Improved isotonic_regression class.

dlib/optimization/isotonic_regression.h

@@ -5,6 +5,7 @@
 
 #include "isotonic_regression_abstract.h"
 #include <vector>
+#include <utility>
 
 namespace dlib
 {
@@ -22,22 +23,13 @@ namespace dlib
             iterator obegin
         )
         {
-            for (auto i = begin; i != end; ++i)
-            {
-                blocks.emplace_back(*i);
-                while (blocks.size() > 1 && prev_block().avg > current_block().avg)
-                {
-                    // merge the last two blocks.
-                    prev_block() = prev_block() + current_block();
-                    blocks.pop_back();
-                }
-            }
+            do_isotonic_regression(begin, end);
 
-            // unpack results to output
+            // unpack blocks to output
             for (auto& block : blocks)
             {
                 for (size_t k = 0; k < block.num; ++k)
-                    *obegin++ = block.avg;
+                    set_val(*obegin++, block.avg);
             }
 
             blocks.clear();
@@ -47,8 +39,108 @@ namespace dlib
             std::vector<double>& vect
         ) { (*this)(vect.begin(), vect.end(), vect.begin()); }
 
+        template <typename T, typename U>
+        void operator() (
+            std::vector<std::pair<T,U>>& vect
+        ) { (*this)(vect.begin(), vect.end(), vect.begin()); }
+
+
+        template <
+            typename const_iterator, 
+            typename iterator
+            >
+        void fit_with_linear_output_interpolation (
+            const_iterator begin,
+            const_iterator end,
+            iterator obegin
+        )
+        {
+            do_isotonic_regression(begin, end);
+
+            // Unpack blocks to output, but here instead of producing the step function
+            // output we linearly interpolate.  Note that this actually fits the data less
+            // than the step-function, but in many applications might be closer to what you
+            // really when when using isotonic_regression than the step function.
+            for (size_t i = 0; i < blocks.size(); ++i)
+            {
+                auto& block = blocks[i];
+
+                double prev = (blocks.front().avg + block.avg)/2;
+                if (i > 0)
+                    prev = (blocks[i-1].avg+block.avg)/2;
+
+                double next = (blocks.back().avg + block.avg)/2;
+                if (i+1 < blocks.size())
+                    next = (blocks[i+1].avg+block.avg)/2;
+
+                for (size_t k = 0; k < block.num; ++k)
+                {
+                    const auto mid = block.num/2.0;
+                    if (k < mid)
+                    {
+                        const double alpha = k/mid;
+                        set_val(*obegin++, (1-alpha)*prev + alpha*block.avg);
+                    }
+                    else
+                    {
+                        const double alpha = k/mid-1;
+                        set_val(*obegin++, alpha*next + (1-alpha)*block.avg);
+                    }
+                }
+            }
+
+            blocks.clear();
+        }
+
+        void fit_with_linear_output_interpolation (
+            std::vector<double>& vect
+        ) { fit_with_linear_output_interpolation(vect.begin(), vect.end(), vect.begin()); }
+
+        template <typename T, typename U>
+        void fit_with_linear_output_interpolation (
+            std::vector<std::pair<T,U>>& vect
+        ) { fit_with_linear_output_interpolation(vect.begin(), vect.end(), vect.begin()); }
+
     private:
 
+        template <
+            typename const_iterator
+            >
+        void do_isotonic_regression (
+            const_iterator begin,
+            const_iterator end
+        )
+        {
+            blocks.clear();
+
+            // Do the actual isotonic regression.  The output is a step-function and is
+            // stored in the vector of blocks.
+            for (auto i = begin; i != end; ++i)
+            {
+                blocks.emplace_back(get_val(*i));
+                while (blocks.size() > 1 && prev_block().avg > current_block().avg)
+                {
+                    // merge the last two blocks.
+                    prev_block() = prev_block() + current_block();
+                    blocks.pop_back();
+                }
+            }
+        }
+
+
+        template <typename T>
+        static double get_val(const T& v) { return v;}
+
+        template <typename T, typename U>
+        static double get_val(const std::pair<T,U>& v) { return v.second;}
+
+        template <typename T>
+        static void set_val(T& v, double val) { v = val;}
+
+        template <typename T, typename U>
+        static void set_val(std::pair<T,U>& v, double val) { v.second = val;}
+
+
 
