HMC random walks #38

TolisChal · 2019-08-20T07:38:13Z

This PR contains:

Implementations of HMC random walk when the hamiltonian is the logbarrier function of the polytope. We have 4 different implementation of the same random walk.
Implementation of the HMC random walk with reflections when the hamiltonian is the spherical Gaussian distribution. We use this walk to CG volume computation method yielding a faster method for H-polytopes

…ssed locally.

…o compute the first point of the CDHR and not to test in every step if this is the first point. We initialize min_plus to max double value and max_min to min double value to avoid checking in every iteration if they are seted for the first time. We add comments in Polytope class and we remove useles comments from samplers.h header file.

… type in volume.h header file and in vol.cpp file. We have to use template number type in nikolic implementation in order to use different number type than double. Pass all the test locally.

… irrespective of the number type we use for the implementaion. We can use float or long double. We made adjustments in tests cpp files in order to use the template number type. We pass all the tests for both double and float.

… declared in volume.h. Now we can use float or long double in Rcpp interface as well.

Use eigen matrix for the H-polytope representation. Able to use float or long double.

… in order to sample the first point for the coordinate directions HnR. Make adjustments in get_annealing_schedule() and in volume_gaussian_annealing() for the first point coordinate case.

…he ratios in gaussian_annealing_volume. Use const bool for the computation of the first coord point.

…e class point_on_Dsphere. We move functions point_in_Dsphere, point_on_Dsphere from gaussian_samplers.h to samplers.h and implement function get_direction() to avoid duplicated code. We use these functions for RDHR and Gausian Ball Walk and first random point in convex polytope.

…r max nuber of iterations and tolerance in gaussian_samplers.h and gaussian_annealing.h header files.

…an(), we have moved the output arguments at the end of the argument list. We change all not returning functions to void in gaussian_annealing.h and in gaussian_samplers.h. We add comments for the variable k in get_next_gaussian and iterators for the fn vector. We have renamed its to iterators. We change rand_exp_range_coord() to return a double.

…positions and values of the window W.

… Add to Rcpp interface the choice to run CV algo.

…for the CV algo. The new pdf is in doc folder.

…for CV algo in volumeCV_test.cpp. We have passed succesfully the test locally

…t_coord_point() to compute first gaussian point, overriding is not needed.

… set error=0.2 and number of experiments=20 in the CV volume test. We passed the test succesfully locally.

…ith eigen matrix in Vpolytope class. Use template number type in VPolytope class.

…tope by a point c. Create function linear_transformIt() for both H and V polytopes, to apply a linear transformation to the polytope. We use these functions to the main rounding functionin order to use the latter for Vpolytopes as well. We replace all the shiftings in CV algorithm with the function shift().

…turns the distances between the chebychev center and the facets. In Vpolytope it computes an upper bound for the number of facets (number of d-1-faces of the cyclic polytope) and for each facet returns the radius of the chebychev ball as lower bound for the distance, so we can compute an upper bound for the a_0 of the first gaussian. CV algo can now be used for the volume approximation of v polytopes.We add boolean deltaset to var_g to be able to set the radius of the ball walk when the a_i changes (that was a bug that is fixed).

…olytope returns false and V polytope returns its vertices if num_of_vertices<20*dimension. We modified rotating() function to use linear_transformIt() polytope's function, now we can rotate a V polytope as well.

…lytope classes. We remove Init(int d) and constructor with argument int d in V polytope. We create upper_bound_of_hyperplanes in V polytope class. We change input and output number type to NT in factorial function, because of overflow.

…g double. In main file we have to incluse the disearable header file. Fix some bugs for the template number types in vol.cpp and in struct var_g in volume.h. We define const boolean using_float because we use it in get_first_gaussian() function, to guarantee convergence: If float is used we set tol=1e-06 otherwise we set it 1e-07.

… with long double.

…end of solve_LP() function that computes the chebychev center of a H-polytope. We add some comments in samplers.h and use NT number type where it is not able to use a tamplate.

…ve to give the path to the ext file and a flag Vpoly=TRUE.

