From 45e0e5f8e947a9c1b4a995477c43a006ec2df43f Mon Sep 17 00:00:00 2001 From: Pierre-Emmanuel Viel Date: Tue, 17 Dec 2013 12:51:58 +0100 Subject: [PATCH 1/5] Pick centers in KMeans++ with a probability proportional to their distance^2, instead of simple distance, to previous centers --- .../opencv2/flann/hierarchical_clustering_index.h | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h b/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h index ce2d622450..02fc278448 100644 --- a/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h +++ b/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h @@ -210,8 +210,11 @@ private: assert(index >=0 && index < n); centers[0] = dsindices[index]; + // Computing distance^2 will have the advantage of even higher probability further to pick new centers + // far from previous centers (and this complies to "k-means++: the advantages of careful seeding" article) for (int i = 0; i < n; i++) { closestDistSq[i] = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols); + closestDistSq[i] *= closestDistSq[i]; currentPot += closestDistSq[i]; } @@ -237,7 +240,10 @@ private: // Compute the new potential double newPot = 0; - for (int i = 0; i < n; i++) newPot += std::min( distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols), closestDistSq[i] ); + for (int i = 0; i < n; i++) { + DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols); + newPot += std::min( dist*dist, closestDistSq[i] ); + } // Store the best result if ((bestNewPot < 0)||(newPot < bestNewPot)) { @@ -249,7 +255,10 @@ private: // Add the appropriate center centers[centerCount] = dsindices[bestNewIndex]; currentPot = bestNewPot; - for (int i = 0; i < n; i++) closestDistSq[i] = std::min( distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols), closestDistSq[i] ); + for (int i = 0; i < n; i++) { + DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols); + closestDistSq[i] = std::min( dist*dist, closestDistSq[i] ); + } } centers_length = centerCount; From 5aeeaa6fce4016fd626f31f56025cf83ff07576a Mon Sep 17 00:00:00 2001 From: Pierre-Emmanuel Viel Date: Tue, 17 Dec 2013 13:04:49 +0100 Subject: [PATCH 2/5] Apply to KMeansIndex KMeanspp the same modification as in HierarchicalClusteringIndex --- modules/flann/include/opencv2/flann/kmeans_index.h | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/modules/flann/include/opencv2/flann/kmeans_index.h b/modules/flann/include/opencv2/flann/kmeans_index.h index 3fea956a74..3bf12047cd 100644 --- a/modules/flann/include/opencv2/flann/kmeans_index.h +++ b/modules/flann/include/opencv2/flann/kmeans_index.h @@ -211,6 +211,7 @@ public: for (int i = 0; i < n; i++) { closestDistSq[i] = distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols); + closestDistSq[i] *= closestDistSq[i]; currentPot += closestDistSq[i]; } @@ -236,7 +237,10 @@ public: // Compute the new potential double newPot = 0; - for (int i = 0; i < n; i++) newPot += std::min( distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols), closestDistSq[i] ); + for (int i = 0; i < n; i++) { + DistanceType dist = distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols); + newPot += std::min( dist*dist, closestDistSq[i] ); + } // Store the best result if ((bestNewPot < 0)||(newPot < bestNewPot)) { @@ -248,7 +252,10 @@ public: // Add the appropriate center centers[centerCount] = indices[bestNewIndex]; currentPot = bestNewPot; - for (int i = 0; i < n; i++) closestDistSq[i] = std::min( distance_(dataset_[indices[i]], dataset_[indices[bestNewIndex]], dataset_.cols), closestDistSq[i] ); + for (int i = 0; i < n; i++) { + DistanceType dist = distance_(dataset_[indices[i]], dataset_[indices[bestNewIndex]], dataset_.cols); + closestDistSq[i] = std::min( dist*dist, closestDistSq[i] ); + } } centers_length = centerCount; From fa749de0dcb27d3b666eda341b43e8c13f66be8e Mon Sep 17 00:00:00 2001 From: Pierre-Emmanuel Viel Date: Tue, 17 Dec 2013 13:26:55 +0100 Subject: [PATCH 3/5] As some processed distances are already ^2, use template to select whether or not we have to ^2 in KMeanspp --- .../include/opencv2/flann/kmeans_index.h | 62 ++++++++++++++++++- 1 file changed, 59 insertions(+), 3 deletions(-) diff --git