From 083d287f4e36f3fd44dad58483516aa7f7536193 Mon Sep 17 00:00:00 2001
From: mkirsz <s1351949@sms.ed.ac.uk>
Date: Tue, 24 Dec 2024 23:49:26 +0000
Subject: [PATCH] Fixing of ML params, also shifted cos cutoff
---
include/tadah/models/cutoffs.h | 15 +++++
include/tadah/models/descriptors/d_base.h | 1 +
include/tadah/models/m_blr_train.h | 63 +++++++++++++++++--
include/tadah/models/m_krr_train.h | 73 +++++++++++++++++++----
include/tadah/models/m_train.h | 64 ++++++++++++++++++--
src/cutoffs.cpp | 42 +++++++++++++
src/m_train.cpp | 4 ++
tests/test_cutoffs.cpp | 43 +++++++++++++
tests/test_svd.cpp | 1 -
9 files changed, 286 insertions(+), 20 deletions(-)
create mode 100644 src/m_train.cpp
diff --git a/include/tadah/models/cutoffs.h b/include/tadah/models/cutoffs.h
index 5403645..882f0c9 100644
--- a/include/tadah/models/cutoffs.h
+++ b/include/tadah/models/cutoffs.h
@@ -139,5 +139,20 @@ class Cut_Poly2 : public Cut_Base {
double calc(double r);
double calc_prime(double r);
};
+class Cut_Cos_S : public Cut_Base {
+ private:
+ std::string lab = "Cut_Cos_S";
+ double rcut, rcut_sq, rcut_inv;
+ double rcut_inner;
+ public:
+ Cut_Cos_S();
+ Cut_Cos_S(double rcut);
+ std::string label() ;
+ void set_rcut(const double r);
+ double get_rcut();
+ double get_rcut_sq();
+ double calc(double r);
+ double calc_prime(double r);
+};
//template<> inline Registry<Cut_Base,double>::Map Registry<Cut_Base,double>::registry{};
#endif
diff --git a/include/tadah/models/descriptors/d_base.h b/include/tadah/models/descriptors/d_base.h
index e3499b8..e02737c 100644
--- a/include/tadah/models/descriptors/d_base.h
+++ b/include/tadah/models/descriptors/d_base.h
@@ -41,6 +41,7 @@ protected:
public:
std::vector<std::string> keys; // keys required by this descriptor
size_t nparams; // number of params which follow TYPExB
+ std::vector<bool> is_optimizable;
static void get_grid(const Config &c, std::string key, v_type &v);
static v_type get_grid(std::vector<std::string>&); // expand joined grids
int verbose;
diff --git a/include/tadah/models/m_blr_train.h b/include/tadah/models/m_blr_train.h
index d78bd43..fdbc457 100644
--- a/include/tadah/models/m_blr_train.h
+++ b/include/tadah/models/m_blr_train.h
@@ -1,6 +1,8 @@
#ifndef M_BLR_TRAIN_H
#define M_BLR_TRAIN_H
+#include "tadah/core/core_types.h"
+#include "tadah/core/utils.h"
#include <tadah/models/m_train.h>
#include <tadah/models/m_blr_core.h>
#include <tadah/models/linear_regressor.h>
@@ -9,6 +11,7 @@
#include <tadah/core/config.h>
#include <iostream>
+#include <vector>
/**
* @class M_BLR_Train
@@ -54,13 +57,63 @@ public:
* @param T Target vector.
* @throws std::runtime_error if the object is already trained.
