Blame view
src/transform/fmllr-diag-gmm-test.cc
8.58 KB
8dcb6dfcb
first commit
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 |
// transform/fmllr-diag-gmm-test.cc
// Copyright 2009-2011 Microsoft Corporation
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include "util/common-utils.h"
#include "gmm/diag-gmm.h"
#include "transform/fmllr-diag-gmm.h"
namespace kaldi {
void InitRandomGmm (DiagGmm *gmm_in) {
int32 num_gauss = 5 + rand () % 4;
int32 dim = 10 + Rand() % 10;
DiagGmm &gmm(*gmm_in);
gmm.Resize(num_gauss, dim);
Matrix<BaseFloat> inv_vars(num_gauss, dim),
means(num_gauss, dim);
Vector<BaseFloat> weights(num_gauss);
for (int32 i = 0; i < num_gauss; i++) {
for (int32 j = 0; j < dim; j++) {
inv_vars(i, j) = Exp(RandGauss() * (1.0 / (1 + j)));
means(i, j) = RandGauss() * (1.0 / (1 + j));
}
weights(i) = Exp(RandGauss());
}
weights.Scale(1.0 / weights.Sum());
gmm.SetWeights(weights);
gmm.SetInvVarsAndMeans(inv_vars, means);
gmm.ComputeGconsts();
}
// This test is statistical and relies on some identities
// related to the Aikake criterion.
void UnitTestFmllrDiagGmm() {
using namespace kaldi;
DiagGmm gmm;
InitRandomGmm(&gmm);
int32 dim = gmm.Dim();
int32 npoints = dim*(dim+1)*5;
Matrix<BaseFloat> rand_points(npoints, dim);
for (int32 i = 0; i < npoints; i++) {
SubVector<BaseFloat> row(rand_points, i);
gmm.Generate(&row);
}
Matrix<BaseFloat> cur_xform(dim, dim+1);
cur_xform.SetUnit(); // set diag to unit.
int32 niters = 5;
BaseFloat objf_change_tot = 0.0, objf_change, count;
for (int32 j = 0; j < niters; j++) {
FmllrOptions opts;
FmllrDiagGmmAccs stats(dim, j % 2 == 0 ? opts : FmllrOptions());
for (int32 i = 0; i < npoints; i++) {
SubVector<BaseFloat> row(rand_points, i);
if (j == 0) { // split this case off to exercise more of the code.
stats.AccumulateForGmm(gmm, row, 1.0);
} else {
Vector<BaseFloat> xformed_row(row);
ApplyAffineTransform(cur_xform, &xformed_row);
Vector<BaseFloat> posteriors(gmm.NumGauss());
gmm.ComponentPosteriors(xformed_row, &posteriors);
stats.AccumulateFromPosteriors(gmm, row, posteriors);
}
}
stats.Update(opts, &cur_xform, &objf_change, &count);
{ // Test for ApplyFeatureTransformToStats:
BaseFloat objf_change_tmp, count_tmp;
ApplyFeatureTransformToStats(cur_xform, &stats);
Matrix<BaseFloat> mat(dim, dim+1);
mat.SetUnit();
stats.Update(opts, &mat, &objf_change_tmp, &count_tmp);
// After we apply this transform to the stats, there should
// be nothing to gain from further transforming the data.
KALDI_ASSERT(objf_change_tmp/count_tmp < 0.01);
}
KALDI_LOG << "Objf change on iter " << j << " is " << objf_change;
objf_change_tot += objf_change;
}
KALDI_ASSERT(ApproxEqual(count, npoints));
int32 num_params = dim*(dim+1);
BaseFloat expected_objf_change = 0.5 * num_params;
KALDI_LOG << "Expected objf change is: not much more than " << expected_objf_change
<<", seen: " << objf_change_tot;
KALDI_ASSERT(objf_change_tot < 2.0 * expected_objf_change); // or way too much.
// This test relies on statistical laws and if it fails it does not *necessarily*
// mean that something is wrong.
}
// This is a test for the diagonal update and also of ApplyModelTransformToStats().
void UnitTestFmllrDiagGmmDiagonal() {
using namespace kaldi;
DiagGmm gmm;
InitRandomGmm(&gmm);
int32 dim = gmm.Dim();
int32 npoints = dim*(dim+1)*5;
Matrix<BaseFloat> rand_points(npoints, dim);
for (int32 i = 0; i < npoints; i++) {
SubVector<BaseFloat> row(rand_points, i);
gmm.Generate(&row);
}
Matrix<BaseFloat> cur_xform(dim, dim+1);
cur_xform.SetUnit(); // set diag to unit.
