Introduction

Deep learning is a fast growing area. The new approaches to build neural networks usually introduce new types of layers. They could be modifications of existing ones or implement outstanding researching ideas.

OpenCV gives an opportunity to import and run networks from different deep learning frameworks. There are a number of the most popular layers. However you can face a problem that your network cannot be imported using OpenCV because of unimplemented layers.

The first solution is to create a feature request at https://github.com/opencv/opencv/issues mentioning details such a source of model and type of new layer. A new layer could be implemented if OpenCV community shares this need.

The second way is to define a custom layer so OpenCV's deep learning engine will know how to use it. This tutorial is dedicated to show you a process of deep learning models import customization.

Define a custom layer in C++

Deep learning layer is a building block of network's pipeline. It has connections to input blobs and produces results to output blobs. There are trained weights and hyper-parameters. Layers' names, types, weights and hyper-parameters are stored in files are generated by native frameworks during training. If OpenCV mets unknown layer type it throws an exception trying to read a model:

Unspecified error: Can't create layer "layer_name" of type "MyType" in function getLayerInstance

To import the model correctly you have to derive a class from cv::dnn::Layer with the following methods:

class MyLayer : public cv::dnn::Layer
{
public:
    MyLayer(const cv::dnn::LayerParams &params);
    static cv::Ptr<cv::dnn::Layer> create(cv::dnn::LayerParams& params);
    virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
                                 const int requiredOutputs,
                                 std::vector<std::vector<int> > &outputs,
                                 std::vector<std::vector<int> > &internals) const CV_OVERRIDE;
    virtual void forward(cv::InputArrayOfArrays inputs,
                         cv::OutputArrayOfArrays outputs,
                         cv::OutputArrayOfArrays internals) CV_OVERRIDE;
    virtual void finalize(cv::InputArrayOfArrays inputs,
                          cv::OutputArrayOfArrays outputs) CV_OVERRIDE;
};

And register it before the import:

#include <opencv2/dnn/layer.details.hpp>  // CV_DNN_REGISTER_LAYER_CLASS
static inline void loadNet()
{
    CV_DNN_REGISTER_LAYER_CLASS(Interp, InterpLayer);
    // ...

Note: MyType is a type of unimplemented layer from the thrown exception.

Let's see what all the methods do:

Constructor

MyLayer(const cv::dnn::LayerParams &params);

Retrieves hyper-parameters from cv::dnn::LayerParams. If your layer has trainable weights they will be already stored in the Layer's member cv::dnn::Layer::blobs.

A static method create

static cv::Ptr<cv::dnn::Layer> create(cv::dnn::LayerParams& params);

This method should create an instance of you layer and return cv::Ptr with it.

Output blobs' shape computation

    virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
                                 const int requiredOutputs,
                                 std::vector<std::vector<int> > &outputs,
                                 std::vector<std::vector<int> > &internals) const CV_OVERRIDE;

Returns layer's output shapes depends on input shapes. You may request an extra memory using internals.

Run a layer

    virtual void forward(cv::InputArrayOfArrays inputs,
                         cv::OutputArrayOfArrays outputs,
                         cv::OutputArrayOfArrays internals) CV_OVERRIDE;

Implement a layer's logic here. Compute outputs for given inputs.

Note: OpenCV manages memory allocated for layers. In the most cases the same memory can be reused between layers. So your forward implementation should not rely that the second invocation of forward will has the same data at outputs and internals.

Optional finalize method

virtual void finalize(cv::InputArrayOfArrays inputs,

cv::OutputArrayOfArrays outputs) CV_OVERRIDE;

The chain of methods are the following: OpenCV deep learning engine calls create method once then it calls getMemoryShapes for an every created layer then you can make some preparations depends on known input dimensions at cv::dnn::Layer::finalize. After network was initialized only forward method is called for an every network's input.

Note: Varying input blobs' sizes such height or width or batch size you make OpenCV reallocate all the internal memory. That leads efficiency gaps. Try to initialize and deploy models using a fixed batch size and image's dimensions.