*/
- void train(phi_type &Phi, t_type &T) {
- if (trained) {
- throw std::runtime_error("This object is already trained!");
+ void train(phi_type &Phi, t_type &T) {
+ if (trained) {
+ throw std::runtime_error("This object is already trained!");
+ }
+
+ if (config.exist("FIXWEIGHT") && config.exist("FIXINDEX")) {
+ std::vector<std::string> indices_str(config.size("FIXINDEX"));
+ config.get("FIXINDEX", indices_str);
+
+ // Parse indices and perform checks
+ std::vector<size_t> indices = parse_indices(indices_str);
+
+ // Check that the min index is >= 1 and max index is <= Phi.cols()
+ if (!indices.empty()) {
+ size_t max_index = *std::max_element(indices.begin(), indices.end());
+ if (*std::min_element(indices.begin(), indices.end()) < 1 || max_index > Phi.cols()) {
+ throw std::runtime_error("FIXINDEX: Indices out of bounds: valid range is from 1 to " + std::to_string(Phi.cols()));
}
- LinearRegressor::train(config, Phi, T, weights, Sigma);
- trained = true;
+ }
+
+ // Adjust for 0-based indexing
+ for (auto &i : indices) i--;
+
+ t_type w_f(config.size("FIXWEIGHT"));
+ if (w_f.size() != indices.size()) {
+ throw std::runtime_error("FIXWEIGHT and FIXINDEX differ in size: " + std::to_string(w_f.size()) + " and " + std::to_string(indices.size()));
+ }
+
+ config.get("FIXWEIGHT", w_f);
+
+ t_type T_r;
+ auto move_map = prep_train_with_residuals(Phi, T, indices, w_f, T_r);
+
+ // Resize Phi for training
+ Phi.resize(Phi.rows(), Phi.cols() - indices.size());
+
+ LinearRegressor::train(config, Phi, T_r, weights, Sigma);
+
+ t_type w_temp(weights.size() + indices.size());
+
+ for (size_t i = 0; i < w_f.size(); ++i) {
+ w_temp[w_temp.size() - w_f.size() + i] = w_f[i];
+ }
+
+ for (size_t i = 0; i < weights.size(); ++i) {
+ w_temp[i] = weights[i];
+ }
+
+ weights = w_temp;
+ reverse_vector(weights, move_map);
+ trained = true;
+ }
+ else {
+ LinearRegressor::train(config, Phi, T, weights, Sigma);
+ trained = true;
}
+ }
};
#endif // M_BLR_TRAIN_H
diff --git a/include/tadah/models/m_krr_train.h b/include/tadah/models/m_krr_train.h
index 24e7d83..d4e7903 100644
--- a/include/tadah/models/m_krr_train.h
+++ b/include/tadah/models/m_krr_train.h
@@ -60,21 +60,74 @@ public:
* @param T Target vector.
* @throws std::runtime_error if the object is already trained.
*/
- void train(phi_type &Phi, t_type &T) {
- if (trained) {
- throw std::runtime_error("This object is already trained!");
- }
- if (kernel.get_label() != "Kern_Linear") {
- ekm.project(Phi);
+ void train(phi_type &Phi, t_type &T) {
+ if (trained) {
+ throw std::runtime_error("This object is already trained!");
+ }
+
+ if (config.exist("FIXWEIGHT") && config.exist("FIXINDEX")) {
+ std::vector<std::string> indices_str(config.size("FIXINDEX"));
+ config.get("FIXINDEX", indices_str);
+
+ // Parse indices and perform checks
+ std::vector<size_t> indices = parse_indices(indices_str);
+
+ // Check that the min index is >= 1 and max index is <= Phi.cols()
+ if (!indices.empty()) {
+ size_t max_index = *std::max_element(indices.begin(), indices.end());
+ if (*std::min_element(indices.begin(), indices.end()) < 1 || max_index > Phi.cols()) {
+ throw std::runtime_error("FIXINDEX: Indices out of bounds: valid range is from 1 to " + std::to_string(Phi.cols()));
}
- LinearRegressor::train(config, Phi, T, weights, Sigma);
+ }
+
+ // Adjust for 0-based indexing
+ for (auto &i : indices) i--;
+
+ t_type w_f(config.size("FIXWEIGHT"));
+ if (w_f.size() != indices.size()) {
+ throw std::runtime_error("FIXWEIGHT and FIXINDEX differ in size: " + std::to_string(w_f.size()) + " and " + std::to_string(indices.size()));
+ }
+
+ config.get("FIXWEIGHT", w_f);
- if (kernel.get_label() != "Kern_Linear") {
- weights = ekm.EKM_mat * weights;
+ t_type T_r;
+ auto move_map = prep_train_with_residuals(Phi, T, indices, w_f, T_r);
+
+ reverse_columns(Phi, move_map);
+
+ for (const auto &i : indices) {
+ for (size_t j=0; j<Phi.rows(); ++j) {
+ Phi(j,i)=0;
}
- trained = true;
+ }
+
+ if (kernel.get_label() != "Kern_Linear") {
+ ekm.project(Phi);
+ }
+ LinearRegressor::train(config, Phi, T_r, weights, Sigma);
+ if (kernel.get_label() != "Kern_Linear") {
+ weights = ekm.EKM_mat * weights;
+ }
+
+ // reverse_vector(weights, move_map);
+ for (size_t i=0; i<indices.size(); ++i) {
+ weights[indices[i]]=w_f[i];
+ }
+ trained = true;
+ }
+ else {
+ if (kernel.get_label() != "Kern_Linear") {
+ ekm.project(Phi);
+ }
+ LinearRegressor::train(config, Phi, T, weights, Sigma);
+ if (kernel.get_label() != "Kern_Linear") {
+ weights = ekm.EKM_mat * weights;
+ }
+ trained = true;
}
+ }
+
/**
* @brief Standard KRR training using covariance matrix computation.