…rtices of a V-polytope as an input. Remove variable telescopic product from volesti and set initial volume equl to the volume of the chebychev ball to avoid overflow in some cases (in each iteration we multiply vol with the current ratio).

vissarion

Thanks for this PR, it seems that (at least some versions of) the new random walks are very promising. However, the code needs more work, so looking forward to your feedback. Please explain for which walks you add an R interface and why. Also you say the new volume algorithm is faster, please include here few simple runs for illustration. Finally, tests and docs and comments on the code regarding references of the implemented algorithms.

vissarion · 2019-09-11T13:52:34Z

R-proj/src/sample_points.cpp

@@ -10,6 +10,7 @@
 #include <Rcpp.h>
 #include <RcppEigen.h>
 #include <chrono>
+#define VOLESTI_DEBUG


Should this be removed?

vissarion · 2019-09-11T13:53:18Z

R-proj/src/sample_points.cpp

@@ -101,7 +104,8 @@ Rcpp::NumericMatrix sample_points(Rcpp::Nullable<Rcpp::Reference> P = R_NilValue

    int type, dim, numpoints;
    NT radius = 1.0, delta = -1.0;
-    bool set_mean_point = false, cdhr = true, rdhr = false, ball_walk = false, gaussian = false;
+    bool set_mean_point = false, cdhr = false, rdhr = false, ball_walk = false, gaussian = false, gibbs = false,


Similar to other PR's please change this to use enums instead of many bool vars.

see e.g. #39

vissarion · 2019-09-11T13:54:44Z

R-proj/src/sample_points.cpp

            rdhr = true;
-            ball_walk = false;
+        } else if (Rcpp::as<std::string>(WalkType).compare(std::string("HMC"))==0) {
+            if (type != 1) throw Rcpp::exception("HMC can be applies only to H-polytopes!");


"can be applied ..."

vissarion · 2019-09-11T13:55:06Z

R-proj/src/sample_points.cpp

@@ -239,6 +248,7 @@ Rcpp::NumericMatrix sample_points(Rcpp::Nullable<Rcpp::Reference> P = R_NilValue
                    InnerBall = HP.ComputeInnerBall();
                    if (!set_mean_point) MeanPoint = InnerBall.first;
                }
+                //if (hmc_barrier && !gaussian) a = 1.0/NT(Rcpp::as<MT>(Rcpp::as<Rcpp::Reference>(P).field("A")).rows());


is this needed?

vissarion · 2019-09-11T13:58:23Z

include/convex_bodies/polytopes.h

@@ -419,6 +419,7 @@ class HPolytope{
    bool get_points_for_rounding (T &randPoints) {
        return false;
    }
+


are those new lines needed or made by mistake by some IDE?

vissarion · 2019-09-11T14:02:02Z

include/samplers/gaussian_samplers.h

@@ -327,4 +346,47 @@ void gaussian_ball_walk(Point &p,
    }
 }

+/*


is this needed?

vissarion · 2019-09-11T14:04:33Z

include/samplers/hmc_heur_sam.h

@@ -0,0 +1,112 @@
+// VolEsti (volume computation and sampling library)


I propose to use a less cryptic name for the file i.e. sampler instead of sam, what is heur stands for?

vissarion · 2019-09-11T14:05:12Z

include/samplers/hmc_refl.h

@@ -0,0 +1,164 @@
+// VolEsti (volume computation and sampling library)


filename, hmc_reflections.h ?

vissarion · 2019-09-11T14:05:36Z

include/samplers/hmc_refl.h

+
+// Licensed under GNU LGPL.3, see LICENCE file
+
+#ifndef HMC_REFL_H


also please update those wrt the filename

vissarion · 2019-09-11T14:06:13Z

include/samplers/hmc_rk4.h

@@ -0,0 +1,167 @@
+// VolEsti (volume computation and sampling library)


filename. rk4 ?!

vissarion · 2021-09-21T12:49:32Z

Parts of this PR appear in #142 and #144

TolisChal added 30 commits July 31, 2018 14:22

use template number type everywhere in H-polytope class. All tests pa…

c082552

…ssed locally.

Use template number type in BallIntersevtPolytope class and NT number…

7bac1ca

… type in volume.h header file and in vol.cpp file. We have to use template number type in nikolic implementation in order to use different number type than double. Pass all the test locally.

Make adjustments in vol_R.cpp in order to use NT number type as it is…

de9ee55

… declared in volume.h. Now we can use float or long double in Rcpp interface as well.