Example: custom layer from Caffe

Let's create a custom layer Interp from https://github.com/cdmh/deeplab-public. It's just a simple resize that takes an input blob of size N x C x Hi x Wi and returns an output blob of size N x C x Ho x Wo where N is a batch size, C is a number of channels, Hi x Wi and Ho x Wo are input and output height x width correspondingly. This layer has no trainable weights but it has hyper-parameters to specify an output size.

In example,

layer {
  name: "output"
  type: "Interp"
  bottom: "input"
  top: "output"
  interp_param {
    height: 9
    width: 8
  }
}

This way our implementation can look like:

class InterpLayer : public cv::dnn::Layer
{
public:
    InterpLayer(const cv::dnn::LayerParams &params) : Layer(params)
    {
        outWidth = params.get<int>("width", 0);
        outHeight = params.get<int>("height", 0);
    }
    static cv::Ptr<cv::dnn::Layer> create(cv::dnn::LayerParams& params)
    {
        return cv::Ptr<cv::dnn::Layer>(new InterpLayer(params));
    }
    virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
                                 const int requiredOutputs,
                                 std::vector<std::vector<int> > &outputs,
                                 std::vector<std::vector<int> > &internals) const CV_OVERRIDE
    {
        CV_UNUSED(requiredOutputs); CV_UNUSED(internals);
        std::vector<int> outShape(4);
        outShape[0] = inputs[0][0];  // batch size
        outShape[1] = inputs[0][1];  // number of channels
        outShape[2] = outHeight;
        outShape[3] = outWidth;
        outputs.assign(1, outShape);
        return false;
    }
    // Implementation of this custom layer is based on https://github.com/cdmh/deeplab-public/blob/master/src/caffe/layers/interp_layer.cpp
    virtual void forward(cv::InputArrayOfArrays inputs_arr,
                         cv::OutputArrayOfArrays outputs_arr,
                         cv::OutputArrayOfArrays internals_arr) CV_OVERRIDE
    {
        if (inputs_arr.depth() == CV_16S)
        {
            // In case of DNN_TARGET_OPENCL_FP16 target the following method
            // converts data from FP16 to FP32 and calls this forward again.
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }
        std::vector<cv::Mat> inputs, outputs;
        inputs_arr.getMatVector(inputs);
        outputs_arr.getMatVector(outputs);
        cv::Mat& inp = inputs[0];
        cv::Mat& out = outputs[0];
        const float* inpData = (float*)inp.data;
        float* outData = (float*)out.data;
        const int batchSize = inp.size[0];
        const int numChannels = inp.size[1];
        const int inpHeight = inp.size[2];
        const int inpWidth = inp.size[3];
        const float rheight = (outHeight > 1) ? static_cast<float>(inpHeight - 1) / (outHeight - 1) : 0.f;
        const float rwidth = (outWidth > 1) ? static_cast<float>(inpWidth - 1) / (outWidth - 1) : 0.f;
        for (int h2 = 0; h2 < outHeight; ++h2)
        {
            const float h1r = rheight * h2;
            const int h1 = static_cast<int>(h1r);
            const int h1p = (h1 < inpHeight - 1) ? 1 : 0;
            const float h1lambda = h1r - h1;
            const float h0lambda = 1.f - h1lambda;
            for (int w2 = 0; w2 < outWidth; ++w2)
            {
                const float w1r = rwidth * w2;
                const int w1 = static_cast<int>(w1r);
                const int w1p = (w1 < inpWidth - 1) ? 1 : 0;
                const float w1lambda = w1r - w1;
                const float w0lambda = 1.f - w1lambda;
                const float* pos1 = inpData + h1 * inpWidth + w1;
                float* pos2 = outData + h2 * outWidth + w2;
                for (int c = 0; c < batchSize * numChannels; ++c)
                {
                    pos2[0] =
                      h0lambda * (w0lambda * pos1[0] + w1lambda * pos1[w1p]) +
                      h1lambda * (w0lambda * pos1[h1p * inpWidth] + w1lambda * pos1[h1p * inpWidth + w1p]);
                    pos1 += inpWidth * inpHeight;
                    pos2 += outWidth * outHeight;
                }
            }
        }
    }
private:
    int outWidth, outHeight;
};

Next we need to register a new layer type and try to import the model.

