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 2123 : Metrics::Metrics(const torch::Tensor& samples, const std::vector<std::string>& features, const std::string& className, const int classNumStates)
12 2123 : : samples(samples)
13 2123 : , className(className)
14 2123 : , features(features)
15 2123 : , classNumStates(classNumStates)
16 : {
17 2123 : }
18 : //samples is n+1xm std::vector used to fit the model
19 96 : 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 96 : : samples(torch::zeros({ static_cast<int>(vsamples.size() + 1), static_cast<int>(vsamples[0].size()) }, torch::kInt32))
21 96 : , className(className)
22 96 : , features(features)
23 96 : , classNumStates(classNumStates)
24 : {
25 768 : for (int i = 0; i < vsamples.size(); ++i) {
26 2688 : samples.index_put_({ i, "..." }, torch::tensor(vsamples[i], torch::kInt32));
27 : }
28 384 : samples.index_put_({ -1, "..." }, torch::tensor(labels, torch::kInt32));
29 864 : }
30 478 : std::vector<int> Metrics::SelectKBestWeighted(const torch::Tensor& weights, bool ascending, unsigned k)
31 : {
32 : // Return the K Best features
33 478 : auto n = features.size();
34 478 : if (k == 0) {
35 6 : k = n;
36 : }
37 : // compute scores
38 478 : scoresKBest.clear();
39 478 : featuresKBest.clear();
40 1434 : auto label = samples.index({ -1, "..." });
41 10522 : for (int i = 0; i < n; ++i) {
42 30132 : scoresKBest.push_back(mutualInformation(label, samples.index({ i, "..." }), weights));
43 10044 : featuresKBest.push_back(i);
44 : }
45 : // sort & reduce scores and features
46 478 : if (ascending) {
47 94 : sort(featuresKBest.begin(), featuresKBest.end(), [&](int i, int j)
48 2088 : { return scoresKBest[i] < scoresKBest[j]; });
49 94 : sort(scoresKBest.begin(), scoresKBest.end(), std::less<double>());
50 94 : if (k < n) {
51 154 : for (int i = 0; i < n - k; ++i) {
52 110 : featuresKBest.erase(featuresKBest.begin());
53 110 : scoresKBest.erase(scoresKBest.begin());
54 : }
55 : }
56 : } else {
57 384 : sort(featuresKBest.begin(), featuresKBest.end(), [&](int i, int j)
58 64808 : { return scoresKBest[i] > scoresKBest[j]; });
59 384 : sort(scoresKBest.begin(), scoresKBest.end(), std::greater<double>());
60 384 : featuresKBest.resize(k);
61 384 : scoresKBest.resize(k);
62 : }
63 956 : return featuresKBest;
64 11000 : }
65 48 : std::vector<double> Metrics::getScoresKBest() const
66 : {
67 48 : return scoresKBest;
68 : }
69 :
70 152 : torch::Tensor Metrics::conditionalEdge(const torch::Tensor& weights)
71 : {
72 152 : auto result = std::vector<double>();
73 152 : auto source = std::vector<std::string>(features);
74 152 : source.push_back(className);
75 152 : auto combinations = doCombinations(source);
76 : // Compute class prior
77 152 : auto margin = torch::zeros({ classNumStates }, torch::kFloat);
78 828 : for (int value = 0; value < classNumStates; ++value) {
79 2704 : auto mask = samples.index({ -1, "..." }) == value;
80 676 : margin[value] = mask.sum().item<double>() / samples.size(1);
81 676 : }
82 4164 : for (auto [first, second] : combinations) {
83 4012 : int index_first = find(features.begin(), features.end(), first) - features.begin();
84 4012 : int index_second = find(features.begin(), features.end(), second) - features.begin();
85 4012 : double accumulated = 0;
86 23820 : for (int value = 0; value < classNumStates; ++value) {
87 79232 : auto mask = samples.index({ -1, "..." }) == value;
