lsst.jointcal  21.0.0-20-g3b2d1f0+f7c7c85b61
FitterBase.cc
Go to the documentation of this file.
1 // -*- LSST-C++ -*-
2 /*
3  * This file is part of jointcal.
4  *
5  * Developed for the LSST Data Management System.
6  * This product includes software developed by the LSST Project
7  * (https://www.lsst.org).
8  * See the COPYRIGHT file at the top-level directory of this distribution
9  * for details of code ownership.
10  *
11  * This program is free software: you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation, either version 3 of the License, or
14  * (at your option) any later version.
15  *
16  * This program is distributed in the hope that it will be useful,
17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19  * GNU General Public License for more details.
20  *
21  * You should have received a copy of the GNU General Public License
22  * along with this program. If not, see <https://www.gnu.org/licenses/>.
23  */
24 
25 #include <vector>
26 #include "Eigen/Core"
27 
28 #include <boost/math/tools/minima.hpp>
29 
30 #include "lsst/log/Log.h"
31 
32 #include "lsst/jointcal/Chi2.h"
33 #include "lsst/jointcal/CcdImage.h"
38 
39 namespace lsst {
40 namespace jointcal {
41 
43  Chi2Statistic chi2;
44  accumulateStatImageList(_associations->getCcdImageList(), chi2);
46  // chi2.ndof contains the number of squares.
47  // So subtract the number of parameters.
48  chi2.ndof -= _nTotal;
49  return chi2;
50 }
51 
53  FittedStarList &fsOutliers, double &cut) const {
54  // collect chi2 contributions
55  Chi2List chi2List;
56  chi2List.reserve(_associations->getMaxMeasuredStars() + _associations->refStarList.size());
57  // contributions from measurement terms:
58  accumulateStatImageList(_associations->ccdImageList, chi2List);
59  // and from reference terms
60  accumulateStatRefStars(chi2List);
61 
62  // compute some statistics
63  size_t nval = chi2List.size();
64  if (nval == 0) return 0;
65  sort(chi2List.begin(), chi2List.end());
66  double median = (nval & 1) ? chi2List[nval / 2].chi2
67  : 0.5 * (chi2List[nval / 2 - 1].chi2 + chi2List[nval / 2].chi2);
68  auto averageAndSigma = chi2List.computeAverageAndSigma();
69  LOGLS_DEBUG(_log, "findOutliers chi2 stat: mean/median/sigma " << averageAndSigma.first << '/' << median
70  << '/' << averageAndSigma.second);
71  cut = averageAndSigma.first + nSigmaCut * averageAndSigma.second;
72  /* For each of the parameters, we will not remove more than 1
73  measurement that contributes to constraining it. Keep track using
74  of what we are touching using an integer vector. This is the
75  trick that Marc Betoule came up to for outlier removals in "star
76  flats" fits. */
77  Eigen::VectorXi affectedParams(_nTotal);
78  affectedParams.setZero();
79 
80  std::size_t nOutliers = 0; // returned to the caller
81  // start from the strongest outliers.
82  for (auto chi2 = chi2List.rbegin(); chi2 != chi2List.rend(); ++chi2) {
83  if (chi2->chi2 < cut) break; // because the array is sorted.
84  IndexVector indices;
85  /* now, we want to get the indices of the parameters this chi2
86  term depends on. We have to figure out which kind of term it
87  is; we use for that the type of the star attached to the Chi2Star. */
88  auto measuredStar = std::dynamic_pointer_cast<MeasuredStar>(chi2->star);
89  std::shared_ptr<FittedStar> fittedStar; // To add to fsOutliers if it is a reference outlier.
90  if (measuredStar == nullptr) {
91  // it is a reference outlier
92  fittedStar = std::dynamic_pointer_cast<FittedStar>(chi2->star);
93  if (fittedStar->getMeasurementCount() == 0) {
94  LOGLS_WARN(_log, "FittedStar with no measuredStars found as an outlier: "
95  << *fittedStar << " chi2: " << chi2->chi2);
96  continue;
97  }
98  if (_nStarParams == 0) {
100  "RefStar is outlier but not removed when not fitting FittedStar-RefStar values: "
101  << *(fittedStar->getRefStar()) << " chi2: " << chi2->chi2);
102  continue;
103  }
104  // NOTE: Stars contribute twice to astrometry (x,y), but once to photometry (flux),
105  // NOTE: but we only need to mark one index here because both will be removed with that star.
