < layer->input_size(); input_id++) {
                String input_op_name = layer->input(input_id);

                if (input_op_name == layer_name) {
                    layer->set_input(input_id, removed_layer_input);
                }
        }
    }

    if (remove_from_net)
        net.mutable_node()->DeleteSubrange(layer_index, 1);
}

class TFImporter {
public:
    TFImporter(const char *model, const char *config = NULL);
    TFImporter(const char *dataModel, size_t lenModel,
               const char *dataConfig = NULL, size_t lenConfig = 0);

    void populateNet(Net dstNet);

private:
    void kernelFromTensor(const tensorflow::TensorProto &tensor, Mat &dstBlob);

    void connect(const std::map<String, int>& layers_name_id_map, Net& network, const Pin& outPin,
                 const int input_layer_id, const int input_blob_id);
    void connectToAllBlobs(const std::map<String, int>& layer_id, Net& network, const Pin& outPin,
                           const int input_layer_id, const int input_blobs_count);
    const tensorflow::TensorProto& getConstBlob(const tensorflow::NodeDef &layer, std::map<String, int> const_layers,
                                                int input_blob_index = -1, int* actual_inp_blob_idx = 0);


    // Binary serialized TensorFlow graph includes weights.
    tensorflow::GraphDef netBin;
    // Optional text definition of TensorFlow graph. More flexible than binary format
    // and may be used to build the network using binary format only as a weights storage.
    // This approach is similar to Caffe's `.prorotxt` and `.caffemodel`.
    tensorflow::GraphDef netTxt;

    std::vector<String> netInputsNames;
};

TFImporter::TFImporter(const char *model, const char *config)
{
    if (model && model[0])
        ReadTFNetParamsFromBinaryFileOrDie(model, &netBin);
    if (config && config[0])
        ReadTFNetParamsFromTextFileOrDie(config, &netTxt);
}

TFImporter::TFImporter(const char *dataModel, size_t lenModel,
                       const char *dataConfig, size_t lenConfig)
{
    if (dataModel != NULL && lenModel > 0)
        ReadTFNetParamsFromBinaryBufferOrDie(dataModel, lenModel, &netBin);
    if (dataConfig != NULL && lenConfig > 0)
        ReadTFNetParamsFromTextBufferOrDie(dataConfig, lenConfig, &netTxt);
}

void TFImporter::kernelFromTensor(const tensorflow::TensorProto &tensor, Mat &dstBlob)
{
    MatShape shape;
    blobShapeFromTensor(tensor, shape);
    int dims = (int)shape.size();

    // TODO: other blob types
    CV_Assert(tensor.dtype() == tensorflow::DT_FLOAT ||
              tensor.dtype() == tensorflow::DT_HALF);
    CV_Assert(dims == 4);

    // REORDER kernel HWIO to OIHW
    swap(shape[0], shape[2]); // IWHO
    swap(shape[1], shape[3]); // IOHW
    swap(shape[0], shape[1]); // OIHW

    dstBlob.create(shape, CV_32F);

    Mat tensorContent = getTensorContent(tensor);
    int size = tensorContent.total();
    CV_Assert(size == (int)dstBlob.total());

    float *dstData = dstBlob.ptr<float>();
    const float *data = reinterpret_cast<const float*>(tensorContent.data);

    int out_c = shape[0], input_c = shape[1], height = shape[2], width = shape[3];
    int total = out_c*input_c*height*width;
    for(int i_oc = 0; i_oc < out_c; i_oc++) {
        for(int i_ic = 0; i_ic < input_c; i_ic++) {
            for(int i_h = 0; i_h < height; i_h++) {
                for(int i_w = 0; i_w < width; i_w++) {
                    int dst_i = input_c*height*width*i_oc + height*width*i_ic + width*i_h + i_w;
                    int src_i = out_c*input_c*width*i_h + out_c*input_c*i_w + out_c*i_ic + i_oc;
                    CV_Assert(dst_i < total);
                    CV_Assert(src_i < total);
                   dstData[dst_i] = data[src_i];
                }
            }
        }
    }
}

void TFImporter::connect(const std::map<String, int>& layers_name_id_map, Net& network, const Pin& outPin,
             const int input_layer_id, const int input_blob_id)
{
    std::map<String, int>::const_iterator it = layers_name_id_map.find(outPin.name);
    if (it == layers_name_id_map.end())
        CV_Error(Error::StsError, "Input layer not found: " + outPin.name);

    std::vector<String>::iterator inpNameIt = std::find(netInputsNames.begin(), netInputsNames.end(), outPin.name);
    int blobIndex;
    if (inpNameIt == netInputsNames.end())
        blobIndex = outPin.blobIndex;
    else
        blobIndex = inpNameIt - netInputsNames.begin();
    network.connect(it->second, blobIndex, input_layer_id, input_blob_id);
}

void TFImporter::connectToAllBlobs(const std::map<String, int>& layer_id, Net& network, const Pin& outPin,
                     const int input_layer_id, const int input_blobs_count)
{
    for (int input_blob_id = 0; input_blob_id < input_blobs_count; input_blob_id++)
        connect(layer_id, network, outPin, input_layer_id, input_blob_id);
}

const tensorflow::TensorProto& TFImporter::getConstBlob(const tensorflow::NodeDef &layer, std::map<String, int> const_layers,
                                              int input_blob_index, int* actual_inp_blob_idx) {
    if (input_blob_index == -1) {
        for(int i = 0; i < layer.input_size(); i++) {
            Pin input = parsePin(layer.input(i));
            if (const_layers.find(input.name) != const_layers.end()) {
                if (input_blob_index != -1)
                    CV_Error(Error::StsError, "More than one input is Const op");

                input_blob_index = i;
            }
        }
    }

    if (input_blob_index == -1)
        CV_Error(Error::StsError, "Const input blob for weights not found");

    Pin kernel_inp = parsePin(layer.input(input_blob_index));
    if (const_layers.find(kernel_inp.name) == const_layers.end())
        CV_Error(Error::StsError, "Const kernel input not found");
    if (kernel_inp.blobIndex != 0)
        CV_Error(Error::StsError, "Unsupported kernel input");

    if(actual_inp_blob_idx) {
        *actual_inp_blob_idx = input_blob_index;
    }

    int nodeIdx = const_layers.at(kernel_inp.name);
    if (nodeIdx < netBin.node_size() && netBin.node(nodeIdx).name() == kernel_inp.name)
    {
        return netBin.node(nodeIdx).attr().at("value").tensor();
    }
    else
    {
        CV_Assert(nodeIdx < netTxt.node_size(),
                  netTxt.node(nodeIdx).name() == kernel_inp.name);
        return netTxt.node(nodeIdx).attr().at("value").tensor();
    }
}

static void addConstNodes(tensorflow::GraphDef& net, std::map<String, int>& const_layers,
                          std::set<String>& layers_to_ignore)
{
    for (int li = 0; li < net.node_size(); li++)
    {
        const tensorflow::NodeDef &layer = net.node(li);
        String name = layer.name();
        String type = layer.op();

        if (type == "Dequantize")
        {
            // Example of Dequantize node:
            //   name: "conv2d_1/bias"
            //   op: "Dequantize"
            //   input: "conv2d_1/bias_quantized_const" (tensor of dtype DT_QUINT8)
            //   input: "conv2d_1/bias_quantized_min"
            //   input: "conv2d_1/bias_quantized_max"
            //   attr { key: "T" value { type: DT_QUINT8 } }   (quantized type)
            //   attr { key: "mode" value { s: "MIN_FIRST" } } (quantization technique)
            CV_Assert(layer.input_size() == 3);
            for (int i = 0; i < 3; ++i)
                CV_Assert(const_layers.find(layer.input(i)) != const_layers.end());
            CV_Assert(hasLayerAttr(layer, "mode") &&
                      getLayerAttr(layer, "mode").s() == "MIN_FIRST");

            int tensorId = const_layers[layer.input(0)];
            int minId = const_layers[layer.input(1)];
            int maxId = const_layers[layer.input(2)];

            tensorflow::TensorProto* tensor = net.mutable_node(tensorId)
                                                ->mutable_attr()->at("value")
                                                 .mutable_tensor();
            CV_Assert(tensor->dtype() == tensorflow::DT_QUINT8);

