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 "Ensemble.h"
8 :
9 : namespace bayesnet {
10 :
11 324 : Ensemble::Ensemble(bool predict_voting) : Classifier(Network()), n_models(0), predict_voting(predict_voting)
12 : {
13 :
14 324 : };
15 : const std::string ENSEMBLE_NOT_FITTED = "Ensemble has not been fitted";
16 40 : void Ensemble::trainModel(const torch::Tensor& weights)
17 : {
18 40 : n_models = models.size();
19 660 : for (auto i = 0; i < n_models; ++i) {
20 : // fit with std::vectors
21 620 : models[i]->fit(dataset, features, className, states);
22 : }
23 40 : }
24 56 : std::vector<int> Ensemble::compute_arg_max(std::vector<std::vector<double>>& X)
25 : {
26 56 : std::vector<int> y_pred;
27 12400 : for (auto i = 0; i < X.size(); ++i) {
28 12344 : auto max = std::max_element(X[i].begin(), X[i].end());
29 24688 : y_pred.push_back(std::distance(X[i].begin(), max));
30 : }
31 112 : return y_pred;
32 56 : }
33 424 : torch::Tensor Ensemble::compute_arg_max(torch::Tensor& X)
34 : {
35 424 : auto y_pred = torch::argmax(X, 1);
36 848 : return y_pred;
37 424 : }
38 164 : torch::Tensor Ensemble::voting(torch::Tensor& votes)
39 : {
40 : // Convert m x n_models tensor to a m x n_class_states with voting probabilities
41 164 : auto y_pred_ = votes.accessor<int, 2>();
42 164 : std::vector<int> y_pred_final;
43 164 : int numClasses = states.at(className).size();
44 : // votes is m x n_models with the prediction of every model for each sample
45 164 : auto result = torch::zeros({ votes.size(0), numClasses }, torch::kFloat32);
46 164 : auto sum = std::reduce(significanceModels.begin(), significanceModels.end());
47 42084 : for (int i = 0; i < votes.size(0); ++i) {
48 : // n_votes store in each index (value of class) the significance added by each model
49 : // i.e. n_votes[0] contains how much value has the value 0 of class. That value is generated by the models predictions
50 41920 : std::vector<double> n_votes(numClasses, 0.0);
51 375272 : for (int j = 0; j < n_models; ++j) {
52 333352 : n_votes[y_pred_[i][j]] += significanceModels.at(j);
53 : }
54 41920 : result[i] = torch::tensor(n_votes);
55 41920 : }
56 : // To only do one division and gain precision
57 164 : result /= sum;
58 328 : return result;
59 164 : }
60 100 : std::vector<std::vector<double>> Ensemble::predict_proba(std::vector<std::vector<int>>& X)
61 : {
62 100 : if (!fitted) {
63 24 : throw std::logic_error(ENSEMBLE_NOT_FITTED);
64 : }
65 76 : return predict_voting ? predict_average_voting(X) : predict_average_proba(X);
66 : }
67 452 : torch::Tensor Ensemble::predict_proba(torch::Tensor& X)
68 : {
69 452 : if (!fitted) {
70 24 : throw std::logic_error(ENSEMBLE_NOT_FITTED);
71 : }
72 428 : return predict_voting ? predict_average_voting(X) : predict_average_proba(X);
73 : }
74 68 : std::vector<int> Ensemble::predict(std::vector<std::vector<int>>& X)
75 : {
76 68 : auto res = predict_proba(X);
77 104 : return compute_arg_max(res);
78 52 : }
79 436 : torch::Tensor Ensemble::predict(torch::Tensor& X)
80 : {
81 436 : auto res = predict_proba(X);
82 840 : return compute_arg_max(res);
83 420 : }
84 296 : torch::Tensor Ensemble::predict_average_proba(torch::Tensor& X)
85 : {
86 296 : auto n_states = models[0]->getClassNumStates();
87 296 : torch::Tensor y_pred = torch::zeros({ X.size(1), n_states }, torch::kFloat32);
88 296 : auto threads{ std::vector<std::thread>() };
89 296 : std::mutex mtx;
90 1764 : for (auto i = 0; i < n_models; ++i) {
91 1468 : threads.push_back(std::thread([&, i]() {
92 1468 : auto ypredict = models[i]->predict_proba(X);
93 1468 : std::lock_guard<std::mutex> lock(mtx);
94 1468 : y_pred += ypredict * significanceModels[i];
95 1468 : }));
96 : }
97 1764 : for (auto& thread : threads) {
98 1468 : thread.join();
99 : }
100 296 : auto sum = std::reduce(significanceModels.begin(), significanceModels.end());
101 296 : y_pred /= sum;
102 592 : return y_pred;
103 296 : }
104 44 : std::vector<std::vector<double>> Ensemble::predict_average_proba(std::vector<std::vector<int>>& X)
105 : {
106 44 : auto n_states = models[0]->getClassNumStates();
107 44 : std::vector<std::vector<double>> y_pred(X[0].size(), std::vector<double>(n_states, 0.0));
108 44 : auto threads{ std::vector<std::thread>() };
109 44 : std::mutex mtx;