106  indices.push_back(fittedStar->getIndexInMatrix());
107  LOGLS_TRACE(_log, "Removing refStar " << *(fittedStar->getRefStar()) << " chi2: " << chi2->chi2);
108  /* One might think it would be useful to account for PM
109  parameters here, but it is just useless */
110  } else {
111  // it is a measurement outlier
112  auto tempFittedStar = measuredStar->getFittedStar();
113  if (tempFittedStar->getMeasurementCount() == 1 && tempFittedStar->getRefStar() == nullptr) {
114  LOGLS_WARN(_log, "FittedStar with 1 measuredStar and no refStar found as an outlier: "
115  << *tempFittedStar);
116  continue;
117  }
118  getIndicesOfMeasuredStar(*measuredStar, indices);
119  LOGLS_TRACE(_log, "Removing measStar " << *measuredStar << " chi2: " << chi2->chi2);
120  }
121 
122  /* Find out if we already discarded a stronger outlier
123  constraining some parameter this one constrains as well. If
124  yes, we keep this one, because this stronger outlier could be
125  causing the large chi2 we have in hand. */
126  bool drop_it = true;
127  for (auto const &i : indices) {
128  if (affectedParams(i) != 0) {
129  drop_it = false;
130  }
131  }
132 
133  if (drop_it) // store the outlier in one of the lists:
134  {
135  if (measuredStar == nullptr) {
136  // reference term
137  fsOutliers.push_back(fittedStar);
138  } else {
139  // measurement term
140  msOutliers.push_back(measuredStar);
141  }
142  // mark the parameters as directly changed when we discard this chi2 term.
143  for (auto const &i : indices) {
144  affectedParams(i)++;
145  }
146  nOutliers++;
147  }
148  } // end loop on measurements/references
149  LOGLS_INFO(_log, "findOutliers: found " << msOutliers.size() << " meas outliers and " << fsOutliers.size()
150  << " ref outliers ");
151 
152  return nOutliers;
153 }
154 
155 namespace {
157 SparseMatrixD createHessian(std::size_t nParTot, TripletList const &tripletList) {
158  SparseMatrixD jacobian(nParTot, tripletList.getNextFreeIndex());
159  jacobian.setFromTriplets(tripletList.begin(), tripletList.end());
160  return jacobian * jacobian.transpose();
161 }
162 
164 void dumpMatrixAndGradient(SparseMatrixD const &matrix, Eigen::VectorXd const &grad,
165  std::string const &dumpFile, LOG_LOGGER _log) {
166  std::string ext = ".txt";
167  Eigen::MatrixXd matrixDense(matrix);
168  std::string dumpMatrixPath = dumpFile + "-mat" + ext;
169  std::ofstream matfile(dumpMatrixPath);
170  matfile << matrixDense << std::endl;
171  std::string dumpGradPath = dumpFile + "-grad" + ext;
172  std::ofstream gradfile(dumpGradPath);
173  gradfile << grad << std::endl;
174  LOGLS_INFO(_log, "Dumped Hessian, gradient to: '" << dumpMatrixPath << "', '" << dumpGradPath << "'");
175 }
176 } // namespace
177 
178 MinimizeResult FitterBase::minimize(std::string const &whatToFit, double nSigmaCut,
179  double sigmaRelativeTolerance, bool doRankUpdate, bool const doLineSearch,
180  std::string const &dumpMatrixFile) {
181  assignIndices(whatToFit);
182 
184 
185  // TODO : write a guesser for the number of triplets
186  std::size_t nTrip = (_lastNTrip) ? _lastNTrip : 1e6;
187  TripletList tripletList(nTrip);
188  Eigen::VectorXd grad(_nTotal);
189  grad.setZero();
190  double scale = 1.0;
191 
192  // Fill the triplets
193  leastSquareDerivatives(tripletList, grad);
194  _lastNTrip = tripletList.size();
195 
196  LOGLS_DEBUG(_log, "End of triplet filling, ntrip = " << tripletList.size());
197 
198  SparseMatrixD hessian = createHessian(_nTotal, tripletList);
199  tripletList.clear(); // we don't need it any more after we have the hessian.