            Mat qMin = getTensorContent(net.node(minId).attr().at("value").tensor());
            Mat qMax = getTensorContent(net.node(maxId).attr().at("value").tensor());
            CV_Assert(qMin.total() == 1, qMin.type() == CV_32FC1,
                      qMax.total() == 1, qMax.type() == CV_32FC1);

            Mat content = getTensorContent(*tensor);

            float minVal = qMin.at<float>(0);
            float rangeScale = (qMax.at<float>(0) - minVal) / 255;
            CV_Assert(rangeScale >= 0);
            content.convertTo(content, CV_32FC1, rangeScale,
                              rangeScale * cvRound(minVal / rangeScale));

            tensor->set_dtype(tensorflow::DT_FLOAT);
            tensor->set_tensor_content(content.data, content.total() * content.elemSize1());

            net.mutable_node(tensorId)->set_name(name);
            CV_Assert(const_layers.insert(std::make_pair(name, tensorId)).second);
            layers_to_ignore.insert(name);
            continue;
        }
        else if (type != "Const")
            continue;  // only Const parameters are supported

        if (layer.attr().find("value") != layer.attr().end())
        {
            CV_Assert(const_layers.insert(std::make_pair(name, li)).second);
        }
        layers_to_ignore.insert(name);
    }
}

static int getDataLayout(const tensorflow::NodeDef& layer)
{
    if (hasLayerAttr(layer, "data_format"))
    {
        std::string format = getLayerAttr(layer, "data_format").s();
        if (format == "NHWC" || format == "channels_last")
            return DATA_LAYOUT_NHWC;
        else if (format == "NCHW" || format == "channels_first")
            return DATA_LAYOUT_NCHW;
        else
            CV_Error(Error::StsParseError, "Unknown data_format value: " + format);
    }
    return DATA_LAYOUT_UNKNOWN;
}

static inline std::string getNodeName(const std::string& tensorName)
{
    return tensorName.substr(0, tensorName.rfind(':'));
}

// If all inputs of specific layer have the same data layout we can say that
// this layer's output has this data layout too. Returns DATA_LAYOUT_UNKNOWN otherwise.
static int predictOutputDataLayout(const tensorflow::GraphDef& net,
                                   const tensorflow::NodeDef& layer,
                                   const std::map<String, int>& data_layouts)
{
    int layout = getDataLayout(layer);
    if (layout != DATA_LAYOUT_UNKNOWN)
        return layout;

    // Determine layout by layer's inputs
    std::map<String, int>::const_iterator it;
    for (int i = 0, n = layer.input_size(); i < n; ++i)
    {
        it = data_layouts.find(getNodeName(layer.input(i)));
        if (it != data_layouts.end())
        {
            if (layout != DATA_LAYOUT_UNKNOWN)
            {
                if (it->second != layout && it->second != DATA_LAYOUT_UNKNOWN)
                    return DATA_LAYOUT_UNKNOWN;
            }
            else
                layout = it->second;
        }
    }

    if (layout != DATA_LAYOUT_UNKNOWN)
        return layout;

    // Determine layout by layer's consumers recursively.
    it = data_layouts.find(layer.name());
    CV_Assert(it != data_layouts.end());
    return it->second;
}

void TFImporter::populateNet(Net dstNet)
{
    RemoveIdentityOps(netBin);
    RemoveIdentityOps(netTxt);

    if (!netTxt.ByteSize())
        simplifySubgraphs(netBin);

    std::set<String> layers_to_ignore;

    tensorflow::GraphDef& net = netTxt.ByteSize() != 0 ? netTxt : netBin;

    int layersSize = net.node_size();

    std::map<String, int> data_layouts;
    // Pre-fill data layouts where they are set explicitly.
    // Assuming that nodes are in topological order
    for (int i = net.node_size() - 1; i >= 0; --i)
    {
        const tensorflow::NodeDef& layer = net.node(i);
        std::string name = layer.name();

        int layout = getDataLayout(layer);
        std::map<String, int>::iterator it = data_layouts.find(name);
        if (it != data_layouts.end())
        {
            if (layout != DATA_LAYOUT_UNKNOWN)
            {
                if (it->second == DATA_LAYOUT_UNKNOWN)
                    it->second = layout;
                else if (it->second != layout)
                {
                    it->second = DATA_LAYOUT_UNKNOWN;
                    layout = DATA_LAYOUT_UNKNOWN;
                }
            }
            else
                layout = it->second;
        }
        else
            data_layouts[name] = layout;

        // Specify input layers to have the same data layout.
        for (int j = 0; j < layer.input_size(); ++j)
        {
            name = getNodeName(layer.input(j));
            it = data_layouts.find(name);
            if (it != data_layouts.end())
            {
                if (layout != DATA_LAYOUT_UNKNOWN)
                {
                    if (it->second == DATA_LAYOUT_UNKNOWN)
                        it->second = layout;
                    else if (it->second != layout)
                        it->second = DATA_LAYOUT_UNKNOWN;
                }
            }
            else
                data_layouts[name] = layout;
        }
    }

    // find all Const layers for params
    std::map<String, int> value_id;
    addConstNodes(netBin, value_id, layers_to_ignore);
    addConstNodes(netTxt, value_id, layers_to_ignore);

    std::map<String, int> layer_id;

    for (int li = 0; li < layersSize; li++)
    {
        tensorflow::NodeDef layer = net.node(li);
        String name = layer.name();
        String type = layer.op();
        LayerParams layerParams;

        if(layers_to_ignore.find(name) != layers_to_ignore.end())
            continue;

        int predictedLayout = predictOutputDataLayout(net, layer, data_layouts);
        data_layouts[name] = predictedLayout;

        if (type == "Conv2D" || type == "SpaceToBatchND" || type == "DepthwiseConv2dNative")
        {
            // The first node of dilated convolution subgraph.
            // Extract input node, dilation rate and paddings.
            std::string input = layer.input(0);
            if (type == "SpaceToBatchND")
            {
                // op: "SpaceToBatchND"
                // input: "input"
                // input: "SpaceToBatchND/block_shape"
                // input: "SpaceToBatchND/paddings"
                CV_Assert(layer.input_size() == 3);

                DictValue dilation = parseDims(getConstBlob(layer, value_id, 1));
                CV_Assert(dilation.size() == 2);
                layerParams.set("dilation_h", dilation.get<int>(0));
                layerParams.set("dilation_w", dilation.get<int>(1));

