Merge pull request #1179 from arrybn:statistics

modules/dnn/include/opencv2/dnn/dnn.hpp

@@ -135,6 +135,8 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
                                      const int requiredOutputs,
                                      std::vector<MatShape> &outputs,
                                      std::vector<MatShape> &internals) const;
+        virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                               const std::vector<MatShape> &outputs) const {(void)inputs; (void)outputs; return 0;}
 
         CV_PROP String name; //!< Name of the layer instance, can be used for logging or other internal purposes.
         CV_PROP String type; //!< Type name which was used for creating layer by layer factory.
@@ -323,6 +325,50 @@ namespace dnn //! This namespace is used for dnn module functionlaity.
                                      const int layerId,
                                      std::vector<MatShape>* inLayerShapes,
                                      std::vector<MatShape>* outLayerShapes) const;
+         /** @brief Computes FLOP for whole loaded model with specified input shapes.
+          * @param netInputShapes vector of shapes for all net inputs.
+          * @returns computed FLOP.
+          */
+         CV_WRAP int64 getFLOPS(const std::vector<MatShape>& netInputShapes) const;
+         /** @overload */
+         CV_WRAP int64 getFLOPS(const MatShape& netInputShape) const;
+         /** @overload */
+         CV_WRAP int64 getFLOPS(const int layerId,
+                               const std::vector<MatShape>& netInputShapes) const;
+         /** @overload */
+         CV_WRAP int64 getFLOPS(const int layerId,
+                               const MatShape& netInputShape) const;
+
+         /** @brief Returns list of types for layer used in model.
+          * @param layersTypes output parameter for returning types.
+          */
+         CV_WRAP void getLayerTypes(std::vector<String>& layersTypes) const;
+
+         /** @brief Returns count of layers of specified type.
+          * @param layerType type.
+          * @returns count of layers
+          */
+         CV_WRAP int getLayersCount(const String& layerType) const;
+
+         /** @brief Computes bytes number which are requered to store
+          * all weights and intermediate blobs for model.
+          * @param netInputShapes vector of shapes for all net inputs.
+          * @param weights output parameter to store resulting bytes for weights.
+          * @param blobs output parameter to store resulting bytes for intermediate blobs.
+          */
+         CV_WRAP void getMemoryConsumption(const std::vector<MatShape>& netInputShapes,
+                                           size_t& weights, size_t& blobs) const;
+         /** @overload */
+         CV_WRAP void getMemoryConsumption(const MatShape& netInputShape,
+                                           size_t& weights, size_t& blobs) const;
+         /** @overload */
+         CV_WRAP void getMemoryConsumption(const int layerId,
+                                           const std::vector<MatShape>& netInputShapes,
+                                           size_t& weights, size_t& blobs) const;
+         /** @overload */
+         CV_WRAP void getMemoryConsumption(const int layerId,
+                                           const MatShape& netInputShape,
+                                           size_t& weights, size_t& blobs) const;
     private:
 
         struct Impl;

modules/dnn/src/dnn.cpp

@@ -876,6 +876,144 @@ void Net::getLayerShapes(const Net::Impl::ShapesVec& netInputShapes,
         *outLayerShapes = shapes.out;
 }
 