int32 niters = 2;
BaseFloat objf_change_tot = 0.0, objf_change, count;
FmllrOptions opts;
opts.update_type = "diag";
for (int32 j = 0; j < niters; j++) {
FmllrDiagGmmAccs stats(dim, j % 2 == 0 ? opts : FmllrOptions());
for (int32 i = 0; i < npoints; i++) {
SubVector<BaseFloat> row(rand_points, i);
if (j == 0) { // split this case off to exercise more of the code.
stats.AccumulateForGmm(gmm, row, 1.0);
} else {
Vector<BaseFloat> xformed_row(row);
ApplyAffineTransform(cur_xform, &xformed_row);
Vector<BaseFloat> posteriors(gmm.NumGauss());
gmm.ComponentPosteriors(xformed_row, &posteriors);
stats.AccumulateFromPosteriors(gmm, row, posteriors);
}
}
stats.Update(opts, &cur_xform, &objf_change, &count);
{ // Test for ApplyModelTransformToStats:
BaseFloat objf_change_tmp, count_tmp;
ApplyModelTransformToStats(cur_xform, &stats);
Matrix<BaseFloat> mat(dim, dim+1);
mat.SetUnit();
stats.Update(opts, &mat, &objf_change_tmp, &count_tmp);
// After we apply this transform to the stats, there should
// be nothing to gain from further transforming the data.
KALDI_ASSERT(objf_change_tmp/count_tmp < 0.01);
}
KALDI_LOG << "Objf change on iter " << j << " is " << objf_change;
objf_change_tot += objf_change;
}
KALDI_ASSERT(ApproxEqual(count, npoints));
int32 num_params = dim*2;
BaseFloat expected_objf_change = 0.5 * num_params;
KALDI_LOG << "Expected objf change is: not much more than " << expected_objf_change
<<", seen: " << objf_change_tot;
KALDI_ASSERT(objf_change_tot < 2.0 * expected_objf_change); // or way too much.
// This test relies on statistical laws and if it fails it does not *necessarily*
// mean that something is wrong.
}
// This is a test for the offset-only update and also of ApplyModelTransformToStats().
void UnitTestFmllrDiagGmmOffset() {
using namespace kaldi;
DiagGmm gmm;
InitRandomGmm(&gmm);
int32 dim = gmm.Dim();
int32 npoints = dim*(dim+1)*5;
Matrix<BaseFloat> rand_points(npoints, dim);
for (int32 i = 0; i < npoints; i++) {
SubVector<BaseFloat> row(rand_points, i);
gmm.Generate(&row);
}
Matrix<BaseFloat> cur_xform(dim, dim+1);
cur_xform.SetUnit(); // set diag to unit.
int32 niters = 2;
BaseFloat objf_change_tot = 0.0, objf_change, count;
FmllrOptions opts;
opts.update_type = "offset";
for (int32 j = 0; j < niters; j++) {
FmllrDiagGmmAccs stats(dim, j % 2 == 0 ? opts : FmllrOptions());
for (int32 i = 0; i < npoints; i++) {
SubVector<BaseFloat> row(rand_points, i);
if (j == 0) { // split this case off to exercise more of the code.
stats.AccumulateForGmm(gmm, row, 1.0);
} else {
Vector<BaseFloat> xformed_row(row);
ApplyAffineTransform(cur_xform, &xformed_row);
Vector<BaseFloat> posteriors(gmm.NumGauss());
gmm.ComponentPosteriors(xformed_row, &posteriors);
stats.AccumulateFromPosteriors(gmm, row, posteriors);
}
}
stats.Update(opts, &cur_xform, &objf_change, &count);
{ // Test for ApplyModelTransformToStats:
BaseFloat objf_change_tmp, count_tmp;
ApplyModelTransformToStats(cur_xform, &stats);
Matrix<BaseFloat> mat(dim, dim+1);
mat.SetUnit();
stats.Update(opts, &mat, &objf_change_tmp, &count_tmp);
// After we apply this transform to the stats, there should
// be nothing to gain from further transforming the data.
KALDI_ASSERT(objf_change_tmp/count_tmp < 0.01);
}
KALDI_LOG << "Objf change on iter " << j << " is " << objf_change;
objf_change_tot += objf_change;
}
KALDI_ASSERT(ApproxEqual(count, npoints));
int32 num_params = dim;
BaseFloat expected_objf_change = 0.5 * num_params;
KALDI_LOG << "Expected objf change is: not much more than " << expected_objf_change
<<", seen: " << objf_change_tot;
KALDI_ASSERT(objf_change_tot < 2.0 * expected_objf_change); // or way too much.
// This test relies on statistical laws and if it fails it does not *necessarily*
// mean that something is wrong.
}
} // namespace kaldi ends here
int main() {
for (int i = 0; i < 2; i++) { // did more iterations when first testing...
kaldi::UnitTestFmllrDiagGmmOffset();
kaldi::UnitTestFmllrDiagGmmDiagonal();
kaldi::UnitTestFmllrDiagGmm();
}
std::cout << "Test OK.
";
}
|