                Mat paddings;
                parseTensor<int>(getConstBlob(layer, value_id, 2), paddings);

                // paddings is a 2x2 matrix: [[top, bot], [left, right]]
                layerParams.set("pad_h", paddings.at<float>(0));
                layerParams.set("pad_w", paddings.at<float>(2));

                StrIntVector next_layers = getNextLayers(net, name, "Conv2D");
                if (next_layers.empty())
                {
                    next_layers = getNextLayers(net, name, "DepthwiseConv2dNative");
                }
                CV_Assert(next_layers.size() == 1);
                layer = net.node(next_layers[0].second);
                layers_to_ignore.insert(next_layers[0].first);
                name = layer.name();
                type = layer.op();
            }

            layerParams.set("bias_term", false);
            layerParams.blobs.resize(1);

            StrIntVector next_layers = getNextLayers(net, name, "BiasAdd");
            if (next_layers.size() == 1) {
                layerParams.set("bias_term", true);
                layerParams.blobs.resize(2);

                int weights_layer_index = next_layers[0].second;

                blobFromTensor(getConstBlob(net.node(weights_layer_index), value_id), layerParams.blobs[1]);
                ExcludeLayer(net, weights_layer_index, 0, false);
                layers_to_ignore.insert(next_layers[0].first);
            }

            const tensorflow::TensorProto& kernelTensor = getConstBlob(layer, value_id);
            kernelFromTensor(kernelTensor, layerParams.blobs[0]);
            releaseTensor(const_cast<tensorflow::TensorProto*>(&kernelTensor));
            int* kshape = layerParams.blobs[0].size.p;
            if (type == "DepthwiseConv2dNative")
            {
                const int chMultiplier = kshape[0];
                const int inCh = kshape[1];
                const int height = kshape[2];
                const int width = kshape[3];

                Mat copy = layerParams.blobs[0].clone();
                float* src = (float*)copy.data;
                float* dst = (float*)layerParams.blobs[0].data;
                for (int i = 0; i < chMultiplier; ++i)
                    for (int j = 0; j < inCh; ++j)
                        for (int s = 0; s < height * width; ++s)
                            {
                                int src_i = (i * inCh + j) * height * width + s;
                                int dst_i = (j * chMultiplier + i) * height* width + s;
                                dst[dst_i] = src[src_i];
                            }
                // TODO Use reshape instead
                kshape[0] = inCh * chMultiplier;
                kshape[1] = 1;
                size_t* kstep = layerParams.blobs[0].step.p;
                kstep[0] = kstep[1]; // fix steps too
            }
            layerParams.set("kernel_h", kshape[2]);
            layerParams.set("kernel_w", kshape[3]);
            layerParams.set("num_output", kshape[0]);

            setStrides(layerParams, layer);
            setPadding(layerParams, layer);

            // The final node of dilated convolution subgraph.
            next_layers = getNextLayers(net, name, "BatchToSpaceND");
            if (!next_layers.empty())
            {
                layerParams.set("pad_mode", "");  // We use padding values.
                CV_Assert(next_layers.size() == 1);
                ExcludeLayer(net, next_layers[0].second, 0, false);
                layers_to_ignore.insert(next_layers[0].first);
            }

            int id = dstNet.addLayer(name, "Convolution", layerParams);
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, parsePin(input), id, 0);

            if (data_layouts[name] == DATA_LAYOUT_UNKNOWN)
                data_layouts[name] = DATA_LAYOUT_NHWC;
        }
        else if (type == "BiasAdd" || type == "Add")
        {
            bool haveConst = false;
            for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)
            {
                Pin input = parsePin(layer.input(ii));
                haveConst = value_id.find(input.name) != value_id.end();
            }
            CV_Assert(!haveConst || layer.input_size() == 2);

            if (haveConst)
            {
                Mat values = getTensorContent(getConstBlob(layer, value_id));
                CV_Assert(values.type() == CV_32FC1);

                int id;
                if (values.total() == 1)  // is a scalar.
                {
                    layerParams.set("shift", values.at<float>(0));
                    id = dstNet.addLayer(name, "Power", layerParams);
                }
                else  // is a vector
                {
                    layerParams.blobs.resize(1, values);
                    id = dstNet.addLayer(name, "Shift", layerParams);
                }
                layer_id[name] = id;

                // one input only
                connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
            }
            else
            {
                layerParams.set("operation", "sum");
                int id = dstNet.addLayer(name, "Eltwise", layerParams);
                layer_id[name] = id;

                for (int ii = 0; ii < layer.input_size(); ii++)
                {
                    Pin inp = parsePin(layer.input(ii));
                    if (layer_id.find(inp.name) == layer_id.end())
                        CV_Error(Error::StsError, "Input layer not found: " + inp.name);
                    connect(layer_id, dstNet, inp, id, ii);
                }
            }
        }
        else if (type == "Sub")
        {
            bool haveConst = false;
            for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)
            {
                Pin input = parsePin(layer.input(ii));
                haveConst = value_id.find(input.name) != value_id.end();
            }
            CV_Assert(haveConst);

            Mat values = getTensorContent(getConstBlob(layer, value_id));
            CV_Assert(values.type() == CV_32FC1);
            values *= -1.0f;

            int id;
            if (values.total() == 1)  // is a scalar.
            {
                layerParams.set("shift", values.at<float>(0));
                id = dstNet.addLayer(name, "Power", layerParams);
            }
            else  // is a vector
            {
                layerParams.blobs.resize(1, values);
                id = dstNet.addLayer(name, "Shift", layerParams);
            }
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
        }
        else if (type == "MatMul")
        {
            CV_Assert(layer.input_size() == 2);

            layerParams.set("bias_term", false);
            layerParams.blobs.resize(1);

            StrIntVector next_layers = getNextLayers(net, name, "BiasAdd");
            if (next_layers.empty())
            {
                next_layers = getNextLayers(net, name, "Add");
            }
            if (next_layers.size() == 1) {
                layerParams.set("bias_term", true);
                layerParams.blobs.resize(2);

                int weights_layer_index = next_layers[0].second;
                blobFromTensor(getConstBlob(net.node(weights_layer_index), value_id), layerParams.blobs[1]);
                ExcludeLayer(net, weights_layer_index, 0, false);
                layers_to_ignore.insert(next_layers[0].first);
            }

            int kernel_blob_index = -1;
            const tensorflow::TensorProto& kernelTensor = getConstBlob(layer, value_id, -1, &kernel_blob_index);
            blobFromTensor(kernelTensor, layerParams.blobs[0]);
            releaseTensor(const_cast<tensorflow::TensorProto*>(&kernelTensor));

            if (kernel_blob_index == 1) { // In this case output is computed by x*W formula - W should be transposed
                Mat data = layerParams.blobs[0].t();
                layerParams.blobs[0] = data.clone();
            }

            layerParams.set("num_output", layerParams.blobs[0].size[0]);

            int id = dstNet.addLayer(name, "InnerProduct", layerParams);
            layer_id[name] = id;

