| 1 | // $Id: CrossSplitter.cc 558 2006-03-10 15:58:16Z peter $ |
|---|
| 2 | |
|---|
| 3 | |
|---|
| 4 | #include <c++_tools/classifier/CrossSplitter.h> |
|---|
| 5 | #include <c++_tools/classifier/DataLookup2D.h> |
|---|
| 6 | #include <c++_tools/classifier/Target.h> |
|---|
| 7 | #include <c++_tools/random/random.h> |
|---|
| 8 | |
|---|
| 9 | #include <algorithm> |
|---|
| 10 | #include <cassert> |
|---|
| 11 | #include <utility> |
|---|
| 12 | #include <vector> |
|---|
| 13 | |
|---|
| 14 | namespace theplu { |
|---|
| 15 | namespace classifier { |
|---|
| 16 | |
|---|
| 17 | CrossSplitter::CrossSplitter(const Target& target, const DataLookup2D& data, |
|---|
| 18 | const size_t N, const size_t k) |
|---|
| 19 | : k_(k), state_(0), target_(target), weighted_(false) |
|---|
| 20 | { |
|---|
| 21 | assert(target.size()>1); |
|---|
| 22 | assert(target.size()==data.columns()); |
|---|
| 23 | |
|---|
| 24 | build(target, N, k); |
|---|
| 25 | |
|---|
| 26 | for (size_t i=0; i<N; i++){ |
|---|
| 27 | |
|---|
| 28 | // Dynamically allocated. Must be deleted in destructor. |
|---|
| 29 | training_data_.push_back(data.training_data(training_index_[i])); |
|---|
| 30 | training_weight_.push_back |
|---|
| 31 | (new MatrixLookup(training_data_.back()->rows(), |
|---|
| 32 | training_data_.back()->columns(),1)); |
|---|
| 33 | validation_data_.push_back(data.validation_data(training_index_[i], |
|---|
| 34 | validation_index_[i])); |
|---|
| 35 | validation_weight_.push_back |
|---|
| 36 | (new MatrixLookup(validation_data_.back()->rows(), |
|---|
| 37 | validation_data_.back()->columns(),1)); |
|---|
| 38 | |
|---|
| 39 | |
|---|
| 40 | training_target_.push_back(Target(target,training_index_[i])); |
|---|
| 41 | validation_target_.push_back(Target(target,validation_index_[i])); |
|---|
| 42 | } |
|---|
| 43 | assert(training_data_.size()==N); |
|---|
| 44 | assert(training_weight_.size()==N); |
|---|
| 45 | assert(training_target_.size()==N); |
|---|
| 46 | assert(validation_data_.size()==N); |
|---|
| 47 | assert(validation_weight_.size()==N); |
|---|
| 48 | assert(validation_target_.size()==N); |
|---|
| 49 | } |
|---|
| 50 | |
|---|
| 51 | CrossSplitter::CrossSplitter(const Target& target, const DataLookup2D& data, |
|---|
| 52 | const MatrixLookup& weight, |
|---|
| 53 | const size_t N, const size_t k) |
|---|
| 54 | : k_(k), state_(0), target_(target), weighted_(true) |
|---|
| 55 | { |
|---|
| 56 | assert(target.size()>1); |
|---|
| 57 | assert(target.size()==data.columns()); |
|---|
| 58 | |
|---|
| 59 | build(target, N, k); |
|---|
| 60 | |
|---|
| 61 | for (size_t i=0; i<N; i++){ |
|---|
| 62 | |
|---|
| 63 | // Dynamically allocated. Must be deleted in destructor. |
|---|
| 64 | training_data_.push_back(data.training_data(training_index_[i])); |
|---|
| 65 | validation_data_.push_back(data.validation_data(training_index_[i], |
|---|
| 66 | validation_index_[i])); |
|---|
| 67 | training_weight_.push_back(weight.training_data(training_index_[i])); |
|---|
| 68 | validation_weight_.push_back(weight.validation_data(training_index_[i], |