200 
201  LOGLS_DEBUG(_log, "Starting factorization, hessian: dim="
202  << hessian.rows() << " non-zeros=" << hessian.nonZeros()
203  << " filling-frac = " << hessian.nonZeros() / std::pow(hessian.rows(), 2));
204 
205  if (dumpMatrixFile != "") {
206  if (hessian.rows() * hessian.cols() > 2e8) {
207  LOGLS_WARN(_log, "Hessian matrix is too big to dump to file, with rows, columns: "
208  << hessian.rows() << ", " << hessian.cols());
209  } else {
210  dumpMatrixAndGradient(hessian, grad, dumpMatrixFile, _log);
211  }
212  }
213 
215  if (chol.info() != Eigen::Success) {
216  LOGLS_ERROR(_log, "minimize: factorization failed ");
217  return MinimizeResult::Failed;
218  }
219 
220  std::size_t totalMeasOutliers = 0;
221  std::size_t totalRefOutliers = 0;
222  double oldChi2 = computeChi2().chi2;
223  double oldSigmaCut = 0;
224  double sigmaCut;
225 
226  while (true) {
227  Eigen::VectorXd delta = chol.solve(grad);
228  if (doLineSearch) {
229  scale = _lineSearch(delta);
230  }
231  offsetParams(scale * delta);
232  Chi2Statistic currentChi2(computeChi2());
233  LOGLS_DEBUG(_log, currentChi2);
234  if (!isfinite(currentChi2.chi2)) {
235  LOGL_ERROR(_log, "chi2 is not finite. Aborting outlier rejection.");
236  returnCode = MinimizeResult::NonFinite;
237  break;
238  }
239  if (currentChi2.chi2 > oldChi2 && totalMeasOutliers + totalRefOutliers != 0) {
240  LOGL_WARN(_log, "chi2 went up, skipping outlier rejection loop");
241  returnCode = MinimizeResult::Chi2Increased;
242  break;
243  }
244  oldChi2 = currentChi2.chi2;
245 
246  if (nSigmaCut == 0) break; // no rejection step to perform
247  MeasuredStarList msOutliers;
248  FittedStarList fsOutliers;
249  // keep nOutliers so we don't have to sum msOutliers.size()+fsOutliers.size() twice below.
250  std::size_t nOutliers = findOutliers(nSigmaCut, msOutliers, fsOutliers, sigmaCut);
251  double relChange = 1 - sigmaCut / oldSigmaCut;
252  LOGLS_DEBUG(_log, "findOutliers chi2 cut level: " << sigmaCut << ", relative change: " << relChange);
253  // If sigmaRelativeTolerance is set and at least one iteration has been done, break loop when the
254  // fractional change in sigmaCut levels is less than the sigmaRelativeTolerance parameter.
255  if ((sigmaRelativeTolerance > 0) && (oldSigmaCut > 0 && relChange < sigmaRelativeTolerance)){
256  LOGLS_INFO(_log, "Iterations stopped with chi2 cut at " << sigmaCut << " and relative change of "
257  << relChange);
258  break;
259  }
260  totalMeasOutliers += msOutliers.size();
261  totalRefOutliers += fsOutliers.size();
262  oldSigmaCut = sigmaCut;
263  if (nOutliers == 0) break;
264  TripletList outlierTriplets(nOutliers);
265  grad.setZero(); // recycle the gradient
266  // compute the contributions of outliers to derivatives
267  outliersContributions(msOutliers, fsOutliers, outlierTriplets, grad);
268  // Remove significant outliers
269  removeMeasOutliers(msOutliers);
270  removeRefOutliers(fsOutliers);
271  if (doRankUpdate) {
272  // convert triplet list to eigen internal format
273  SparseMatrixD H(_nTotal, outlierTriplets.getNextFreeIndex());
274  H.setFromTriplets(outlierTriplets.begin(), outlierTriplets.end());
275  chol.update(H, false /* means downdate */);
276  // The contribution of outliers to the gradient is the opposite
277  // of the contribution of all other terms, because they add up to 0
278  grad *= -1;
279  } else {
280  // don't reuse tripletList because we want a new nextFreeIndex.
281  TripletList nextTripletList(_lastNTrip);
282  grad.setZero();
283  // Rebuild the matrix and gradient
284  leastSquareDerivatives(nextTripletList, grad);
285  _lastNTrip = nextTripletList.size();
286  LOGLS_DEBUG(_log, "Triplets recomputed, ntrip = " << nextTripletList.size());
287 
288  hessian = createHessian(_nTotal, nextTripletList);
289  nextTripletList.clear(); // we don't need it any more after we have the hessian.