            // one input only
            int input_blob_index = kernel_blob_index == 0 ? 1 : 0;
            connect(layer_id, dstNet, parsePin(layer.input(input_blob_index)), id, 0);
            data_layouts[name] = DATA_LAYOUT_UNKNOWN;
        }
        else if (type == "Reshape")
        {
            Pin inpId = parsePin(layer.input(0));
            Mat newShape = getTensorContent(getConstBlob(layer, value_id, 1));

            if (newShape.total() != 4 && data_layouts[layer.input(0)] == DATA_LAYOUT_NHWC)
            {
                LayerParams permLP;
                int order[] = {0, 2, 3, 1};  // From OpenCV's NCHW to NHWC.
                permLP.set("order", DictValue::arrayInt<int*>(order, 4));

                std::string permName = name + "/nchw";
                CV_Assert(layer_id.find(permName) == layer_id.end());
                int permId = dstNet.addLayer(permName, "Permute", permLP);
                layer_id[permName] = permId;
                connect(layer_id, dstNet, inpId, permId, 0);
                inpId = Pin(permName);
            }
            else if (newShape.total() == 4 && data_layouts[layer.input(0)] == DATA_LAYOUT_NHWC)
            {
                // NHWC->NCHW
                std::swap(*newShape.ptr<int32_t>(0, 2), *newShape.ptr<int32_t>(0, 3));
                std::swap(*newShape.ptr<int32_t>(0, 1), *newShape.ptr<int32_t>(0, 2));
            }
            layerParams.set("dim", DictValue::arrayInt<int*>(newShape.ptr<int>(), newShape.total()));

            int id = dstNet.addLayer(name, "Reshape", layerParams);
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, inpId, id, 0);
            data_layouts[name] = DATA_LAYOUT_UNKNOWN;
        }
        else if (type == "Flatten" || type == "Squeeze")
        {
            Pin inpId = parsePin(layer.input(0));
            int inpLayout = data_layouts[layer.input(0)];
            if (type == "Squeeze")
            {
                CV_Assert(hasLayerAttr(layer, "squeeze_dims"));
                const tensorflow::AttrValue& dims = getLayerAttr(layer, "squeeze_dims");
                if (inpLayout == DATA_LAYOUT_NHWC)
                {
                    if (dims.list().i_size() != 2 || dims.list().i(0) != 1 || dims.list().i(1) != 2)
                        CV_Error(Error::StsNotImplemented, "Unsupported squeeze configuration");
                }
                else if (inpLayout == DATA_LAYOUT_NCHW)
                {
                    if (dims.list().i_size() != 2 || dims.list().i(0) != 2 || dims.list().i(1) != 3)
                        CV_Error(Error::StsNotImplemented, "Unsupported squeeze configuration");
                }
                else
                    CV_Error(Error::StsNotImplemented, "Unsupported squeeze configuration");
            }
            if (inpLayout == DATA_LAYOUT_NHWC)
            {
                LayerParams permLP;
                int order[] = {0, 2, 3, 1};  // From OpenCV's NCHW to NHWC.
                permLP.set("order", DictValue::arrayInt<int*>(order, 4));

                std::string permName = name + "/nchw";
                CV_Assert(layer_id.find(permName) == layer_id.end());
                int permId = dstNet.addLayer(permName, "Permute", permLP);
                layer_id[permName] = permId;
                connect(layer_id, dstNet, inpId, permId, 0);
                inpId = Pin(permName);
            }
            int id = dstNet.addLayer(name, "Flatten", layerParams);
            layer_id[name] = id;
            connect(layer_id, dstNet, inpId, id, 0);
            data_layouts[name] = DATA_LAYOUT_UNKNOWN;
        }
        else if (type == "Transpose")
        {
            Mat perm = getTensorContent(getConstBlob(layer, value_id, 1));
            CV_Assert(perm.type() == CV_32SC1);
            int* permData = (int*)perm.data;
            if (perm.total() == 4)
            {
                // Only NHWC <-> NCHW permutations are allowed. OpenCV is always
                // keep NCHW layout this way.
                if (data_layouts[layer.input(0)] == DATA_LAYOUT_NHWC)
                {
                    if (permData[0] == 0 && permData[1] == 3 && permData[2] == 1 && permData[3] == 2)
                    {
                        // in TensorFlow: NHWC->NCHW
                        // in OpenCV: NCHW->NCHW
                        data_layouts[name] = DATA_LAYOUT_NCHW;
                    }
                    else if (permData[0] == 0 && permData[1] == 1 && permData[2] == 2 && permData[3] == 3)
                    {
                        // in TensorFlow: NHWC->NHWC
                        // in OpenCV: NCHW->NCHW
                        data_layouts[name] = DATA_LAYOUT_NHWC;
                    }
                    else
                        CV_Error(Error::StsParseError, "Only NHWC <-> NCHW permutations are allowed.");
                }
                else if (data_layouts[layer.input(0)] == DATA_LAYOUT_NCHW)
                {
                    if (permData[0] == 0 && permData[1] == 2 && permData[2] == 3 && permData[3] == 1)
                    {
                        // in TensorFlow: NCHW->NHWC
                        // in OpenCV: NCHW->NCHW
                        data_layouts[name] = DATA_LAYOUT_NHWC;
                    }
                    else if (permData[0] == 0 && permData[1] == 1 && permData[2] == 2 && permData[3] == 3)
                    {
                        // in TensorFlow: NCHW->NCHW
                        // in OpenCV: NCHW->NCHW
                        data_layouts[name] = DATA_LAYOUT_NCHW;
                    }
                    else
                        CV_Error(Error::StsParseError, "Only NHWC <-> NCHW permutations are allowed.");
                }
                int id = dstNet.addLayer(name, "Identity", layerParams);
                layer_id[name] = id;
                connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
            }
            else
            {
                layerParams.set("order", DictValue::arrayInt<int*>(permData, perm.total()));

                int id = dstNet.addLayer(name, "Permute", layerParams);
                layer_id[name] = id;

                // one input only
                connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
                data_layouts[name] = DATA_LAYOUT_UNKNOWN;
            }
        }
        else if (type == "Const")
        {
        }
        else if (type == "LRN")
        {
            if(hasLayerAttr(layer, "alpha")) {
                layerParams.set("alpha", getLayerAttr(layer, "alpha").f());
            }
            if(hasLayerAttr(layer, "beta")) {
                layerParams.set("beta", getLayerAttr(layer, "beta").f());
            }
            if(hasLayerAttr(layer, "depth_radius")) {
                int radius = (int)getLayerAttr(layer, "depth_radius").i();
                layerParams.set("local_size", 2*radius + 1);
            }
            if(hasLayerAttr(layer, "bias")) {
                layerParams.set("bias", getLayerAttr(layer, "bias").f());
            }
            layerParams.set("norm_by_size", false);

            int id = dstNet.addLayer(name, "LRN", layerParams);
            layer_id[name] = id;

            connectToAllBlobs(layer_id, dstNet, parsePin(layer.input(0)), id, layer.input_size());
        }
        else if (type == "Concat" || type == "ConcatV2")
        {
            int axisId = (type == "Concat" ? 0 : layer.input_size() - 1);
            int axis = getConstBlob(layer, value_id, axisId).int_val().Get(0);

            if (data_layouts[name] == DATA_LAYOUT_NHWC)
                axis = toNCHW(axis);
            layerParams.set("axis", axis);

            int id = dstNet.addLayer(name, "Concat", layerParams);
            layer_id[name] = id;


            int from = (type == "Concat" ? 1 : 0);
            int to = (type == "Concat" ? layer.input_size() : layer.input_size() - 1);