         struct block_t
         {

dlib/optimization/isotonic_regression_abstract.h

@@ -4,6 +4,7 @@
 #ifdef DLIB_ISOTONIC_ReGRESSION_ABSTRACT_H_
 
 #include <vector>
+#include <utility>
 
 namespace dlib
 {
@@ -31,9 +32,10 @@ namespace dlib
         );
         /*!
             requires
-                - [begin,end) is an iterator range of float or doubles.
-                - obegin points to an iterator range at least std::distance(begin,end) in
-                  size which is capable of storing double or float values.
+                - [begin,end) is an iterator range of float or doubles or a range of
+                  std::pair<T,double> or std::pair<T,float> where T an be anything.
+                - obegin points to an iterator range at least std::distance(begin,end). 
+                - obegin points to an iterator range of objects of type float, double, std::pair<T,float>, or std::pair<T,double>.
             ensures
                 - Given the range of real values stored in [begin,end), this method performs isotonic regression
                   on this data and writes the results to obegin.  To be specific:
@@ -44,6 +46,10 @@ namespace dlib
                       OUT[i] <= OUT[i+1], i.e. that OUT is monotonic.
                 - It is OK for [begin,end) to overlap with the range pointed to by obegin.
                   That is, this function can run in-place.
+                - Note that when the inputs or outputs are std::pairs this algorithm only
+                  looks at the .second field of the pair.  It therefore still treats these
+                  iterator ranges as ranges of reals since it only looks at the .second
+                  field, which is a real number.  The .first field is entirely ignored.
         !*/
 
         void operator() (
@@ -56,6 +62,63 @@ namespace dlib
                 - #vect.size() == vect.size()
         !*/
 
+        template <typename T, typename U>
+        void operator() (
+            std::vector<std::pair<T,U>>& vect
+        ) { (*this)(vect.begin(), vect.end(), vect.begin()); }
+        /*!
+            ensures
+                - performs in-place isotonic regression.  Therefore, #vect will contain the
+                  isotonic regression of vect.
+                - #vect.size() == vect.size()
+        !*/
+
+
+        template <
+            typename const_iterator, 
+            typename iterator
+            >
+        void fit_with_linear_output_interpolation (
+            const_iterator begin,
+            const_iterator end,
+            iterator obegin
+        );
+        /*!
+            requires
+                - [begin,end) is an iterator range of float or doubles or a range of
+                  std::pair<T,double> or std::pair<T,float> where T an be anything.
+                - obegin points to an iterator range at least std::distance(begin,end). 
+                - obegin points to an iterator range of objects of type float, double, std::pair<T,float>, or std::pair<T,double>.
+            ensures
+                - This function behaves just like (*this)(begin,end,obegin) except that the
+                  output is interpolated.  To explain, not that the optimal output of
+                  isotonic regression is a step function.  However, in many applications
+                  that isn't really what you want.  You want something smoother.  So
+                  fit_with_linear_output_interpolation() does isotonic regression and then
+                  linearly interpolates the step function into a piecewise linear function.
+        !*/
+
+        void fit_with_linear_output_interpolation (
+            std::vector<double>& vect
+        ) { fit_with_linear_output_interpolation(vect.begin(), vect.end(), vect.begin()); }
+        /*!
+            ensures
+                - performs in-place isotonic regression.  Therefore, #vect will contain the
+                  isotonic regression of vect.
+                - #vect.size() == vect.size()
+        !*/
+
+        template <typename T, typename U>
+        void fit_with_linear_output_interpolation (
+            std::vector<std::pair<T,U>>& vect
+        ) { fit_with_linear_output_interpolation(vect.begin(), vect.end(), vect.begin()); }
+        /*!
+            ensures
+                - performs in-place isotonic regression.  Therefore, #vect will contain the
+                  isotonic regression of vect.
+                - #vect.size() == vect.size()
+        !*/
+
     };
 }