290 
292  "Restarting factorization, hessian: dim="
293  << hessian.rows() << " non-zeros=" << hessian.nonZeros()
294  << " filling-frac = " << hessian.nonZeros() / std::pow(hessian.rows(), 2));
295  chol.compute(hessian);
296  if (chol.info() != Eigen::Success) {
297  LOGLS_ERROR(_log, "minimize: factorization failed ");
298  return MinimizeResult::Failed;
299  }
300  }
301  }
302 
303  if (totalMeasOutliers + totalRefOutliers > 0) {
304  _associations->cleanFittedStars();
305  }
306 
307  // only print the outlier summary if outlier rejection was turned on.
308  if (nSigmaCut != 0) {
309  LOGLS_INFO(_log, "Number of outliers (Measured + Reference = Total): "
310  << totalMeasOutliers << " + " << totalRefOutliers << " = "
311  << totalMeasOutliers + totalRefOutliers);
312  }
313  return returnCode;
314 }
315 
317  TripletList &tripletList, Eigen::VectorXd &grad) {
318  for (auto &outlier : msOutliers) {
319  MeasuredStarList tmp;
320  tmp.push_back(outlier);
321  const CcdImage &ccdImage = outlier->getCcdImage();
322  leastSquareDerivativesMeasurement(ccdImage, tripletList, grad, &tmp);
323  }
324  leastSquareDerivativesReference(fsOutliers, tripletList, grad);
325 }
326 
328  for (auto &measuredStar : outliers) {
329  auto fittedStar = measuredStar->getFittedStar();
330  measuredStar->setValid(false);
331  fittedStar->getMeasurementCount()--; // could be put in setValid
332  }
333 }
334 
336  for (auto &fittedStar : outliers) {
337  fittedStar->setRefStar(nullptr);
338  }
339 }
340 
341 void FitterBase::leastSquareDerivatives(TripletList &tripletList, Eigen::VectorXd &grad) const {
342  auto ccdImageList = _associations->getCcdImageList();
343  for (auto const &ccdImage : ccdImageList) {
344  leastSquareDerivativesMeasurement(*ccdImage, tripletList, grad);
345  }
346  leastSquareDerivativesReference(_associations->fittedStarList, tripletList, grad);
347 }
348 
349 void FitterBase::saveChi2Contributions(std::string const &baseName) const {
350  std::string replaceStr = "{type}";
351  auto pos = baseName.find(replaceStr);
352  std::string measFilename(baseName);
353  measFilename.replace(pos, replaceStr.size(), "-meas.csv");
354  std::string refFilename(baseName);
355  refFilename.replace(pos, replaceStr.size(), "-ref.csv");
356  saveChi2MeasContributions(measFilename);
357  saveChi2RefContributions(refFilename);
358 }
359 
360 double FitterBase::_lineSearch(Eigen::VectorXd const &delta) {
361  auto func = [this, &delta](double scale) {
362  auto offset = scale * delta;
363  offsetParams(offset);
364  auto chi2 = computeChi2();
365  // reset the system to where it was before offsetting.
366  offsetParams(-offset);
367  return chi2.chi2;
368  };
369  // The maximum theoretical precision is half the number of bits in the mantissa (see boost docs).
370  auto bits = std::numeric_limits<double>::digits / 2;
371  auto result = boost::math::tools::brent_find_minima(func, -1.0, 2.0, bits);
372  LOGLS_DEBUG(_log, "Line search scale factor: " << result.first);
373  return result.first;
374 }
375 
376 } // namespace jointcal
377 } // namespace lsst
Eigen::SparseMatrix< double, 0, Eigen::Index > SparseMatrixD
Definition: Eigenstuff.h:35
#define LOGLS_WARN(logger, message)
#define LOGL_WARN(logger, message...)
#define LOGLS_INFO(logger, message)
#define LOGLS_ERROR(logger, message)
#define LOGL_ERROR(logger, message...)
#define LOGLS_DEBUG(logger, message)
#define LOGLS_TRACE(logger, message)
T begin(T... args)
void update(SparseMatrixD const &H, bool UpOrDown)
Definition: Eigenstuff.h:68
Handler of an actual image from a single CCD.