CV_DNN_REGISTER_LAYER_CLASS(Interp, InterpLayer);

cv::dnn::Net caffeNet = cv::dnn::readNet("/path/to/config.prototxt", "/path/to/weights.caffemodel");

Example: custom layer from TensorFlow

This is an example of how to import a network with tf.image.resize_bilinear operation. This is also a resize but with an implementation different from OpenCV's or Interp above.

Let's create a single layer network:

inp = tf.placeholder(tf.float32, [2, 3, 4, 5], 'input')

resized = tf.image.resize_bilinear(inp, size=[9, 8], name='resize_bilinear')

OpenCV sees that TensorFlow's graph in the following way:

node {
  name: "input"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "resize_bilinear/size"
  op: "Const"
  attr {
    key: "dtype"
    value {
      type: DT_INT32
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_INT32
        tensor_shape {
          dim {
            size: 2
          }
        }
        tensor_content: "\t\000\000\000\010\000\000\000"
      }
    }
  }
}
node {
  name: "resize_bilinear"
  op: "ResizeBilinear"
  input: "input:0"
  input: "resize_bilinear/size"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "align_corners"
    value {
      b: false
    }
  }
}
library {
}

Custom layers import from TensorFlow is designed to put all layer's attr into cv::dnn::LayerParams but input Const blobs into cv::dnn::Layer::blobs. In our case resize's output shape will be stored in layer's blobs[0].

class ResizeBilinearLayer CV_FINAL : public cv::dnn::Layer
{
public:
    ResizeBilinearLayer(const cv::dnn::LayerParams &params) : Layer(params)
    {
        CV_Assert(!params.get<bool>("align_corners", false));
        CV_Assert(!blobs.empty());
        for (size_t i = 0; i < blobs.size(); ++i)
            CV_Assert(blobs[i].type() == CV_32SC1);
        // There are two cases of input blob: a single blob which contains output
        // shape and two blobs with scaling factors.
        if (blobs.size() == 1)
        {
            CV_Assert(blobs[0].total() == 2);
            outHeight = blobs[0].at<int>(0, 0);
            outWidth = blobs[0].at<int>(0, 1);
            factorHeight = factorWidth = 0;
        }
        else
        {
            CV_Assert(blobs.size() == 2); CV_Assert(blobs[0].total() == 1); CV_Assert(blobs[1].total() == 1);
            factorHeight = blobs[0].at<int>(0, 0);
            factorWidth = blobs[1].at<int>(0, 0);
            outHeight = outWidth = 0;
        }
    }
    static cv::Ptr<cv::dnn::Layer> create(cv::dnn::LayerParams& params)
    {
        return cv::Ptr<cv::dnn::Layer>(new ResizeBilinearLayer(params));
    }
    virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
                                 const int,
                                 std::vector<std::vector<int> > &outputs,
                                 std::vector<std::vector<int> > &) const CV_OVERRIDE
    {
        std::vector<int> outShape(4);
        outShape[0] = inputs[0][0];  // batch size
        outShape[1] = inputs[0][1];  // number of channels
        outShape[2] = outHeight != 0 ? outHeight : (inputs[0][2] * factorHeight);
        outShape[3] = outWidth != 0 ? outWidth : (inputs[0][3] * factorWidth);
        outputs.assign(1, outShape);
        return false;
    }
    virtual void finalize(cv::InputArrayOfArrays, cv::OutputArrayOfArrays outputs_arr) CV_OVERRIDE
    {
        std::vector<cv::Mat> outputs;
        outputs_arr.getMatVector(outputs);
        if (!outWidth && !outHeight)
        {
            outHeight = outputs[0].size[2];
            outWidth = outputs[0].size[3];
        }
    }
    // This implementation is based on a reference implementation from
    // https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h
    virtual void forward(cv::InputArrayOfArrays inputs_arr,
                         cv::OutputArrayOfArrays outputs_arr,
                         cv::OutputArrayOfArrays internals_arr) CV_OVERRIDE
    {
        if (inputs_arr.depth() == CV_16S)
        {
            // In case of DNN_TARGET_OPENCL_FP16 target the following method
            // converts data from FP16 to FP32 and calls this forward again.
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }
        std::vector<cv::Mat> inputs, outputs;
        inputs_arr.getMatVector(inputs);
        outputs_arr.getMatVector(outputs);
        cv::Mat& inp = inputs[0];
        cv::Mat& out = outputs[0];
        const float* inpData = (float*)inp.data;
        float* outData = (float*)out.data;
        const int batchSize = inp.size[0];
        const int numChannels = inp.size[1];
        const int inpHeight = inp.size[2];
        const int inpWidth = inp.size[3];
        float heightScale = static_cast<float>(inpHeight) / outHeight;
        float widthScale = static_cast<float>(inpWidth) / outWidth;
        for (int b = 0; b < batchSize; ++b)
        {
            for (int y = 0; y < outHeight; ++y)
            {
                float input_y = y * heightScale;
                int y0 = static_cast<int>(std::floor(input_y));
                int y1 = std::min(y0 + 1, inpHeight - 1);
                for (int x = 0; x < outWidth; ++x)
                {
                    float input_x = x * widthScale;
                    int x0 = static_cast<int>(std::floor(input_x));
                    int x1 = std::min(x0 + 1, inpWidth - 1);
                    for (int c = 0; c < numChannels; ++c)
                    {
                        float interpolation =
                            inpData[offset(inp.size, c, x0, y0, b)] * (1 - (input_y - y0)) * (1 - (input_x - x0)) +
                            inpData[offset(inp.size, c, x0, y1, b)] * (input_y - y0) * (1 - (input_x - x0)) +
                            inpData[offset(inp.size, c, x1, y0, b)] * (1 - (input_y - y0)) * (input_x - x0) +
                            inpData[offset(inp.size, c, x1, y1, b)] * (input_y - y0) * (input_x - x0);
                        outData[offset(out.size, c, x, y, b)] = interpolation;
                    }
                }
            }
        }
    }
private:
    static inline int offset(const cv::MatSize& size, int c, int x, int y, int b)
    {
        return x + size[3] * (y + size[2] * (c + size[1] * b));
    }
    int outWidth, outHeight, factorWidth, factorHeight;
};