a/modules/flann/include/opencv2/flann/kmeans_index.h b/modules/flann/include/opencv2/flann/kmeans_index.h index 3bf12047cd..460dc64be9 100644 --- a/modules/flann/include/opencv2/flann/kmeans_index.h +++ b/modules/flann/include/opencv2/flann/kmeans_index.h @@ -53,6 +53,62 @@ namespace cvflann { +template +struct squareDistance +{ + typedef typename Distance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist*dist; } +}; + + +template +struct squareDistance, ElementType> +{ + typedef typename L2_Simple::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + +template +struct squareDistance, ElementType> +{ + typedef typename L2::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + + +template +struct squareDistance, ElementType> +{ + typedef typename MinkowskiDistance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + +template +struct squareDistance, ElementType> +{ + typedef typename HellingerDistance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + +template +struct squareDistance, ElementType> +{ + typedef typename ChiSquareDistance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + + +template +typename Distance::ResultType ensureSquareDistance( typename Distance::ResultType dist ) +{ + typedef typename Distance::ElementType ElementType; + + squareDistance dummy; + return dummy( dist ); +} + + + struct KMeansIndexParams : public IndexParams { KMeansIndexParams(int branching = 32, int iterations = 11, @@ -211,7 +267,7 @@ public: for (int i = 0; i < n; i++) { closestDistSq[i] = distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols); - closestDistSq[i] *= closestDistSq[i]; + closestDistSq[i] = ensureSquareDistance( closestDistSq[i] ); currentPot += closestDistSq[i]; } @@ -239,7 +295,7 @@ public: double newPot = 0; for (int i = 0; i < n; i++) { DistanceType dist = distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols); - newPot += std::min( dist*dist, closestDistSq[i] ); + newPot += std::min( ensureSquareDistance(dist), closestDistSq[i] ); } // Store the best result @@ -254,7 +310,7 @@ public: currentPot = bestNewPot; for (int i = 0; i < n; i++) { DistanceType dist = distance_(dataset_[indices[i]], dataset_[indices[bestNewIndex]], dataset_.cols); - closestDistSq[i] = std::min( dist*dist, closestDistSq[i] ); + closestDistSq[i] = std::min( ensureSquareDistance(dist), closestDistSq[i] ); } } From 0d19685f9544ddf2668fa899ce74580fd9d1039f Mon Sep 17 00:00:00 2001 From: Pierre-Emmanuel Viel Date: Tue, 17 Dec 2013 13:34:20 +0100 Subject: [PATCH 4/5] Move templates in dist.h in order to share them between KMeansIndex and HierarchicalClusteringIndex classes. --- modules/flann/include/opencv2/flann/dist.h | 60 +++++++++++++++++++ .../flann/hierarchical_clustering_index.h | 6 +- .../include/opencv2/flann/kmeans_index.h | 56 ----------------- 3 files changed, 63 insertions(+), 59 deletions(-) diff --git a/modules/flann/include/opencv2/flann/dist.h b/modules/flann/include/opencv2/flann/dist.h index 80ae2dc916..2afceb8893 100644 --- a/modules/flann/include/opencv2/flann/dist.h +++ b/modules/flann/include/opencv2/flann/dist.h @@ -812,6 +812,66 @@ struct ZeroIterator }; + +/* + * Depending on processed distances, some of them are already squared (e.g. L2) + * and some are not (e.g.Hamming). In KMeans++ for instance we want to be sure + * we are working on ^2 distances, thus following templates to ensure that. + */ +template +struct squareDistance +{ + typedef typename Distance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist*dist; } +}; + + +template +struct squareDistance, ElementType> +{ + typedef typename L2_Simple::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + +template +struct squareDistance, ElementType> +{ + typedef typename L2::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + + +template +struct squareDistance, ElementType> +{ + typedef typename MinkowskiDistance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + +template +struct squareDistance, ElementType> +{ + typedef typename HellingerDistance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + +template +struct squareDistance, ElementType> +{ + typedef typename ChiSquareDistance::ResultType ResultType; + ResultType operator()( ResultType