Merge pull request #6 from TolisChal/use_eigen_matrix

52da8e8

Use eigen matrix for the H-polytope representation. Able to use float or long double.

merge develop into schedule_annealing. Override gaussian_next_point()…

e884e5c

… in order to sample the first point for the coordinate directions HnR. Make adjustments in get_annealing_schedule() and in volume_gaussian_annealing() for the first point coordinate case.

move computation of the first coordinate point outside the loops of t…

b726611

…he ratios in gaussian_annealing_volume. Use const bool for the computation of the first coord point.

Implement Welford's algorithm for mean and variance. Use constants fo…

98eb4ed

…r max nuber of iterations and tolerance in gaussian_samplers.h and gaussian_annealing.h header files.

We use std min_element and max_element iterators for the min and max …

f43c112

…positions and values of the window W.

change to void all the non returning functions in the implementation.…

f1bb83d

… Add to Rcpp interface the choice to run CV algo.

update R package documentation in rder to explain the new parameters …

b097a69

…for the CV algo. The new pdf is in doc folder.

Update readme file to explain how you can run CV algo. Create a test …

d57a408

…for CV algo in volumeCV_test.cpp. We have passed succesfully the test locally

fix some mistakes in readme file

3c4ade6

Add comments in functions that they had not any. Create gaussian_firs…

feadc6e

…t_coord_point() to compute first gaussian point, overriding is not needed.

add some comments in gaussin_samlrs.h and in gaussian_annealing.h. We…

9a045cc

… set error=0.2 and number of experiments=20 in the CV volume test. We passed the test succesfully locally.

add some comments and remove unused variables from vol.cpp

392650c

merge schedule annealing into v_polytopes branch. Replace stdMatrix w…

2e75945

…ith eigen matrix in Vpolytope class. Use template number type in VPolytope class.

Add two test: 1) volesti volume test with float and 2) CV volume test…

d4c304a

… with long double.

Add comments in polytopes.h and in solve_lp.h. We add a check at the …

839ecc9

…end of solve_LP() function that computes the chebychev center of a H-polytope. We add some comments in samplers.h and use NT number type where it is not able to use a tamplate.

add to R interface the choice of giving as input a V-polytope. You ha…

8d592d5

…ve to give the path to the ext file and a flag Vpoly=TRUE.

delete excecutables and CMAKEFILES from the test folder.

8a36c44

TolisChal added 13 commits July 14, 2019 20:10

fix bug for rk4 solver for hmc-barrirer sampling.

1f6e909

add scaling of the polytope for the hmc-barrier.

522a1dc

improve rk4 method performance

5ed1e03

improve code performance in rk4

484c1e5

improve code in rk4 by implementing rk4 method iteratively

7ad9e25

implment rk2 to compute the hmc-barrier trajectory

90a1aeb

add hmc gaussian with reflections

cd0b449

complete the implementation of HMC gaussian with reflections

d21cfd3

fix bugs in hmc gaussian with reflections

e742cee

add hmc_reflections implementation in gaussian_samplers.h

7ab2e79

use HMC gaussian with reflection for volume computation

99ed971

scale polytope in volume cg algo so that radius of cheb ball equals to 1

0797d19

fix c++ examples bugs

60f3deb

vissarion self-requested a review September 11, 2019 13:51

vissarion requested changes Sep 11, 2019

View reviewed changes

vissarion added the enhancement label Feb 5, 2020

vissarion force-pushed the develop branch 11 times, most recently from e4ac563 to ca784f5 Compare February 27, 2020 15:07

vissarion force-pushed the develop branch from 835f601 to 1bd54f2 Compare October 26, 2020 10:48

vissarion closed this Sep 21, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

HMC random walks #38

HMC random walks #38

TolisChal commented Aug 20, 2019

vissarion left a comment

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion Sep 11, 2019

vissarion commented Sep 21, 2021

@@ @@ -327,4 +346,47 @@ void gaussian_ball_walk(Point &p, @@
                   }
               }
+              /*

		@@ -0,0 +1,112 @@
		// VolEsti (volume computation and sampling library)

		@@ -0,0 +1,164 @@
		// VolEsti (volume computation and sampling library)


		// Licensed under GNU LGPL.3, see LICENCE file

		#ifndef HMC_REFL_H

		@@ -0,0 +1,167 @@
		// VolEsti (volume computation and sampling library)

HMC random walks #38

HMC random walks #38

Conversation

TolisChal commented Aug 20, 2019

vissarion left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vissarion commented Sep 21, 2021