            // input(0) or input(n-1) is concat_dim
            for (int ii = from; ii < to; ii++)
            {
                Pin inp = parsePin(layer.input(ii));
                if (layer_id.find(inp.name) == layer_id.end())
                    CV_Error(Error::StsError, "Input layer not found: " + inp.name);
                connect(layer_id, dstNet, inp, id, ii - from);
            }
        }
        else if (type == "MaxPool")
        {
            layerParams.set("pool", "max");

            setKSize(layerParams, layer);
            setStrides(layerParams, layer);
            setPadding(layerParams, layer);

            int id = dstNet.addLayer(name, "Pooling", layerParams);
            layer_id[name] = id;

            connectToAllBlobs(layer_id, dstNet, parsePin(layer.input(0)), id, layer.input_size());
        }
        else if (type == "AvgPool")
        {
            layerParams.set("pool", "ave");
            layerParams.set("ave_pool_padded_area", false);

            setKSize(layerParams, layer);
            setStrides(layerParams, layer);
            setPadding(layerParams, layer);

            int id = dstNet.addLayer(name, "Pooling", layerParams);
            layer_id[name] = id;

            connectToAllBlobs(layer_id, dstNet, parsePin(layer.input(0)), id, layer.input_size());
        }
        else if (type == "Placeholder")
        {
            if (!hasLayerAttr(layer, "dtype") ||
                getLayerAttr(layer, "dtype").type() != tensorflow::DT_BOOL)  // If input is not a train/test flag.
            {
                netInputsNames.push_back(name);
                layer_id[name] = 0;
            }
        }
        else if (type == "Split") {
            // TODO: determining axis index remapping by input dimensions order of input blob
            // TODO: slicing input may be Const op
            // TODO: slicing kernels for convolutions - in current implementation it is impossible
            // TODO: add parsing num of slices parameter
            CV_Assert(layer.input_size() == 2);
            // num_split
            // 1st blob is dims tensor
            int axis = getConstBlob(layer, value_id, 0).int_val().Get(0);
            layerParams.set("axis", toNCHW(axis));

            int id = dstNet.addLayer(name, "Slice", layerParams);
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, parsePin(layer.input(1)), id, 0);
        }
        else if (type == "Slice")
        {
            // op: "Slice"
            // input: "input_node"
            // input: "Slice/begin"
            // input: "Slice/size"
            CV_Assert(layer.input_size() == 3);
            Mat begins = getTensorContent(getConstBlob(layer, value_id, 1));
            Mat sizes = getTensorContent(getConstBlob(layer, value_id, 2));
            CV_Assert(!begins.empty(), !sizes.empty(), begins.type() == CV_32SC1,
                      sizes.type() == CV_32SC1);

            if (begins.total() == 4 && data_layouts[name] == DATA_LAYOUT_NHWC)
            {
                // Swap NHWC parameters' order to NCHW.
                std::swap(*begins.ptr<int32_t>(0, 2), *begins.ptr<int32_t>(0, 3));
                std::swap(*begins.ptr<int32_t>(0, 1), *begins.ptr<int32_t>(0, 2));
                std::swap(*sizes.ptr<int32_t>(0, 2), *sizes.ptr<int32_t>(0, 3));
                std::swap(*sizes.ptr<int32_t>(0, 1), *sizes.ptr<int32_t>(0, 2));
            }
            layerParams.set("begin", DictValue::arrayInt((int*)begins.data, begins.total()));
            layerParams.set("size", DictValue::arrayInt((int*)sizes.data, sizes.total()));

            int id = dstNet.addLayer(name, "Slice", layerParams);
            layer_id[name] = id;

            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
        }
        else if (type == "Mul")
        {
            bool haveConst = false;
            for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)
            {
                Pin input = parsePin(layer.input(ii));
                haveConst = value_id.find(input.name) != value_id.end();
            }
            CV_Assert(!haveConst || layer.input_size() == 2);

            if (haveConst)
            {
                // Multiplication by constant.
                CV_Assert(layer.input_size() == 2);
                Mat scaleMat = getTensorContent(getConstBlob(layer, value_id));
                CV_Assert(scaleMat.type() == CV_32FC1);

                int id;
                if (scaleMat.total() == 1)  // is a scalar.
                {
                    // Try to match with a LeakyRelu:
                    // node {
                    //   name: "LeakyRelu/mul"
                    //   op: "Mul"
                    //   input: "LeakyRelu/alpha"
                    //   input: "input"
                    // }
                    // node {
                    //   name: "LeakyRelu/Maximum"
                    //   op: "Maximum"
                    //   input: "LeakyRelu/mul"
                    //   input: "input"
                    // }
                    StrIntVector next_layers = getNextLayers(net, name, "Maximum");
                    if (!next_layers.empty())
                    {
                        int maximumLayerIdx = next_layers[0].second;
                        ExcludeLayer(net, maximumLayerIdx, 0, false);
                        layers_to_ignore.insert(next_layers[0].first);

                        layerParams.set("negative_slope", scaleMat.at<float>(0));
                        id = dstNet.addLayer(name, "ReLU", layerParams);
                    }
                    else
                    {
                        // Just a multiplication.
                        layerParams.set("scale", scaleMat.at<float>(0));
                        id = dstNet.addLayer(name, "Power", layerParams);
                    }
                }
                else  // is a vector
                {
                    layerParams.blobs.resize(1, scaleMat);

                   StrIntVector next_layers = getNextLayers(net, name, "Add");
                   if (!next_layers.empty())
                   {
                       layerParams.set("bias_term", true);
                       layerParams.blobs.resize(2);

                       int weights_layer_index = next_layers[0].second;
                       blobFromTensor(getConstBlob(net.node(weights_layer_index), value_id), layerParams.blobs.back());
                       ExcludeLayer(net, weights_layer_index, 0, false);
                       layers_to_ignore.insert(next_layers[0].first);
                   }

                    if (hasLayerAttr(layer, "axis"))
                        layerParams.set("axis", getLayerAttr(layer, "axis").i());

                    id = dstNet.addLayer(name, "Scale", layerParams);
                }
                layer_id[name] = id;

                Pin inp0 = parsePin(layer.input(0));
                if (layer_id.find(inp0.name) != layer_id.end())
                    // First operand is a constant.
                    connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
                else
                    connect(layer_id, dstNet, parsePin(layer.input(1)), id, 0);
            }
            else
            {
                layerParams.set("operation", "prod");
                int id = dstNet.addLayer(name, "Eltwise", layerParams);
                layer_id[name] = id;

                for (int ii = 0; ii < layer.input_size(); ii++)
                {
                    Pin inp = parsePin(layer.input(ii));
                    if (layer_id.find(inp.name) == layer_id.end())
                        CV_Error(Error::StsError, "Input layer not found: " + inp.name);
                    connect(layer_id, dstNet, inp, id, ii);
                }
            }
        }
        else if (type == "Pad")
        {
            Mat paddings = getTensorContent(getConstBlob(layer, value_id, 1));
            CV_Assert(paddings.type() == CV_32SC1);
            if (paddings.total() == 8)
            {
                // Perhabs, we have NHWC padding dimensions order.
                //  N    H    W    C
                // 0 1  2 3  4 5  6 7
                std::swap(*paddings.ptr<int32_t>(0, 2), *paddings.ptr<int32_t>(0, 6));
                std::swap(*paddings.ptr<int32_t>(0, 3), *paddings.ptr<int32_t>(0, 7));
                //  N    C    W    H
                // 0 1  2 3  4 5  6 7
                std::swap(*paddings.ptr<int32_t>(0, 4), *paddings.ptr<int32_t>(0, 6));
                std::swap(*paddings.ptr<int32_t>(0, 5), *paddings.ptr<int32_t>(0, 7));
                //  N    C    H    W
                // 0 1  2 3  4 5  6 7
            }
            layerParams.set("paddings", DictValue::arrayInt<int*>((int*)paddings.data, paddings.total()));

            int id = dstNet.addLayer(name, "Padding", layerParams);
            layer_id[name] = id;