88 59424 : auto first_dataset = samples.index({ index_first, mask });
89 59424 : auto second_dataset = samples.index({ index_second, mask });
90 39616 : auto weights_dataset = weights.index({ mask });
91 39616 : auto mi = mutualInformation(first_dataset, second_dataset, weights_dataset);
92 19808 : auto pb = margin[value].item<double>();
93 19808 : accumulated += pb * mi;
94 19808 : }
95 4012 : result.push_back(accumulated);
96 4012 : }
97 152 : long n_vars = source.size();
98 152 : auto matrix = torch::zeros({ n_vars, n_vars });
99 152 : auto indices = torch::triu_indices(n_vars, n_vars, 1);
100 4164 : for (auto i = 0; i < result.size(); ++i) {
101 4012 : auto x = indices[0][i];
102 4012 : auto y = indices[1][i];
103 4012 : matrix[x][y] = result[i];
104 4012 : matrix[y][x] = result[i];
105 4012 : }
106 304 : return matrix;
107 99868 : }
108 41732 : double Metrics::entropy(const torch::Tensor& feature, const torch::Tensor& weights)
109 : {
110 41732 : torch::Tensor counts = feature.bincount(weights);
111 41732 : double totalWeight = counts.sum().item<double>();
112 41732 : torch::Tensor probs = counts.to(torch::kFloat) / totalWeight;
113 41732 : torch::Tensor logProbs = torch::log(probs);
114 41732 : torch::Tensor entropy = -probs * logProbs;
115 83464 : return entropy.nansum().item<double>();
116 41732 : }
117 : // H(Y|X) = sum_{x in X} p(x) H(Y|X=x)
118 34276 : double Metrics::conditionalEntropy(const torch::Tensor& firstFeature, const torch::Tensor& secondFeature, const torch::Tensor& weights)
119 : {
120 34276 : int numSamples = firstFeature.sizes()[0];
121 34276 : torch::Tensor featureCounts = secondFeature.bincount(weights);
122 34276 : std::unordered_map<int, std::unordered_map<int, double>> jointCounts;
123 34276 : double totalWeight = 0;
124 6993324 : for (auto i = 0; i < numSamples; i++) {
125 6959048 : jointCounts[secondFeature[i].item<int>()][firstFeature[i].item<int>()] += weights[i].item<double>();
126 6959048 : totalWeight += weights[i].item<float>();
127 : }
128 34276 : if (totalWeight == 0)
129 0 : return 0;
130 34276 : double entropyValue = 0;
131 168251 : for (int value = 0; value < featureCounts.sizes()[0]; ++value) {
132 133975 : double p_f = featureCounts[value].item<double>() / totalWeight;
133 133975 : double entropy_f = 0;
134 454356 : for (auto& [label, jointCount] : jointCounts[value]) {
135 320381 : double p_l_f = jointCount / featureCounts[value].item<double>();
136 320381 : if (p_l_f > 0) {
137 320381 : entropy_f -= p_l_f * log(p_l_f);
138 : } else {
139 0 : entropy_f = 0;
140 : }
141 : }
142 133975 : entropyValue += p_f * entropy_f;
143 : }
144 34276 : return entropyValue;
145 34276 : }
146 : // I(X;Y) = H(Y) - H(Y|X)
147 34276 : double Metrics::mutualInformation(const torch::Tensor& firstFeature, const torch::Tensor& secondFeature, const torch::Tensor& weights)
148 : {
149 34276 : return entropy(firstFeature, weights) - conditionalEntropy(firstFeature, secondFeature, weights);
150 : }
151 : /*
152 : Compute the maximum spanning tree considering the weights as distances
153 : and the indices of the weights as nodes of this square matrix using
154 : Kruskal algorithm
155 : */
156 148 : std::vector<std::pair<int, int>> Metrics::maximumSpanningTree(const std::vector<std::string>& features, const torch::Tensor& weights, const int root)
157 : {
158 148 : auto mst = MST(features, weights, root);
159 296 : return mst.maximumSpanningTree();
160 148 : }
161 : }
|
