summaryrefslogtreecommitdiff
path: root/dtrain/test/log_reg_dyer
diff options
context:
space:
mode:
Diffstat (limited to 'dtrain/test/log_reg_dyer')
-rw-r--r--dtrain/test/log_reg_dyer/bin_class.cc4
-rw-r--r--dtrain/test/log_reg_dyer/bin_class.h22
-rw-r--r--dtrain/test/log_reg_dyer/log_reg.cc39
-rw-r--r--dtrain/test/log_reg_dyer/log_reg.h14
4 files changed, 79 insertions, 0 deletions
diff --git a/dtrain/test/log_reg_dyer/bin_class.cc b/dtrain/test/log_reg_dyer/bin_class.cc
new file mode 100644
index 00000000..19bcde25
--- /dev/null
+++ b/dtrain/test/log_reg_dyer/bin_class.cc
@@ -0,0 +1,4 @@
+#include "bin_class.h"
+
+Objective::~Objective() {}
+
diff --git a/dtrain/test/log_reg_dyer/bin_class.h b/dtrain/test/log_reg_dyer/bin_class.h
new file mode 100644
index 00000000..3466109a
--- /dev/null
+++ b/dtrain/test/log_reg_dyer/bin_class.h
@@ -0,0 +1,22 @@
+#ifndef _BIN_CLASS_H_
+#define _BIN_CLASS_H_
+
+#include <vector>
+#include "sparse_vector.h"
+
+struct TrainingInstance {
+ // TODO add other info? loss for MIRA-type updates?
+ SparseVector<double> x_feature_map;
+ bool y;
+};
+
+struct Objective {
+ virtual ~Objective();
+
+ // returns f(x) and f'(x)
+ virtual double ObjectiveAndGradient(const SparseVector<double>& x,
+ const std::vector<TrainingInstance>& training_instances,
+ SparseVector<double>* g) const = 0;
+};
+
+#endif
diff --git a/dtrain/test/log_reg_dyer/log_reg.cc b/dtrain/test/log_reg_dyer/log_reg.cc
new file mode 100644
index 00000000..ec2331fe
--- /dev/null
+++ b/dtrain/test/log_reg_dyer/log_reg.cc
@@ -0,0 +1,39 @@
+#include "log_reg.h"
+
+#include <vector>
+#include <cmath>
+
+#include "sparse_vector.h"
+
+using namespace std;
+
+double LogisticRegression::ObjectiveAndGradient(const SparseVector<double>& x,
+ const vector<TrainingInstance>& training_instances,
+ SparseVector<double>* g) const {
+ double cll = 0;
+ for (int i = 0; i < training_instances.size(); ++i) {
+ const double dotprod = training_instances[i].x_feature_map.dot(x); // TODO no bias, if bias, add x[0]
+ double lp_false = dotprod;
+ double lp_true = -dotprod;
+ if (0 < lp_true) {
+ lp_true += log1p(exp(-lp_true));
+ lp_false = log1p(exp(lp_false));
+ } else {
+ lp_true = log1p(exp(lp_true));
+ lp_false += log1p(exp(-lp_false));
+ }
+ lp_true *= -1;
+ lp_false *= -1;
+ if (training_instances[i].y) { // true label
+ cll -= lp_true;
+ (*g) -= training_instances[i].x_feature_map * exp(lp_false);
+ // (*g)[0] -= exp(lp_false); // bias
+ } else { // false label
+ cll -= lp_false;
+ (*g) += training_instances[i].x_feature_map * exp(lp_true);
+ // g += corpus[i].second * exp(lp_true);
+ }
+ }
+ return cll;
+}
+
diff --git a/dtrain/test/log_reg_dyer/log_reg.h b/dtrain/test/log_reg_dyer/log_reg.h
new file mode 100644
index 00000000..ecc560b8
--- /dev/null
+++ b/dtrain/test/log_reg_dyer/log_reg.h
@@ -0,0 +1,14 @@
+#ifndef _LOG_REG_H_
+#define _LOG_REG_H_
+
+#include <vector>
+#include "sparse_vector.h"
+#include "bin_class.h"
+
+struct LogisticRegression : public Objective {
+ double ObjectiveAndGradient(const SparseVector<double>& x,
+ const std::vector<TrainingInstance>& training_instances,
+ SparseVector<double>* g) const;
+};
+
+#endif