dist ) { return dist; } +}; + + +template +typename Distance::ResultType ensureSquareDistance( typename Distance::ResultType dist ) +{ + typedef typename Distance::ElementType ElementType; + + squareDistance dummy; + return dummy( dist ); +} + } #endif //OPENCV_FLANN_DIST_H_ diff --git a/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h b/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h index 02fc278448..3ccfa5534b 100644 --- a/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h +++ b/modules/flann/include/opencv2/flann/hierarchical_clustering_index.h @@ -214,7 +214,7 @@ private: // far from previous centers (and this complies to "k-means++: the advantages of careful seeding" article) for (int i = 0; i < n; i++) { closestDistSq[i] = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols); - closestDistSq[i] *= closestDistSq[i]; + closestDistSq[i] = ensureSquareDistance( closestDistSq[i] ); currentPot += closestDistSq[i]; } @@ -242,7 +242,7 @@ private: double newPot = 0; for (int i = 0; i < n; i++) { DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols); - newPot += std::min( dist*dist, closestDistSq[i] ); + newPot += std::min( ensureSquareDistance(dist), closestDistSq[i] ); } // Store the best result @@ -257,7 +257,7 @@ private: currentPot = bestNewPot; for (int i = 0; i < n; i++) { DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols); - closestDistSq[i] = std::min( dist*dist, closestDistSq[i] ); + closestDistSq[i] = std::min( ensureSquareDistance(dist), closestDistSq[i] ); } } diff --git a/modules/flann/include/opencv2/flann/kmeans_index.h b/modules/flann/include/opencv2/flann/kmeans_index.h index 460dc64be9..3cbee24404 100644 --- a/modules/flann/include/opencv2/flann/kmeans_index.h +++ b/modules/flann/include/opencv2/flann/kmeans_index.h @@ -53,62 +53,6 @@ namespace cvflann { -template -struct squareDistance -{ - typedef typename Distance::ResultType ResultType; - ResultType operator()( ResultType dist ) { return dist*dist; } -}; - - -template -struct squareDistance, ElementType> -{ - typedef typename L2_Simple::ResultType ResultType; - ResultType operator()( ResultType dist ) { return dist; } -}; - -template -struct squareDistance, ElementType> -{ - typedef typename L2::ResultType ResultType; - ResultType operator()( ResultType dist ) { return dist; } -}; - - -template -struct squareDistance, ElementType> -{ - typedef typename MinkowskiDistance::ResultType ResultType; - ResultType operator()( ResultType dist ) { return dist; } -}; - -template -struct squareDistance, ElementType> -{ - typedef typename HellingerDistance::ResultType ResultType; - ResultType operator()( ResultType dist ) { return dist; } -}; - -template -struct squareDistance, ElementType> -{ - typedef typename ChiSquareDistance::ResultType ResultType; - ResultType operator()( ResultType dist ) { return dist; } -}; - - -template -typename Distance::ResultType ensureSquareDistance( typename Distance::ResultType dist ) -{ - typedef typename Distance::ElementType ElementType; - - squareDistance dummy; - return dummy( dist ); -} - - - struct KMeansIndexParams : public IndexParams { KMeansIndexParams(int branching = 32, int iterations = 11, From eeed9ce24791bea7fa9c58702d9168255007367a Mon Sep 17 00:00:00 2001 From: Alexander Alekhin Date: Fri, 28 Mar 2014 15:11:10 +0400 Subject: [PATCH 5/5] perf report: summary.py: added custom --regressions map --- modules/ts/misc/summary.py | 186 +++++++++++++++++++++++-------------- 1 file changed, 117 insertions(+), 69 deletions(-) diff --git a/modules/ts/misc/summary.py b/modules/ts/misc/summary.py index 4cbb8901f4..9682699850 100755 --- a/modules/ts/misc/summary.py +++ b/modules/ts/misc/summary.py @@ -40,6 +40,7 @@ if __name__ == "__main__": parser.add_option("", "--with-cycles-reduction", action="store_true", dest="calc_cr", default=False, help="output cycle reduction percentages") parser.add_option("", "--with-score", action="store_true", dest="calc_score", default=False, help="output automatic classification of speedups") parser.add_option("", "--progress", action="store_true", dest="progress_mode", default=False, help="enable progress mode") + parser.add_option("", "--regressions", dest="regressions", default=None, metavar="LIST", help="comma-separated