Line data Source code
1 : // ***************************************************************
2 : // SPDX-FileCopyrightText: Copyright 2024 Ricardo Montañana Gómez
3 : // SPDX-FileType: SOURCE
4 : // SPDX-License-Identifier: MIT
5 : // ***************************************************************
6 :
7 : #include "Mst.h"
8 : #include "BayesMetrics.h"
9 : namespace bayesnet {
10 : //samples is n+1xm tensor used to fit the model
11 3957 : Metrics::Metrics(const torch::Tensor& samples, const std::vector<std::string>& features, const std::string& className, const int classNumStates)
12 3957 : : samples(samples)
13 3957 : , features(features)
14 3957 : , className(className)
15 3957 : , classNumStates(classNumStates)
16 : {
17 3957 : }
18 : //samples is n+1xm std::vector used to fit the model
19 176 : Metrics::Metrics(const std::vector<std::vector<int>>& vsamples, const std::vector<int>& labels, const std::vector<std::string>& features, const std::string& className, const int classNumStates)
20 176 : : features(features)
21 176 : , className(className)
22 176 : , classNumStates(classNumStates)
23 352 : , samples(torch::zeros({ static_cast<int>(vsamples.size() + 1), static_cast<int>(vsamples[0].size()) }, torch::kInt32))
24 : {
25 1408 : for (int i = 0; i < vsamples.size(); ++i) {
26 4928 : samples.index_put_({ i, "..." }, torch::tensor(vsamples[i], torch::kInt32));
27 : }
28 704 : samples.index_put_({ -1, "..." }, torch::tensor(labels, torch::kInt32));
29 1584 : }
30 1099 : std::vector<int> Metrics::SelectKBestWeighted(const torch::Tensor& weights, bool ascending, unsigned k)
31 : {
32 : // Return the K Best features
33 1099 : auto n = features.size();
34 1099 : if (k == 0) {
35 0 : k = n;
36 : }
37 : // compute scores
38 1099 : scoresKBest.clear();
39 1099 : featuresKBest.clear();
40 3297 : auto label = samples.index({ -1, "..." });
41 37425 : for (int i = 0; i < n; ++i) {
42 108978 : scoresKBest.push_back(mutualInformation(label, samples.index({ i, "..." }), weights));
43 36326 : featuresKBest.push_back(i);
44 : }
45 : // sort & reduce scores and features
46 1099 : if (ascending) {
47 245 : sort(featuresKBest.begin(), featuresKBest.end(), [&](int i, int j)
48 5931 : { return scoresKBest[i] < scoresKBest[j]; });
49 245 : sort(scoresKBest.begin(), scoresKBest.end(), std::less<double>());
50 245 : if (k < n) {
51 308 : for (int i = 0; i < n - k; ++i) {
52 220 : featuresKBest.erase(featuresKBest.begin());
53 220 : scoresKBest.erase(scoresKBest.begin());
54 : }
55 : }
56 : } else {
57 854 : sort(featuresKBest.begin(), featuresKBest.end(), [&](int i, int j)
58 168709 : { return scoresKBest[i] > scoresKBest[j]; });
59 854 : sort(scoresKBest.begin(), scoresKBest.end(), std::greater<double>());
60 854 : featuresKBest.resize(k);
61 854 : scoresKBest.resize(k);
62 : }
63 2198 : return featuresKBest;
64 38524 : }
65 88 : std::vector<double> Metrics::getScoresKBest() const
66 : {
67 88 : return scoresKBest;
68 : }
69 :
70 374 : torch::Tensor Metrics::conditionalEdge(const torch::Tensor& weights)
71 : {
72 374 : auto result = std::vector<double>();
73 374 : auto source = std::vector<std::string>(features);
74 374 : source.push_back(className);
75 374 : auto combinations = doCombinations(source);
76 : // Compute class prior
77 374 : auto margin = torch::zeros({ classNumStates }, torch::kFloat);
78 2024 : for (int value = 0; value < classNumStates; ++value) {
79 6600 : auto mask = samples.index({ -1, "..." }) == value;
80 1650 : margin[value] = mask.sum().item<double>() / samples.size(1);
81 1650 : }
82 10098 : for (auto [first, second] : combinations) {
83 9724 : int index_first = find(features.begin(), features.end(), first) - features.begin();
84 9724 : int index_second = find(features.begin(), features.end(), second) - features.begin();
85 9724 : double accumulated = 0;
86 57640 : for (int value = 0; value < classNumStates; ++value) {