*
diff --git a/include/tadah/models/m_train.h b/include/tadah/models/m_train.h
index abcf824..09a2113 100644
--- a/include/tadah/models/m_train.h
+++ b/include/tadah/models/m_train.h
@@ -2,6 +2,7 @@
#define M_TRAIN_H
#include <tadah/core/core_types.h>
+#include <tadah/core/utils.h>
/**
* @class M_Train
@@ -11,12 +12,12 @@
*/
class M_Train {
public:
- /**
+ /**
* @brief Virtual destructor for polymorphic deletion.
*/
- virtual ~M_Train() {}
+ virtual ~M_Train() {}
- /**
+ /**
* @brief Pure virtual function to train the model.
*
* Must be implemented by derived classes.
@@ -24,7 +25,62 @@ public:
* @param Phi Design matrix containing input features.
* @param T Target vector for training.
*/
- virtual void train(phi_type &Phi, t_type &T) = 0;
+ virtual void train(phi_type &Phi, t_type &T) = 0;
+
+ /**
+ * Swaps columns and fits a reduced matrix to obtain residuals.
+ *
+ * This function processes a full matrix Phi by swapping columns specified by indices,
+ * creating a new pointer for the fixed matrix Phi_f, and then use w_f weights
+ * to obtain prediction T_f = Phi_f w_f. The function then copmutes the residual vector
+ * T_r = T - T_f. On exit, the Phi matrix is rearranged such that the indices.size()
+ * rightmost columns of the original matrix are the same as Phi_f.
+ * The function return a mapping between columns of original Phi matrix and rearranged.
+ *
+ * @param Phi The full matrix to be reduced and fitted to the adjusted target
+ * vector.
+ * @param T The full target vector to which the reduced matrix Phi will
+ * be fitted.
+ * @param indices A vector of indices specifying which columns in the
+ * original matrix are fixed and should be considered during
+ * the regression.
+ * @param w_f A vector of fixed weights corresponding to the specified
+ * indices. These weights are used to adjust the target
+ * vector, resulting in residuals.
+ */
+ template <typename M>
+ std::vector<size_t> prep_train_with_residuals(
+ Matrix_Base<M> &Phi, t_type &T, std::vector<size_t> &indices, const t_type w_f, t_type &T_r) {
+
+ if (w_f.size() != indices.size()) {
+ throw std::runtime_error("Size of w_f must be equal to size of indices.");
+ }
+
+ std::sort(indices.begin(), indices.end());
+
+ if (indices.back() >= Phi.cols()) {
+ throw std::runtime_error("Largest index is greater than the number of columns in Phi.");
+ }
+
+ std::vector<size_t> move_map = move_columns_in_place(Phi,indices);
+
+ // indices.size() = 1
+ // indices[0] 1
+ // 1 4 7 1 7 4
+ // 2 5 8 -> 2 8 5
+ // 3 6 9 3 9 6
+ // Phi Phi_o Phi_f
+ // move_map = {0,2,1}
+
+ // the fixed columns are to the right
+ MatrixView Phi_f(Phi, Phi.cols()-indices.size());
+
+ t_type T_f = Phi_f * w_f;
+
+ T_r = T-T_f;
+
+ return move_map;
+ }
};
#endif // M_TRAIN_H
diff --git a/src/cutoffs.cpp b/src/cutoffs.cpp
index 768516a..b818cda 100644
--- a/src/cutoffs.cpp
+++ b/src/cutoffs.cpp
@@ -5,6 +5,7 @@
template<> CONFIG::Registry<Cut_Base,double>::Map CONFIG::Registry<Cut_Base,double>::registry{};
CONFIG::Registry<Cut_Base,double>::Register<Cut_Cos> Cut_Cos_1( "Cut_Cos" );
+CONFIG::Registry<Cut_Base,double>::Register<Cut_Cos_S> Cut_Cos_S_1( "Cut_Cos_S" );
CONFIG::Registry<Cut_Base,double>::Register<Cut_Tanh> Cut_Tanh_1( "Cut_Tanh" );
CONFIG::Registry<Cut_Base,double>::Register<Cut_Poly2> Cut_Poly_1( "Cut_Poly2" );
CONFIG::Registry<Cut_Base,double>::Register<Cut_Dummy> Cut_Dummy_1( "Cut_Dummy" );
@@ -165,3 +166,44 @@ double Cut_Poly2::calc_prime(double r) {