110 576 : for (auto i = 0; i < n_models; ++i) {
111 532 : threads.push_back(std::thread([&, i]() {
112 532 : auto ypredict = models[i]->predict_proba(X);
113 532 : assert(ypredict.size() == y_pred.size());
114 532 : assert(ypredict[0].size() == y_pred[0].size());
115 532 : std::lock_guard<std::mutex> lock(mtx);
116 : // Multiply each prediction by the significance of the model and then add it to the final prediction
117 110284 : for (auto j = 0; j < ypredict.size(); ++j) {
118 109752 : std::transform(y_pred[j].begin(), y_pred[j].end(), ypredict[j].begin(), y_pred[j].begin(),
119 739464 : [significanceModels = significanceModels[i]](double x, double y) { return x + y * significanceModels; });
120 : }
121 532 : }));
122 : }
123 576 : for (auto& thread : threads) {
124 532 : thread.join();
125 : }
126 44 : auto sum = std::reduce(significanceModels.begin(), significanceModels.end());
127 : //Divide each element of the prediction by the sum of the significances
128 8436 : for (auto j = 0; j < y_pred.size(); ++j) {
129 51544 : std::transform(y_pred[j].begin(), y_pred[j].end(), y_pred[j].begin(), [sum](double x) { return x / sum; });
130 : }
131 88 : return y_pred;
132 44 : }
133 32 : std::vector<std::vector<double>> Ensemble::predict_average_voting(std::vector<std::vector<int>>& X)
134 : {
135 32 : torch::Tensor Xt = bayesnet::vectorToTensor(X, false);
136 32 : auto y_pred = predict_average_voting(Xt);
137 32 : std::vector<std::vector<double>> result = tensorToVectorDouble(y_pred);
138 64 : return result;
139 32 : }
140 164 : torch::Tensor Ensemble::predict_average_voting(torch::Tensor& X)
141 : {
142 : // Build a m x n_models tensor with the predictions of each model
143 164 : torch::Tensor y_pred = torch::zeros({ X.size(1), n_models }, torch::kInt32);
144 164 : auto threads{ std::vector<std::thread>() };
145 164 : std::mutex mtx;
146 1380 : for (auto i = 0; i < n_models; ++i) {
147 1216 : threads.push_back(std::thread([&, i]() {
148 1216 : auto ypredict = models[i]->predict(X);
149 1216 : std::lock_guard<std::mutex> lock(mtx);
150 3648 : y_pred.index_put_({ "...", i }, ypredict);
151 2432 : }));
152 : }
153 1380 : for (auto& thread : threads) {
154 1216 : thread.join();
155 : }
156 328 : return voting(y_pred);
157 164 : }
158 80 : float Ensemble::score(torch::Tensor& X, torch::Tensor& y)
159 : {
160 80 : auto y_pred = predict(X);
161 72 : int correct = 0;
162 22584 : for (int i = 0; i < y_pred.size(0); ++i) {
163 22512 : if (y_pred[i].item<int>() == y[i].item<int>()) {
164 19668 : correct++;
165 : }
166 : }
167 144 : return (double)correct / y_pred.size(0);
168 72 : }
169 52 : float Ensemble::score(std::vector<std::vector<int>>& X, std::vector<int>& y)
170 : {
171 52 : auto y_pred = predict(X);
172 44 : int correct = 0;
173 11164 : for (int i = 0; i < y_pred.size(); ++i) {
174 11120 : if (y_pred[i] == y[i]) {
175 9276 : correct++;
176 : }
177 : }
178 88 : return (double)correct / y_pred.size();
179 44 : }
180 4 : std::vector<std::string> Ensemble::show() const
181 : {
182 4 : auto result = std::vector<std::string>();
183 20 : for (auto i = 0; i < n_models; ++i) {
184 16 : auto res = models[i]->show();
185 16 : result.insert(result.end(), res.begin(), res.end());
186 16 : }
187 8 : return result;
188 4 : }
189 16 : std::vector<std::string> Ensemble::graph(const std::string& title) const
190 : {
191 16 : auto result = std::vector<std::string>();
192 108 : for (auto i = 0; i < n_models; ++i) {
193 92 : auto res = models[i]->graph(title + "_" + std::to_string(i));
194 92 : result.insert(result.end(), res.begin(), res.end());
195 92 : }
196 32 : return result;
197 16 : }
198 28 : int Ensemble::getNumberOfNodes() const
199 : {
200 28 : int nodes = 0;
201 348 : for (auto i = 0; i < n_models; ++i) {
202 320 : nodes += models[i]->getNumberOfNodes();
203 : }
204 28 : return nodes;
205 : }
206 28 : int Ensemble::getNumberOfEdges() const
207 : {
208 28 : int edges = 0;
209 348 : for (auto i = 0; i < n_models; ++i) {
210 320 : edges += models[i]->getNumberOfEdges();
211 : }
212 28 : return edges;
213 : }
214 4 : int Ensemble::getNumberOfStates() const
215 : {
216 4 : int nstates = 0;
217 20 : for (auto i = 0; i < n_models; ++i) {
218 16 : nstates += models[i]->getNumberOfStates();
219 : }
220 4 : return nstates;
221 : }
222 : }
|