+int64 Net::getFLOPS(const std::vector<MatShape>& netInputShapes) const
+{
+    int64 flops = 0;
+    std::vector<int> ids;
+    std::vector<std::vector<MatShape> > inShapes, outShapes;
+    getLayersShapes(netInputShapes, &ids, &inShapes, &outShapes);
+    CV_Assert(inShapes.size() == outShapes.size());
+    CV_Assert(inShapes.size() == ids.size());
+
+    for(int i = 0; i < ids.size(); i++)
+    {
+        flops += impl->layers[ids[i]].getLayerInstance()->getFLOPS(inShapes[i],
+                                                                   outShapes[i]);
+    }
+
+    return flops;
+}
+
+int64 Net::getFLOPS(const MatShape& netInputShape) const
+{
+    return getFLOPS(std::vector<MatShape>(1, netInputShape));
+}
+
+int64 Net::getFLOPS(const int layerId,
+              const std::vector<MatShape>& netInputShapes) const
+{
+    Impl::MapIdToLayerData::iterator layer = impl->layers.find(layerId);
+    CV_Assert(layer != impl->layers.end());
+
+    Impl::LayerShapes shapes;
+    impl->getLayerShapes(netInputShapes, layerId, shapes);
+
+    return layer->second.getLayerInstance()->getFLOPS(shapes.in, shapes.out);
+}
+
+int64 Net::getFLOPS(const int layerId,
+              const MatShape& netInputShape) const
+{
+    return getFLOPS(layerId, std::vector<MatShape>(1, netInputShape));
+}
+
+void Net::getLayerTypes(std::vector<String>& layersTypes) const
+{
+    layersTypes.clear();
+
+    std::map<String, int> layers;
+    for (Impl::MapIdToLayerData::iterator it = impl->layers.begin();
+         it != impl->layers.end(); it++)
+    {
+        if (layers.find(it->second.type) == layers.end())
+            layers[it->second.type] = 0;
+        layers[it->second.type]++;
+    }
+
+    for (std::map<String, int>::iterator it = layers.begin();
+         it != layers.end(); it++)
+    {
+        layersTypes.push_back(it->first);
+    }
+}
+
+int Net::getLayersCount(const String& layerType) const
+{
+    int count = 0;
+    for (Impl::MapIdToLayerData::iterator it = impl->layers.begin();
+         it != impl->layers.end(); it++)
+    {
+        if (it->second.type == layerType)
+            count++;
+    }
+    return count;
+}
+
+void Net::getMemoryConsumption(const int layerId,
+                               const std::vector<MatShape>& netInputShapes,
+                               size_t& weights, size_t& blobs) const
+{
+    Impl::MapIdToLayerData::iterator layer = impl->layers.find(layerId);
+    CV_Assert(layer != impl->layers.end());
+
+    weights = blobs = 0;
+
+    for(int i = 0; i < layer->second.params.blobs.size(); i++)
+    {
+        const Mat& weightsBlob = layer->second.params.blobs[i];
+        weights += weightsBlob.total()*weightsBlob.elemSize();
+    }
+
+    std::vector<MatShape> outLayerShapes;
+    getLayerShapes(netInputShapes, layerId, 0, &outLayerShapes);
+    for(int i = 0; i < outLayerShapes.size(); i++)
+    {
+        blobs += total(outLayerShapes[i]) * sizeof(float);
+    }
+}
+
+void Net::getMemoryConsumption(const std::vector<MatShape>& netInputShapes,
+                               size_t& weights, size_t& blobs) const
+{
+    std::vector<int> layerIds;
+    std::vector<std::vector<MatShape> > outLayerShapes;
+
+    getLayersShapes(netInputShapes, &layerIds, 0, &outLayerShapes);
+
+    weights = blobs = 0;
+    for(int i = 0; i < layerIds.size(); i++)
+    {
+        Impl::MapIdToLayerData::iterator layer = impl->layers.find(layerIds[i]);
+        CV_Assert(layer != impl->layers.end());
+
+        for(int j = 0; j < layer->second.params.blobs.size(); j++)
+        {
+            const Mat& weightsBlob = layer->second.params.blobs[j];
+            weights += weightsBlob.total()*weightsBlob.elemSize();
+        }
+
+        for(int j = 0; j < outLayerShapes[i].size(); j++)
+        {
+            blobs += total(outLayerShapes[i][j]) * sizeof(float);
+        }
+    }
+}
+
+void Net::getMemoryConsumption(const int layerId,
+                               const MatShape& netInputShape,
+                               size_t& weights, size_t& blobs) const
+{
+    getMemoryConsumption(layerId, std::vector<MatShape>(1, netInputShape),
+                         weights, blobs);
+}
+
+void Net::getMemoryConsumption(const MatShape& netInputShape,
+                               size_t& weights, size_t& blobs) const
+{
+    getMemoryConsumption(std::vector<MatShape>(1, netInputShape),
+                         weights, blobs);
+}
+
 //////////////////////////////////////////////////////////////////////////
 