custom regressions map: \"[r][c]#current-#reference\" (indexes of columns are 0-based, \"r\" - reverse flag, \"c\" - color flag for base data)") parser.add_option("", "--show-all", action="store_true", dest="showall", default=False, help="also include empty and \"notrun\" lines") parser.add_option("", "--match", dest="match", default=None) parser.add_option("", "--match-replace", dest="match_replace", default="") @@ -56,6 +57,24 @@ if __name__ == "__main__": if options.columns: options.columns = [s.strip().replace("\\n", "\n") for s in options.columns.split(",")] + if options.regressions: + assert not options.progress_mode, 'unsupported mode' + + def parseRegressionColumn(s): + """ Format: '[r][c]-' """ + reverse = s.startswith('r') + if reverse: + s = s[1:] + addColor = s.startswith('c') + if addColor: + s = s[1:] + parts = s.split('-', 1) + link = (int(parts[0]), int(parts[1]), reverse, addColor) + assert link[0] != link[1] + return link + + options.regressions = [parseRegressionColumn(s) for s in options.regressions.split(',')] + # expand wildcards and filter duplicates files = [] seen = set() @@ -90,8 +109,18 @@ if __name__ == "__main__": sys.stderr.write("Error: no test data found" + os.linesep) quit() - # find matches setsCount = len(test_sets) + + if options.regressions is None: + reference = -1 if options.progress_mode else 0 + options.regressions = [(i, reference, False, True) for i in range(1, len(test_sets))] + + for link in options.regressions: + (i, ref, reverse, addColor) = link + assert i >= 0 and i < setsCount + assert ref < setsCount + + # find matches test_cases = {} name_extractor = lambda name: str(name) @@ -117,29 +146,29 @@ if __name__ == "__main__": # header tbl.newColumn("name", "Name of Test", align = "left", cssclass = "col_name") - i = 0 - for set in test_sets: - tbl.newColumn(str(i), getSetName(set, i, options.columns, False), align = "center") - i += 1 - metric_sets = test_sets[1:] + for i in range(setsCount): + tbl.newColumn(str(i), getSetName(test_sets[i], i, options.columns, False), align = "center") + + def addHeaderColumns(suffix, description, cssclass): + for link in options.regressions: + (i, ref, reverse, addColor) = link + if reverse: + i, ref = ref, i + current_set = test_sets[i] + current = getSetName(current_set, i, options.columns) + if ref >= 0: + reference_set = test_sets[ref] + reference = getSetName(reference_set, ref, options.columns) + else: + reference = 'previous' + tbl.newColumn(str(i) + '-' + str(ref) + suffix, '%s\nvs\n%s\n(%s)' % (current, reference, description), align='center', cssclass=cssclass) + if options.calc_cr: - i = 1 - for set in metric_sets: - reference = getSetName(test_sets[0], 0, options.columns) if not options.progress_mode else 'previous' - tbl.newColumn(str(i) + "$", getSetName(set, i, options.columns) + "\nvs\n" + reference + "\n(cycles reduction)", align = "center", cssclass = "col_cr") - i += 1 + addHeaderColumns(suffix='$', description='cycles reduction', cssclass='col_cr') if options.calc_relatives: - i = 1 - for set in metric_sets: - reference = getSetName(test_sets[0], 0, options.columns) if not options.progress_mode else 'previous' - tbl.newColumn(str(i) + "%", getSetName(set, i, options.columns) + "\nvs\n" + reference + "\n(x-factor)", align = "center", cssclass = "col_rel") - i += 1 + addHeaderColumns(suffix='%', description='x-factor', cssclass='col_rel') if options.calc_score: - i = 1 - for set in metric_sets: - reference = getSetName(test_sets[0], 0, options.columns) if not options.progress_mode else 'previous' - tbl.newColumn(str(i) + "S", getSetName(set, i, options.columns) + "\nvs\n" + reference + "\n(score)", align = "center", cssclass = "col_name") - i += 1 + addHeaderColumns(suffix='S', description='score', cssclass='col_name') # rows prevGroupName = None @@ -166,68 +195,87 @@ if __name__ == "__main__": if options.intersect_logs: needNewRow = False break - tbl.newCell(str(i), "-") - if options.calc_relatives and i > 0: - tbl.newCell(str(i) + "%", "-") - if options.calc_cr and i > 0: - tbl.newCell(str(i) + "$", "-") - if options.calc_score and i > 0: - tbl.newCell(str(i) + "$", "-") else: status = case.get("status") if status != "run": - tbl.newCell(str(i), status, color = "red") - if status != "notrun": - needNewRow = True - if options.calc_relatives and i > 0: - tbl.newCell(str(i) + "%", "-", color = "red") - if options.calc_cr and i > 0: - tbl.newCell(str(i) + "$", "-", color = "red") - if options.calc_score and i > 0: - tbl.newCell(str(i) + "S", "-", color = "red") + tbl.newCell(str(i), status, color="red") else: val = getter(case, cases[0], options.units) - def getter_fn(fn): - if fn and i > 0 and val: - for j in reversed(range(i)) if options.progress_mode else [0]: - r = cases[j] - if r is not None and r.get("status") == 'run': - return fn(case, r, options.units) - return None - valp = getter_fn(getter_p) if options.calc_relatives or options.progress_mode else None - valcr = getter_fn(getter_cr) if options.calc_cr else None - val_score = getter_fn(getter_score) if options.calc_score else None - if not valp or i == 0: - color = None - elif valp > 1.05: - color = "green" - elif valp < 0.95: - color = "red" - else: - color = None if val: needNewRow = True - tbl.newCell(str(i), formatValue(val, options.metric, options.units), val, color = color) - if options.calc_relatives and i > 0: - tbl.newCell(str(i) + "%", formatValue(valp, "%"), valp, color = color, bold = color) - if options.calc_cr and i > 0: - tbl.newCell(str(i) + "$", formatValue(valcr, "$"), valcr, color = color, bold = color) - if options.calc_score and i > 0: - tbl.newCell(str(i) + "S", formatValue(val_score, "S"), val_score, color = color, bold = color) + tbl.newCell(str(i), formatValue(val, options.metric, options.units), val) + + if needNewRow: + for link in options.regressions: + (i, reference, reverse, addColor) = link + if reverse: + i, reference = reference, i + tblCellID = str(i) + '-' + str(reference) + case = cases[i] + if case is None: + if options.calc_relatives: + tbl.newCell(tblCellID + "%", "-") + if options.calc_cr: + tbl.newCell(tblCellID + "$", "-") + if options.calc_score: + tbl.newCell(tblCellID + "$", "-") + else: + status = case.get("status") + if status != "run": + tbl.newCell(str(i), status, color="red") + if status != "notrun": + needNewRow = True + if options.calc_relatives: + tbl.newCell(tblCellID + "%", "-", color="red") + if options.calc_cr: + tbl.newCell(tblCellID + "$", "-", color="red") + if options.calc_score: + tbl.newCell(tblCellID + "S", "-", color="red") + else: + val = getter(case, cases[0], options.units) + def getRegression(fn): + if fn and val: + for j in reversed(range(i)) if reference < 0 else [reference]: + r = cases[j] + if r is not None and r.get("status") == 'run': + return fn(case, r, options.units) + valp = getRegression(getter_p) if options.calc_relatives or options.progress_mode else None + valcr = getRegression(getter_cr) if options.calc_cr else None + val_score = getRegression(getter_score) if options.calc_score else None + if not valp: + color = None + elif valp > 1.05: + color = 'green' + elif valp < 0.95: + color = 'red' + else: + color = None + if addColor: + if not reverse: + tbl.newCell(str(i), formatValue(val, options.metric, options.units), val, color=color) + else: + r = cases[reference] + if r is not None and r.get("status") == 'run': + val = getter(r, cases[0], options.units) + tbl.newCell(str(reference), formatValue(val, options.metric, options.units), val, color=color) + if options.calc_relatives: + tbl.newCell(tblCellID + "%", formatValue(valp, "%"), valp, color=color, bold=color) + if options.calc_cr: + tbl.newCell(tblCellID + "$", formatValue(valcr, "$"), valcr, color=color, bold=color) + if options.calc_score: + tbl.newCell(tblCellID + "S", formatValue(val_score, "S"), val_score, color = color, bold = color) + if not needNewRow: tbl.trimLastRow() if options.regressionsOnly: for r in reversed(range(len(tbl.rows))): - delete = True - i = 1 - for set in metric_sets: - val = tbl.rows[r].cells[len(tbl.rows[r].cells)-i].value + for i in range(1, len(options.regressions) + 1): + val = tbl.rows[r].cells[len(tbl.rows[r].cells) - i].value if val is not None and val < float(options.regressionsOnly): - delete = False - i += 1 - if (delete): + break + else: tbl.rows.pop(r) # output table