            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
        }
        else if (type == "FusedBatchNorm")
        {
            // op: "FusedBatchNorm"
            // input: "input"
            // input: "BatchNorm/gamma"
            // input: "BatchNorm/beta"
            // input: "BatchNorm/moving_mean"
            // input: "BatchNorm/moving_variance"
            if (layer.input_size() != 5)
                CV_Error(Error::StsNotImplemented,
                         "Expected gamma, beta, mean and std");
            Pin inpId = parsePin(layer.input(0));

            bool isTraining = hasLayerAttr(layer, "is_training") && getLayerAttr(layer, "is_training").b();

            layerParams.blobs.resize(2);

            const tensorflow::TensorProto& gammaTensor = getConstBlob(layer, value_id, 1);
            if (!gammaTensor.tensor_content().empty())
            {
                layerParams.blobs.resize(layerParams.blobs.size() + 1);
                layerParams.set("has_weight", true);
                blobFromTensor(gammaTensor, layerParams.blobs.back());
            }
            else
                layerParams.set("has_weight", false);

            const tensorflow::TensorProto& betaTensor = getConstBlob(layer, value_id, 2);
            if (!betaTensor.tensor_content().empty())
            {
                layerParams.blobs.resize(layerParams.blobs.size() + 1);
                layerParams.set("has_bias", true);
                blobFromTensor(betaTensor, layerParams.blobs.back());
            }
            else
                layerParams.set("has_bias", false);

            Mat mean, std;
            if (isTraining)
            {
                if (layerParams.blobs.size() == 2)
                    CV_Error(Error::StsNotImplemented, "Cannot determine number "
                             "of parameters for batch normalization layer.");
                mean = Mat::zeros(1, layerParams.blobs[3].total(), CV_32F);
                std = Mat::ones(1, layerParams.blobs[3].total(), CV_32F);

                // Add an extra layer: Mean-Variance normalization
                LayerParams mvnParams;
                std::string mvnName = name + "/MVN";
                CV_Assert(layer_id.find(mvnName) == layer_id.end());
                int mvnId = dstNet.addLayer(mvnName, "MVN", mvnParams);
                layer_id[mvnName] = mvnId;
                connect(layer_id, dstNet, inpId, mvnId, 0);
                inpId = Pin(mvnName);
            }
            else
            {
                blobFromTensor(getConstBlob(layer, value_id, 3), mean);
                blobFromTensor(getConstBlob(layer, value_id, 4), std);
            }
            layerParams.blobs[0] = mean;
            layerParams.blobs[1] = std;

            if (hasLayerAttr(layer, "epsilon"))
                layerParams.set("eps", getLayerAttr(layer, "epsilon").f());

            int id = dstNet.addLayer(name, "BatchNorm", layerParams);
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, inpId, id, 0);
        }
        else if (type == "Conv2DBackpropInput")
        {
            // op: "Conv2DBackpropInput"
            // input: "conv2d_transpose/output_shape"
            // input: "weights"
            // input: "input"
            if (layer.input_size() != 3)
                CV_Error(Error::StsNotImplemented,
                         "Expected output shape, weights and input nodes");

            layerParams.set("bias_term", false);
            layerParams.blobs.resize(1);

            StrIntVector next_layers = getNextLayers(net, name, "BiasAdd");
            if (next_layers.size() == 1)
            {
                layerParams.set("bias_term", true);
                layerParams.blobs.resize(2);

                int weights_layer_index = next_layers[0].second;

                blobFromTensor(getConstBlob(net.node(weights_layer_index), value_id), layerParams.blobs[1]);
                ExcludeLayer(net, weights_layer_index, 0, false);
                layers_to_ignore.insert(next_layers[0].first);
            }

            kernelFromTensor(getConstBlob(layer, value_id, 1), layerParams.blobs[0]);

            const int* kshape = layerParams.blobs[0].size.p;
            const int kernelH = kshape[2];
            const int kernelW = kshape[3];
            layerParams.set("kernel_h", kernelH);
            layerParams.set("kernel_w", kernelW);
            layerParams.set("num_output", kshape[1]);

            setStrides(layerParams, layer);
            setPadding(layerParams, layer);

            // For convolution layer, output shape computes as
            // o = 1 + (i - k + 2*p) / s
            // i - input size, o - output size, k - kernel size, p - pad, s - stride
            // In TensorFlow, p == 0 is padMode == 'VALID' or p == (k - 1) / 2
            // considering that k is odd.
            // SAME:  o = 1 + (i - 1) / s
            // VALID: o = 1 + i / s
            // Deconvolution's layer output shape computes as
            // SAME:  o = 1 + (i - 1)*s
            // VALID: o = (i - 1)*s
            // If output_shape differs from formulas above then adjust padding is applied.

            const int strideY = layerParams.get<int>("stride_h");
            const int strideX = layerParams.get<int>("stride_w");
            Mat outShape = getTensorContent(getConstBlob(layer, value_id, 0));
            const int outH = outShape.at<int>(1);
            const int outW = outShape.at<int>(2);
            if (layerParams.get<String>("pad_mode") == "SAME")
            {
                layerParams.set("adj_w", (outW - 1) % strideX);
                layerParams.set("adj_h", (outH - 1) % strideY);
            }
            else if (layerParams.get<String>("pad_mode") == "VALID")
            {
                layerParams.set("adj_w", (outW - kernelW) % strideX);
                layerParams.set("adj_h", (outH - kernelH) % strideY);
            }
            int id = dstNet.addLayer(name, "Deconvolution", layerParams);
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, parsePin(layer.input(2)), id, 0);
        }
        else if (type == "BlockLSTM")
        {
            // op: "BlockLSTM"
            // input: "lstm_block_wrapper/ToInt64/x"  (ignore, number of time stamps)
            // input: "input"
            // input: "lstm_block_wrapper/zeros"      (ignore)
            // input: "lstm_block_wrapper/zeros"      (ignore)
            // input: "lstm_block_wrapper/kernel"
            // input: "lstm_block_wrapper/w_i_diag"
            // input: "lstm_block_wrapper/w_f_diag"
            // input: "lstm_block_wrapper/w_o_diag"
            // input: "lstm_block_wrapper/bias"
            if (layer.input_size() != 9)
                CV_Error(Error::StsNotImplemented, "Unexpected number of input nodes");

            if (hasLayerAttr(layer, "forget_bias"))
                layerParams.set("forget_bias", getLayerAttr(layer, "forget_bias").f());

            if (hasLayerAttr(layer, "forget_bias"))
            {
                float cellClip = getLayerAttr(layer, "cell_clip").f();
                // Cell clip disabled if it's negative.
                if (cellClip >= 0)
                {
                    layerParams.set("use_cell_clip", true);
                    layerParams.set("cell_clip", cellClip);
                }
            }