double rs=r-rcut_inner;
return -30.0*(rs-1.0)*(rs-1.0)*rs*rs;
}
+
+Cut_Cos_S::Cut_Cos_S() {}
+Cut_Cos_S::Cut_Cos_S(double rcut)
+{
+ set_rcut(rcut);
+ test_rcut(rcut);
+}
+
+std::string Cut_Cos_S::label() {
+ return lab;
+}
+
+void Cut_Cos_S::set_rcut(const double r) {
+ test_rcut(r);
+ rcut=r;
+ rcut_sq=r*r;
+ rcut_inv= r<=0 ? 0.0 : 1.0/r;
+ rcut_inner=rcut-1.0;
+}
+double Cut_Cos_S::get_rcut() {
+ return rcut;
+}
+
+double Cut_Cos_S::get_rcut_sq() {
+ return rcut_sq;
+}
+
+double Cut_Cos_S::calc(double r) {
+ if (r>=rcut) return 0.0;
+ else if (r<= rcut_inner) return 1.0;
+ double rs = (r-rcut_inner)/(rcut-rcut_inner);
+ return 0.5*(1+std::cos(M_PI*rs));
+}
+double Cut_Cos_S::calc_prime(double r) {
+ if (r>=rcut || r<= rcut_inner) return 0.0;
+ else {
+ double rs = (r - rcut_inner) / (rcut - rcut_inner);
+ double drs_dr = 1.0 / (rcut - rcut_inner);
+ return -0.5 * M_PI * std::sin(M_PI * rs) * drs_dr;
+ }
+}
diff --git a/src/m_train.cpp b/src/m_train.cpp
new file mode 100644
index 0000000..3bc76b5
--- /dev/null
+++ b/src/m_train.cpp
@@ -0,0 +1,4 @@
+#include <tadah/models/m_train.h>
+
+// void M_Train::train_with_residuals(
+// }
diff --git a/tests/test_cutoffs.cpp b/tests/test_cutoffs.cpp
index 997b03d..aaac9d0 100644
--- a/tests/test_cutoffs.cpp
+++ b/tests/test_cutoffs.cpp
@@ -162,3 +162,46 @@ TEST_CASE( "Testing Cutoffs: Cut_Poly2", "[Cut_Poly2]" ) {
Catch::Matchers::WithinAbs(-1.728, 1e-12));
delete c2b;
}
+TEST_CASE( "Testing Cutoffs: Cut_Cos_s", "[Cut_Cos_S]" ) {
+
+ REQUIRE_NOTHROW(Cut_Cos_S());
+ double rcut2b = 6.2;
+ double rcut2bsq = rcut2b*rcut2b;
+ using Cut = Cut_Cos_S;
+ std::string cuttype = "Cut_Cos_S";
+ Cut_Base *c2b = new Cut( rcut2b );
+
+ REQUIRE( c2b->label() == cuttype );
+
+ REQUIRE( c2b->calc(rcut2b) < std::numeric_limits<double>::min() );
+ REQUIRE( c2b->calc_prime(rcut2b) < std::numeric_limits<double>::min() );
+ REQUIRE( std::abs(c2b->get_rcut()-rcut2b)<std::numeric_limits<double>::min() );
+ REQUIRE( std::abs(c2b->get_rcut_sq()-rcut2bsq)<std::numeric_limits<double>::min() );
+
+ // cutoff cannot be negative
+ double temp = -0.1;
+ REQUIRE_THROWS(Cut( temp ));
+ REQUIRE_THROWS_AS(c2b->set_rcut(temp), std::runtime_error);
+ REQUIRE_THROWS_AS(c2b->test_rcut(temp), std::runtime_error);
+
+ // recheck after resetting cutoff
+ rcut2b=3.4;
+ rcut2bsq=rcut2b*rcut2b;
+ c2b->set_rcut(100000);
+ c2b->set_rcut(rcut2b);
+ REQUIRE( c2b->calc(rcut2b) < std::numeric_limits<double>::min() );
+ REQUIRE( c2b->calc_prime(rcut2b) < std::numeric_limits<double>::min() );
+ REQUIRE( std::abs(c2b->get_rcut()-rcut2b)<std::numeric_limits<double>::min() );
+ REQUIRE( std::abs(c2b->get_rcut_sq()-rcut2bsq)<std::numeric_limits<double>::min() );
+
+ REQUIRE_THAT(c2b->calc(2.0),
+ Catch::Matchers::WithinAbs(1, 1e-12));
+ REQUIRE_THAT(c2b->calc_prime(2.0),
+ Catch::Matchers::WithinAbs(0, 1e-12));
+
+ REQUIRE_THAT(c2b->calc(3.0),
+ Catch::Matchers::WithinAbs(0.3454915028, 1e-10));
+ REQUIRE_THAT(c2b->calc_prime(3.0),
+ Catch::Matchers::WithinAbs(-1.4939160824, 1e-10));
+ delete c2b;
+}
diff --git a/tests/test_svd.cpp b/tests/test_svd.cpp
index f216f20..d434096 100644
--- a/tests/test_svd.cpp
+++ b/tests/test_svd.cpp
@@ -1,7 +1,6 @@
#include "catch2/catch.hpp"
#include <tadah/core/maths.h>
#include <tadah/models/svd.h>
-#include <iomanip>
void multiplyMatrices(const double* A, const double* B, double* C, int m, int n, int p) {
// C = A * B
--
GitLab