Next we register a layer and try to import the model.

CV_DNN_REGISTER_LAYER_CLASS(ResizeBilinear, ResizeBilinearLayer);

cv::dnn::Net tfNet = cv::dnn::readNet("/path/to/graph.pb");

Define a custom layer in Python

The following example shows how to customize OpenCV's layers in Python.

Let's consider Holistically-Nested Edge Detection deep learning model. That was trained with one and only difference comparing to a current version of Caffe framework. Crop layers that receive two input blobs and crop the first one to match spatial dimensions of the second one used to crop from the center. Nowadays Caffe's layer does it from the top-left corner. So using the latest version of Caffe or OpenCV you'll get shifted results with filled borders.

Next we're going to replace OpenCV's Crop layer that makes top-left cropping by a centric one.

Create a class with getMemoryShapes and forward methods

class CropLayer(object):
    def __init__(self, params, blobs):
        self.xstart = 0
        self.xend = 0
        self.ystart = 0
        self.yend = 0
    # Our layer receives two inputs. We need to crop the first input blob
    # to match a shape of the second one (keeping batch size and number of channels)
    def getMemoryShapes(self, inputs):
        inputShape, targetShape = inputs[0], inputs[1]
        batchSize, numChannels = inputShape[0], inputShape[1]
        height, width = targetShape[2], targetShape[3]
        self.ystart = (inputShape[2] - targetShape[2]) // 2
        self.xstart = (inputShape[3] - targetShape[3]) // 2
        self.yend = self.ystart + height
        self.xend = self.xstart + width
        return [[batchSize, numChannels, height, width]]
    def forward(self, inputs):
        return [inputs[0][:,:,self.ystart:self.yend,self.xstart:self.xend]]

Note: Both methods should return lists.

Register a new layer.

cv.dnn_registerLayer('Crop', CropLayer)

That's it! We've replaced an implemented OpenCV's layer to a custom one. You may find a full script in the source code.