            Mat W, Wh, Wx, b;
            blobFromTensor(getConstBlob(layer, value_id, 4), W);
            blobFromTensor(getConstBlob(layer, value_id, 8), b);
            const int outSize = W.cols / 4;

            // IGFO->IFOG
            float* weightData = (float*)W.data;
            for (int i = 0; i < W.rows; ++i)
                for (int j = 0; j < outSize; ++j)
                {
                    std::swap(weightData[i * W.cols + 1 * outSize + j],
                              weightData[i * W.cols + 2 * outSize + j]);
                    std::swap(weightData[i * W.cols + 2 * outSize + j],
                              weightData[i * W.cols + 3 * outSize + j]);
                }
            Wx = W.rowRange(0, W.rows - outSize).t();
            Wh = W.rowRange(W.rows - outSize, W.rows).t();

            layerParams.blobs.resize(3);
            layerParams.blobs[0] = Wh;
            layerParams.blobs[1] = Wx;
            layerParams.blobs[2] = b;

            if (hasLayerAttr(layer, "use_peephole"))
            {
                bool usePeephole = getLayerAttr(layer, "use_peephole").b();
                if (usePeephole)
                {
                    layerParams.set("use_peephole", true);
                    layerParams.blobs.resize(6);
                    for (int i = 0; i < 3; ++i)
                    {
                        Mat w;
                        blobFromTensor(getConstBlob(layer, value_id, 5 + i), w);
                        w = w.reshape(1, w.total());  // Single column.
                        w = Mat::diag(w);  // Make a diagonal matrix.
                        layerParams.blobs[3 + i] = w;
                    }
                }
            }

            int id = dstNet.addLayer(name, "LSTM", layerParams);
            layer_id[name] = id;

            // one input only
            connect(layer_id, dstNet, parsePin(layer.input(1)), id, 0);
            data_layouts[name] = DATA_LAYOUT_UNKNOWN;
        }
        else if (type == "ResizeNearestNeighbor" || type == "ResizeBilinear")
        {
            if (layer.input_size() == 2)
            {
                Mat outSize = getTensorContent(getConstBlob(layer, value_id, 1));
                CV_Assert(outSize.type() == CV_32SC1, outSize.total() == 2);
                layerParams.set("height", outSize.at<int>(0, 0));
                layerParams.set("width", outSize.at<int>(0, 1));
            }
            else if (layer.input_size() == 3)
            {
                Mat factorHeight = getTensorContent(getConstBlob(layer, value_id, 1));
                Mat factorWidth = getTensorContent(getConstBlob(layer, value_id, 2));
                CV_Assert(factorHeight.type() == CV_32SC1, factorHeight.total() == 1,
                          factorWidth.type() == CV_32SC1, factorWidth.total() == 1);
                layerParams.set("zoom_factor_x", factorWidth.at<int>(0));
                layerParams.set("zoom_factor_y", factorHeight.at<int>(0));
            }
            else
                CV_Assert(layer.input_size() == 2 || layer.input_size() == 3);

            if (type == "ResizeNearestNeighbor")
                layerParams.set("interpolation", "nearest");
            else
                layerParams.set("interpolation", "bilinear");

            if (hasLayerAttr(layer, "align_corners"))
                layerParams.set("align_corners", getLayerAttr(layer, "align_corners").b());

            int id = dstNet.addLayer(name, "Resize", layerParams);
            layer_id[name] = id;

            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
        }
        else if (type == "L2Normalize")
        {
            // op: "L2Normalize"
            // input: "input"
            // input: "reduction_indices" (axis)
            CV_Assert(layer.input_size() == 2);
            Mat reductionIndices = getTensorContent(getConstBlob(layer, value_id, 1));
            CV_Assert(reductionIndices.type() == CV_32SC1);

            const int numAxes = reductionIndices.total();
            if (data_layouts[name] == DATA_LAYOUT_NHWC)
                for (int i = 0; i < numAxes; ++i)
                    reductionIndices.at<int>(i) = toNCHW(reductionIndices.at<int>(i));

            cv::sort(reductionIndices, reductionIndices, SORT_ASCENDING);
            for (int i = 1; i < numAxes; ++i)
            {
                CV_Assert(reductionIndices.at<int>(i) == reductionIndices.at<int>(i - 1) + 1);
                // Axes have the same sign.
                CV_Assert(reductionIndices.at<int>(i) * reductionIndices.at<int>(i - 1) >= 0);
            }
            layerParams.set("start_axis", reductionIndices.at<int>(0));
            layerParams.set("end_axis", reductionIndices.at<int>(numAxes - 1));

            int id = dstNet.addLayer(name, "Normalize", layerParams);
            layer_id[name] = id;
            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
        }
        else if (type == "PriorBox")
        {
            if (hasLayerAttr(layer, "min_size"))
                layerParams.set("min_size", getLayerAttr(layer, "min_size").i());
            if (hasLayerAttr(layer, "max_size"))
                layerParams.set("max_size", getLayerAttr(layer, "max_size").i());
            if (hasLayerAttr(layer, "flip"))
                layerParams.set("flip", getLayerAttr(layer, "flip").b());
            if (hasLayerAttr(layer, "clip"))
                layerParams.set("clip", getLayerAttr(layer, "clip").b());
            if (hasLayerAttr(layer, "offset"))
                layerParams.set("offset", getLayerAttr(layer, "offset").f());
            if (hasLayerAttr(layer, "step"))
                layerParams.set("step", getLayerAttr(layer, "step").f());

            const std::string paramNames[] = {"variance", "aspect_ratio", "scales",
                                              "width", "height"};
            for (int i = 0; i < 5; ++i)
            {
                if (hasLayerAttr(layer, paramNames[i]))
                {
                    Mat values = getTensorContent(getLayerAttr(layer, paramNames[i]).tensor());
                    layerParams.set(paramNames[i],
                                    DictValue::arrayReal<float*>((float*)values.data, values.total()));
                }
            }
            int id = dstNet.addLayer(name, "PriorBox", layerParams);
            layer_id[name] = id;
            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
            connect(layer_id, dstNet, parsePin(layer.input(1)), id, 1);
            data_layouts[name] = DATA_LAYOUT_UNKNOWN;
        }
        else if (type == "DetectionOutput")
        {
            // op: "DetectionOutput"
            // input_0: "locations"
            // input_1: "classifications"
            // input_2: "prior_boxes"
            if (hasLayerAttr(layer, "num_classes"))
                layerParams.set("num_classes", getLayerAttr(layer, "num_classes").i());
            if (hasLayerAttr(layer, "share_location"))
                layerParams.set("share_location", getLayerAttr(layer, "share_location").b());
            if (hasLayerAttr(layer, "background_label_id"))
                layerParams.set("background_label_id", getLayerAttr(layer, "background_label_id").i());
            if (hasLayerAttr(layer, "nms_threshold"))
                layerParams.set("nms_threshold", getLayerAttr(layer, "nms_threshold").f());
            if (hasLayerAttr(layer, "top_k"))
                layerParams.set("top_k", getLayerAttr(layer, "top_k").i());
            if (hasLayerAttr(layer, "code_type"))
                layerParams.set("code_type", getLayerAttr(layer, "code_type").s());
            if (hasLayerAttr(layer, "keep_top_k"))
                layerParams.set("keep_top_k", getLayerAttr(layer, "keep_top_k").i());
            if (hasLayerAttr(layer, "confidence_threshold"))
                layerParams.set("confidence_threshold", getLayerAttr(layer, "confidence_threshold").f());
            if (hasLayerAttr(layer, "loc_pred_transposed"))
                layerParams.set("loc_pred_transposed", getLayerAttr(layer, "loc_pred_transposed").b());
            if (hasLayerAttr(layer, "clip"))
                layerParams.set("clip", getLayerAttr(layer, "clip").b());
            if (hasLayerAttr(layer, "variance_encoded_in_target"))
                layerParams.set("variance_encoded_in_target", getLayerAttr(layer, "variance_encoded_in_target").b());