 Importer::~Importer() {}

modules/dnn/src/layers/batch_norm_layer.cpp

@@ -10,6 +10,7 @@ Implementation of Batch Normalization layer.
 */
 
 #include "../precomp.hpp"
+#include <opencv2/dnn/shape_utils.hpp>
 
 namespace cv
 {
@@ -78,6 +79,19 @@ public:
         }
     }
 
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+
+        int64 flops = 0;
+        for(int i = 0; i < inputs.size(); i++)
+        {
+            flops += 3*total(inputs[i]);
+        }
+        return flops;
+    }
+
     bool hasWeights, hasBias;
     float epsilon;
 };

modules/dnn/src/layers/convolution_layer.cpp

@@ -224,6 +224,20 @@ public:
                             dilation.height, dilation.width, outH, outW, dstRow.ptr<float>());
         }
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        CV_Assert(inputs.size() == outputs.size());
+
+        int64 flops = 0;
+        for (int i = 0; i < inputs.size(); i++)
+        {
+            flops += total(outputs[i])*(2*kernel.area()*inputs[i][1] + 1);
+        }
+
+        return flops;
+    }
 };
 
 class DeConvolutionLayerImpl : public BaseConvolutionLayerImpl
@@ -339,6 +353,22 @@ public:
                         dilation.height, dilation.width, dstImg.ptr<float>(), &ofsbuf[0]);
     }
 
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        CV_Assert(inputs.size() == outputs.size());
+
+        float flops = 0;
+        int outChannels = blobs[0].size[0];
+
+        for (int i = 0; i < inputs.size(); i++)
+        {
+            flops += 2*outChannels*kernel.area()*total(inputs[i]);
+        }
+
+        return flops;
+    }
+
     std::vector<int> ofsbuf;
 };
 

modules/dnn/src/layers/elementwise_layers.cpp

@@ -63,6 +63,17 @@ public:
         }
     }
 
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        long flops = 0;
+        for (int i = 0; i < outputs.size(); i++)
+        {
+            flops += total(outputs[i]) * func.getFLOPSPerElement();
+        }
+        return flops;
+    }
+
     Func func;
     bool run_parallel;
 };
@@ -79,6 +90,8 @@ struct ReLUFunctor
     {
         return (x >= (TFloat)0) ? x : (TFloat)slope * x;
     }
+
+    int64 getFLOPSPerElement() const {return 1;}
 };
 
 struct TanHFunctor
@@ -90,6 +103,8 @@ struct TanHFunctor
     {
         return tanh(x);
     }
+
+    int64 getFLOPSPerElement() const {return 1;}
 };
 
 struct SigmoidFunctor
@@ -101,6 +116,8 @@ struct SigmoidFunctor
     {
         return (TFloat)1 / ((TFloat)1 + exp(-x));
     }
+
+    int64 getFLOPSPerElement() const {return 3;}
 };
 
 struct AbsValFunctor
@@ -112,6 +129,8 @@ struct AbsValFunctor
     {
         return abs(x);
     }
+
+    int64 getFLOPSPerElement() const {return 1;}
 };
 
 struct BNLLFunctor
@@ -123,6 +142,8 @@ struct BNLLFunctor
     {
         return log((TFloat)1 + exp(-abs(x)));
     }
+
+    int64 getFLOPSPerElement() const {return 5;}
 };
 
 struct PowerFunctor
@@ -141,6 +162,8 @@ struct PowerFunctor
     {
         return pow((TFloat)shift + (TFloat)scale * x, (TFloat)power);
     }
+
+    int64 getFLOPSPerElement() const {return 3;}
 };
 
 struct PowerFunctor1
@@ -158,6 +181,8 @@ struct PowerFunctor1
     {
         return (TFloat)shift + (TFloat)scale * x;
     }
+
+    int64 getFLOPSPerElement() const {return 2;}
 };
 
 class ChannelsPReLULayerImpl : public ChannelsPReLULayer
@@ -210,6 +235,20 @@ public:
             }
         }
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)inputs; // suppress unused variable warning
+        long flops = 0;
+
+        for (int i = 0; i < outputs.size(); i++)
+        {
+            flops += total(outputs[i]) * 3;
+        }
+
+        return flops;
+    }
 };
 
 #define ACTIVATION_CREATOR_FOR(_Layer, _Functor, ...) \

modules/dnn/src/layers/eltwise_layer.cpp

@@ -143,6 +143,17 @@ public:
                 break;
         }
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        CV_Assert(inputs.size());
+
+        long flops = inputs.size() * total(inputs[0]);
+
+        return flops;
+    }
 };
 