Definition: CcdImage.h:64
Structure to accumulate the chi2 contributions per each star (to help find outliers).
Definition: Chi2.h:100
std::pair< double, double > computeAverageAndSigma()
Compute the average and std-deviation of these chisq values.
Definition: Chi2.cc:33
Simple structure to accumulate chi2 and ndof.
Definition: Chi2.h:52
A list of FittedStar s. Such a list is typically constructed by Associations.
Definition: FittedStar.h:123
void leastSquareDerivatives(TripletList &tripletList, Eigen::VectorXd &grad) const
Evaluates the chI^2 derivatives (Jacobian and gradient) for the current whatToFit setting.
Definition: FitterBase.cc:341
void removeRefOutliers(FittedStarList &outliers)
Remove refStar outliers from the fit. No Refit done.
Definition: FitterBase.cc:335
virtual void getIndicesOfMeasuredStar(MeasuredStar const &measuredStar, IndexVector &indices) const =0
Set the indices of a measured star from the full matrix, for outlier removal.
Chi2Statistic computeChi2() const
Returns the chi2 for the current state.
Definition: FitterBase.cc:42
virtual void saveChi2MeasContributions(std::string const &filename) const =0
Save a CSV file containing residuals of measurement terms.
virtual void leastSquareDerivativesReference(FittedStarList const &fittedStarList, TripletList &tripletList, Eigen::VectorXd &grad) const =0
Compute the derivatives of the reference terms.
virtual void saveChi2Contributions(std::string const &baseName) const
Save the full chi2 term per star that was used in the minimization, for debugging.
Definition: FitterBase.cc:349
virtual void assignIndices(std::string const &whatToFit)=0
Set parameters to fit and assign indices in the big matrix.
virtual void offsetParams(Eigen::VectorXd const &delta)=0
Offset the parameters by the requested quantities.
std::shared_ptr< Associations > _associations
Definition: FitterBase.h:161
virtual void accumulateStatRefStars(Chi2Accumulator &accum) const =0
Compute the chi2 (per star or total, depending on which Chi2Accumulator is used) for RefStars.
virtual void saveChi2RefContributions(std::string const &filename) const =0
Save a CSV file containing residuals of reference terms.
virtual void leastSquareDerivativesMeasurement(CcdImage const &ccdImage, TripletList &tripletList, Eigen::VectorXd &grad, MeasuredStarList const *measuredStarList=nullptr) const =0
Compute the derivatives of the measured stars and model for one CcdImage.
void outliersContributions(MeasuredStarList &msOutliers, FittedStarList &fsOutliers, TripletList &tripletList, Eigen::VectorXd &grad)
Contributions to derivatives from (presumably) outlier terms.
Definition: FitterBase.cc:316
std::size_t findOutliers(double nSigmaCut, MeasuredStarList &msOutliers, FittedStarList &fsOutliers, double &cut) const
Find Measurements and references contributing more than a cut, computed as.
Definition: FitterBase.cc:52
virtual void accumulateStatImageList(CcdImageList const &ccdImageList, Chi2Accumulator &accum) const =0
Compute the chi2 (per star or total, depending on which Chi2Accumulator is used) for measurements.
void removeMeasOutliers(MeasuredStarList &outliers)
Remove measuredStar outliers from the fit. No Refit done.
Definition: FitterBase.cc:327
MinimizeResult minimize(std::string const &whatToFit, double const nSigmaCut=0, double sigmaRelativeTolerance=0, bool const doRankUpdate=true, bool const doLineSearch=false, std::string const &dumpMatrixFile="")
Does a 1 step minimization, assuming a linear model.
Definition: FitterBase.cc:178
A list of MeasuredStar. They are usually filled in Associations::createCcdImage.
Definition: MeasuredStar.h:146
Eigen::Index getNextFreeIndex() const
Definition: Tripletlist.h:47
T clear(T... args)
T end(T... args)
T endl(T... args)
T find(T... args)
T isfinite(T... args)
def scale(algorithm, min, max=None, frame=None)
MinimizeResult
Return value of minimize()
Definition: FitterBase.h:40
Class for a simple mapping implementing a generic AstrometryTransform.
T pow(T... args)
T push_back(T... args)
T rbegin(T... args)
T rend(T... args)
T replace(T... args)
T reserve(T... args)
T size(T... args)
T sort(T... args)