            int id = dstNet.addLayer(name, "DetectionOutput", layerParams);
            layer_id[name] = id;
            for (int i = 0; i < 3; ++i)
                connect(layer_id, dstNet, parsePin(layer.input(i)), id, i);
            data_layouts[name] = DATA_LAYOUT_UNKNOWN;
        }
        else if (type == "Softmax")
        {
            if (hasLayerAttr(layer, "axis"))
                layerParams.set("axis", getLayerAttr(layer, "axis").i());

            int id = dstNet.addLayer(name, "Softmax", layerParams);
            layer_id[name] = id;
            connectToAllBlobs(layer_id, dstNet, parsePin(layer.input(0)), id, layer.input_size());
        }
        else if (type == "CropAndResize")
        {
            // op: "CropAndResize"
            // input: "input"
            // input: "boxes"
            // input: "sizes"
            CV_Assert(layer.input_size() == 3);

            Mat cropSize = getTensorContent(getConstBlob(layer, value_id, 2));
            CV_Assert(cropSize.type() == CV_32SC1, cropSize.total() == 2);

            layerParams.set("height", cropSize.at<int>(0));
            layerParams.set("width", cropSize.at<int>(1));

            int id = dstNet.addLayer(name, "CropAndResize", layerParams);
            layer_id[name] = id;

            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
            connect(layer_id, dstNet, parsePin(layer.input(1)), id, 1);
        }
        else if (type == "Mean")
        {
            Mat indices = getTensorContent(getConstBlob(layer, value_id, 1));
            CV_Assert(indices.type() == CV_32SC1);

            if (indices.total() != 2 || indices.at<int>(0) != 1 || indices.at<int>(1) != 2)
                CV_Error(Error::StsNotImplemented, "Unsupported mode of reduce_mean operation.");

            layerParams.set("pool", "ave");
            layerParams.set("global_pooling", true);

            int id = dstNet.addLayer(name, "Pooling", layerParams);
            layer_id[name] = id;

            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);

            // There are two attributes, "keepdims" and a deprecated "keep_dims".
            bool keepDims = false;
            if (hasLayerAttr(layer, "keepdims"))
                keepDims = getLayerAttr(layer, "keepdims").b();
            else if (hasLayerAttr(layer, "keep_dims"))
                keepDims = getLayerAttr(layer, "keep_dims").b();

            if (!keepDims)
            {
                LayerParams flattenLp;
                std::string flattenName = name + "/flatten";
                CV_Assert(layer_id.find(flattenName) == layer_id.end());
                int flattenId = dstNet.addLayer(flattenName, "Flatten", flattenLp);
                layer_id[flattenName] = flattenId;
                connect(layer_id, dstNet, Pin(name), flattenId, 0);
            }
        }
        else if (type == "ClipByValue")
        {
            // op: "ClipByValue"
            // input: "input"
            // input: "mix"
            // input: "max"
            CV_Assert(layer.input_size() == 3);

            Mat minValue = getTensorContent(getConstBlob(layer, value_id, 1));
            Mat maxValue = getTensorContent(getConstBlob(layer, value_id, 2));
            CV_Assert(minValue.total() == 1, minValue.type() == CV_32F,
                      maxValue.total() == 1, maxValue.type() == CV_32F);

            layerParams.set("min_value", minValue.at<float>(0));
            layerParams.set("max_value", maxValue.at<float>(0));

            int id = dstNet.addLayer(name, "ReLU6", layerParams);
            layer_id[name] = id;

            connect(layer_id, dstNet, parsePin(layer.input(0)), id, 0);
        }
        else if (type == "Abs" || type == "Tanh" || type == "Sigmoid" ||
                 type == "Relu" || type == "Elu" ||
                 type == "Identity" || type == "Relu6")
        {
            std::string dnnType = type;
            if (type == "Abs") dnnType = "AbsVal";
            else if (type == "Tanh") dnnType = "TanH";
            else if (type == "Relu") dnnType = "ReLU";
            else if (type == "Relu6") dnnType = "ReLU6";
            else if (type == "Elu") dnnType = "ELU";

            int id = dstNet.addLayer(name, dnnType, layerParams);
            layer_id[name] = id;
            connectToAllBlobs(layer_id, dstNet, parsePin(layer.input(0)), id, layer.input_size());
        }
        else
        {
            // Importer does not know how to map this TensorFlow's operation onto OpenCV's layer.
            // However we create a layer with the same type and rely that user defined a custom layer.

            // All the attributes are added to LayerParams.
            google::protobuf::Map<std::string, tensorflow::AttrValue> attr = layer.attr();
            for (google::protobuf::Map<std::string, tensorflow::AttrValue>::const_iterator ai = attr.begin();
                 ai != attr.end(); ++ai)
            {
                if (ai->second.value_case() == tensorflow::AttrValue::kS)  // string
                    layerParams.set(ai->first, ai->second.s());
                if (ai->second.value_case() == tensorflow::AttrValue::kI)  // int64
                    layerParams.set(ai->first, ai->second.i());
                if (ai->second.value_case() == tensorflow::AttrValue::kF)  // float
                    layerParams.set(ai->first, ai->second.f());
                if (ai->second.value_case() == tensorflow::AttrValue::kB)  // bool
                    layerParams.set(ai->first, ai->second.b());
            }

            // All the Const input nodes are added to layer's blobs.
            std::vector<std::string> inputsNames;
            for (int i = 0; i < layer.input_size(); ++i)
            {
                // Check if input is a Const node.
                if (value_id.find(layer.input(i)) != value_id.end())
                {
                    Mat blob = getTensorContent(getConstBlob(layer, value_id, i));
                    layerParams.blobs.push_back(blob);
                }
                else
                    inputsNames.push_back(layer.input(i));
            }
            int id = dstNet.addLayer(name, type, layerParams);
            layer_id[name] = id;

            for (int i = 0; i < inputsNames.size(); ++i)
            {
                connect(layer_id, dstNet, parsePin(inputsNames[i]), id, i);
            }
        }
    }
    dstNet.setInputsNames(netInputsNames);
}

} // namespace

#endif //HAVE_PROTOBUF

Net readNetFromTensorflow(const String &model, const String &config)
{
    TFImporter importer(model.c_str(), config.c_str());
    Net net;
    importer.populateNet(net);
    return net;
}

Net readNetFromTensorflow(const char* bufferModel, size_t lenModel,
                          const char* bufferConfig, size_t lenConfig)
{
    TFImporter importer(bufferModel, lenModel, bufferConfig, lenConfig);
    Net net;
    importer.populateNet(net);
    return net;
}

CV__DNN_EXPERIMENTAL_NS_END
}} // namespace