|---|
| 69 | validation_index_[i])); |
|---|
| 70 | |
|---|
| 71 | |
|---|
| 72 | training_target_.push_back(Target(target,training_index_[i])); |
|---|
| 73 | validation_target_.push_back(Target(target,validation_index_[i])); |
|---|
| 74 | } |
|---|
| 75 | assert(training_data_.size()==N); |
|---|
| 76 | assert(training_weight_.size()==N); |
|---|
| 77 | assert(training_target_.size()==N); |
|---|
| 78 | assert(validation_data_.size()==N); |
|---|
| 79 | assert(validation_weight_.size()==N); |
|---|
| 80 | assert(validation_target_.size()==N); |
|---|
| 81 | } |
|---|
| 82 | |
|---|
| 83 | CrossSplitter::~CrossSplitter() |
|---|
| 84 | { |
|---|
| 85 | assert(training_data_.size()==validation_data_.size()); |
|---|
| 86 | for (size_t i=0; i<training_data_.size(); i++) |
|---|
| 87 | delete training_data_[i]; |
|---|
| 88 | for (size_t i=0; i<validation_data_.size(); i++) |
|---|
| 89 | delete validation_data_[i]; |
|---|
| 90 | for (size_t i=0; i<training_weight_.size(); i++) |
|---|
| 91 | delete training_weight_[i]; |
|---|
| 92 | for (size_t i=0; i<validation_weight_.size(); i++) |
|---|
| 93 | delete validation_weight_[i]; |
|---|
| 94 | } |
|---|
| 95 | |
|---|
| 96 | void CrossSplitter::build(const Target& target, size_t N, size_t k) |
|---|
| 97 | { |
|---|
| 98 | std::vector<std::pair<size_t,size_t> > v; |
|---|
| 99 | for (size_t i=0; i<target.size(); i++) |
|---|
| 100 | v.push_back(std::make_pair(target(i),i)); |
|---|
| 101 | // sorting with respect to class |
|---|
| 102 | std::sort(v.begin(),v.end()); |
|---|
| 103 | |
|---|
| 104 | // my_begin[i] is index of first sample of class i |
|---|
| 105 | std::vector<size_t> my_begin; |
|---|
| 106 | my_begin.reserve(target.nof_classes()); |
|---|
| 107 | my_begin.push_back(0); |
|---|
| 108 | for (size_t i=1; i<target.size(); i++) |
|---|
| 109 | while (v[i].first > my_begin.size()-1) |
|---|
| 110 | my_begin.push_back(i); |
|---|
| 111 | my_begin.push_back(target.size()); |
|---|
| 112 | |
|---|
| 113 | random::DiscreteUniform rnd; |
|---|
| 114 | |
|---|
| 115 | for (size_t i=0; i<N; ) { |
|---|
| 116 | // shuffle indices within class each class |
|---|
| 117 | for (size_t j=0; j<target.nof_classes(); j++) |
|---|
| 118 | random_shuffle(v.begin()+my_begin[j],v.begin()+my_begin[j+1],rnd); |
|---|
| 119 | |
|---|
| 120 | for (size_t part=0; part<k && i<N; i++, part++) { |
|---|
| 121 | std::vector<size_t> training_index; |
|---|
| 122 | std::vector<size_t> validation_index; |
|---|
| 123 | for (size_t j=0; j<v.size(); j++) { |
|---|
| 124 | if (j%k==part) |
|---|
| 125 | validation_index.push_back(v[j].second); |
|---|
| 126 | else |
|---|
| 127 | training_index.push_back(v[j].second); |
|---|
| 128 | } |
|---|
| 129 | |
|---|
| 130 | training_index_.push_back(training_index); |
|---|
| 131 | validation_index_.push_back(validation_index); |
|---|
| 132 | } |
|---|
| 133 | } |
|---|
| 134 | assert(training_index_.size()==N); |
|---|
| 135 | assert(validation_index_.size()==N); |
|---|
| 136 | } |
|---|
| 137 | |
|---|
| 138 | }} // of namespace classifier and namespace theplu |
|---|