 Ptr<EltwiseLayer> EltwiseLayer::create(const LayerParams& params)

modules/dnn/src/layers/fully_connected_layer.cpp

@@ -117,6 +117,22 @@ public:
         }
     }
 
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)inputs; // suppress unused variable warning
+        long flops = 0;
+
+        int innerSize = blobs[0].size[1];
+        for(int i = 0; i < outputs.size(); i++)
+        {
+            flops += 3*innerSize*total(outputs[i]);
+        }
+
+        return flops;
+
+    }
+
     bool bias;
 };
 

modules/dnn/src/layers/lrn_layer.cpp

@@ -171,6 +171,35 @@ public:
             }
         }
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        CV_Assert(inputs.size() > 0);
+        long flops = 0;
+
+        for(int i = 0; i < inputs.size(); i++)
+        {
+            if (type == CHANNEL_NRM)
+            {
+                int channels = inputs[i][1];
+                int ksize = (size - 1) / 2;
+
+                flops += inputs[i][0]*(std::min(ksize, channels)*2*total(inputs[i], 2) + channels*4*total(inputs[i], 2));
+
+                if (ksize < channels)
+                {
+                    flops += (size + 2*(channels - size))*total(inputs[i], 2);
+                }
+            }
+            else
+            {
+                flops += total(inputs[i])*(2*size*size + 2);
+            }
+        }
+        return flops;
+    }
 };
 
 Ptr<LRNLayer> LRNLayer::create(const LayerParams& params)

modules/dnn/src/layers/mvn_layer.cpp

@@ -85,6 +85,18 @@ public:
             }
         }
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        long flops = 0;
+        for(int i = 0; i < inputs.size(); i++)
+        {
+            flops += 6*total(inputs[i]) + 3*total(inputs[i], 0, normVariance ? 2 : 1);
+        }
+        return flops;
+    }
 };
 
 Ptr<MVNLayer> MVNLayer::create(const LayerParams& params)

modules/dnn/src/layers/pooling_layer.cpp

@@ -241,6 +241,27 @@ public:
 
         return false;
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)inputs; // suppress unused variable warning
+        long flops = 0;
+
+        for(int i = 0; i < outputs.size(); i++)
+        {
+            if (type == MAX)
+            {
+                if (i%2 == 0)
+                    flops += total(outputs[i])*kernel.area();
+            }
+            else
+            {
+                flops += total(outputs[i])*(kernel.area() + 1);
+            }
+        }
+        return flops;
+    }
 };
 
 Ptr<PoolingLayer> PoolingLayer::create(const LayerParams& params)

modules/dnn/src/layers/prior_box_layer.cpp

@@ -312,6 +312,20 @@ public:
         }
     }
 
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        long flops = 0;
+
+        for (int i = 0; i < inputs.size(); i++)
+        {
+            flops += total(inputs[i], 2) * _numPriors * 4;
+        }
+
+        return flops;
+    }
+
     float _minSize;
     float _maxSize;
 

modules/dnn/src/layers/scale_layer.cpp

@@ -56,6 +56,18 @@ public:
         }
     }
 
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        long flops = 0;
+        for(int i = 0; i < inputs.size(); i++)
+        {
+            flops += 2*total(inputs[i]);
+        }
+        return flops;
+    }
+
     bool hasBias;
 };
 

modules/dnn/src/layers/shift_layer.cpp

@@ -81,6 +81,20 @@ public:
             }
         }
     }
+
+    virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
+                           const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        long flops = 0;
+
+        for(int i= 0; i < inputs.size(); i++)
+        {
+           flops += total(inputs[i]);
+        }
+
+        return flops;
+    }
 };
 
 Ptr<ShiftLayer> ShiftLayer::create(const LayerParams& params)

modules/dnn/src/layers/softmax_layer.cpp

@@ -146,6 +146,20 @@ public:
         }
     }
 
+    int64 getFLOPS(const std::vector<MatShape> &inputs,
+                  const std::vector<MatShape> &outputs) const
+    {
+        (void)outputs; // suppress unused variable warning
+        int64 flops = 0;
+
+        for (int i = 0; i < inputs.size(); i++)
+        {
+            flops += 4*total(inputs[i]);
+        }
+
+        return flops;
+    }
+
     int axisRaw;
 };
 

modules/dnn/test/test_caffe_importer.cpp

@@ -41,6 +41,7 @@
 
 #include "test_precomp.hpp"
 #include "npy_blob.hpp"
+#include <opencv2/dnn/shape_utils.hpp>
 
 namespace cvtest
 {

modules/dnn/test/test_torch_importer.cpp

@@ -43,6 +43,7 @@
 
 #include "test_precomp.hpp"
 #include "npy_blob.hpp"
+#include <opencv2/dnn/shape_utils.hpp>
 
 namespace cvtest
 {