87 191664 : auto mask = samples.index({ -1, "..." }) == value;
88 143748 : auto first_dataset = samples.index({ index_first, mask });
89 143748 : auto second_dataset = samples.index({ index_second, mask });
90 95832 : auto weights_dataset = weights.index({ mask });
91 95832 : auto mi = mutualInformation(first_dataset, second_dataset, weights_dataset);
92 47916 : auto pb = margin[value].item<double>();
93 47916 : accumulated += pb * mi;
94 47916 : }
95 9724 : result.push_back(accumulated);
96 9724 : }
97 374 : long n_vars = source.size();
98 374 : auto matrix = torch::zeros({ n_vars, n_vars });
99 374 : auto indices = torch::triu_indices(n_vars, n_vars, 1);
100 10098 : for (auto i = 0; i < result.size(); ++i) {
101 9724 : auto x = indices[0][i];
102 9724 : auto y = indices[1][i];
103 9724 : matrix[x][y] = result[i];
104 9724 : matrix[y][x] = result[i];
105 9724 : }
106 748 : return matrix;
107 241604 : }
108 : // To use in Python
109 0 : std::vector<float> Metrics::conditionalEdgeWeights(std::vector<float>& weights_)
110 : {
111 0 : const torch::Tensor weights = torch::tensor(weights_);
112 0 : auto matrix = conditionalEdge(weights);
113 0 : std::vector<float> v(matrix.data_ptr<float>(), matrix.data_ptr<float>() + matrix.numel());
114 0 : return v;
115 0 : }
116 101565 : double Metrics::entropy(const torch::Tensor& feature, const torch::Tensor& weights)
117 : {
118 101565 : torch::Tensor counts = feature.bincount(weights);
119 101565 : double totalWeight = counts.sum().item<double>();
120 101565 : torch::Tensor probs = counts.to(torch::kFloat) / totalWeight;
121 101565 : torch::Tensor logProbs = torch::log(probs);
122 101565 : torch::Tensor entropy = -probs * logProbs;
123 203130 : return entropy.nansum().item<double>();
124 101565 : }
125 : // H(Y|X) = sum_{x in X} p(x) H(Y|X=x)
126 91263 : double Metrics::conditionalEntropy(const torch::Tensor& firstFeature, const torch::Tensor& secondFeature, const torch::Tensor& weights)
127 : {
128 91263 : int numSamples = firstFeature.sizes()[0];
129 91263 : torch::Tensor featureCounts = secondFeature.bincount(weights);
130 91263 : std::unordered_map<int, std::unordered_map<int, double>> jointCounts;
131 91263 : double totalWeight = 0;
132 11715815 : for (auto i = 0; i < numSamples; i++) {
133 11624552 : jointCounts[secondFeature[i].item<int>()][firstFeature[i].item<int>()] += weights[i].item<double>();
134 11624552 : totalWeight += weights[i].item<float>();
135 : }
136 91263 : if (totalWeight == 0)
137 0 : return 0;
138 91263 : double entropyValue = 0;
139 311456 : for (int value = 0; value < featureCounts.sizes()[0]; ++value) {
140 220193 : double p_f = featureCounts[value].item<double>() / totalWeight;
141 220193 : double entropy_f = 0;
142 655015 : for (auto& [label, jointCount] : jointCounts[value]) {
143 434822 : double p_l_f = jointCount / featureCounts[value].item<double>();
144 434822 : if (p_l_f > 0) {
145 434822 : entropy_f -= p_l_f * log(p_l_f);
146 : } else {
147 0 : entropy_f = 0;
148 : }
149 : }
150 220193 : entropyValue += p_f * entropy_f;
151 : }
152 91263 : return entropyValue;
153 91263 : }
154 : // I(X;Y) = H(Y) - H(Y|X)
155 91263 : double Metrics::mutualInformation(const torch::Tensor& firstFeature, const torch::Tensor& secondFeature, const torch::Tensor& weights)
156 : {
157 91263 : return entropy(firstFeature, weights) - conditionalEntropy(firstFeature, secondFeature, weights);
158 : }
159 : /*
160 : Compute the maximum spanning tree considering the weights as distances
161 : and the indices of the weights as nodes of this square matrix using
162 : Kruskal algorithm
163 : */
164 319 : std::vector<std::pair<int, int>> Metrics::maximumSpanningTree(const std::vector<std::string>& features, const torch::Tensor& weights, const int root)
165 : {
166 319 : auto mst = MST(features, weights, root);
167 638 : return mst.maximumSpanningTree();
168 319 : }
169 : }
|
