@Override public void process(final DAVOptions options) { final String[] args = getOriginalArgs(); final UseModality<DAVOptions> executed; final int maxSplitIndex = splitPlan.getMaxSplitIndex(); final ProgressLogger logger = new ProgressLogger(LOGGER); logger.expectedUpdates = maxSplitIndex; logger.itemsName = "splits"; logger.priority = Level.INFO; logger.start("Parallel split processing"); final SplitParallelRegion region = new SplitParallelRegion(maxSplitIndex, args, logger); try { getParallelTeam().execute(region); } catch (Exception e) { LOGGER.error("An exception occurred.", e); } logger.stop(); /** Time the duration of the sequence: */ timeService.setModelId(modelId); timeService.stop(); executed = region.getExecuted(); if (executed != null && executed instanceof SequenceMode) { // if we executed SequenceMode final SequenceMode sequenceMode = (SequenceMode) executed; if (evaluateStatistics) { final String label = sequenceMode.getValue("label"); final String statsFilename = sequenceMode.getValue("predictions-filename"); if (statsFilename != null && label != null) { // and the sequence defined the variables "predictions-filename" and "label" try { final List<String> statsModeArgs = new ObjectArrayList<String>( new String[] { "--mode", "stats", "--predictions", statsFilename, "--submission-file", labelPrefix(label) + "-maqcii-submission.txt", "--label", label, "--model-id", modelId, "--dataset-name", options.datasetName, "--other-measures", "prec,rec,F-1,MCC,binary-auc" }); if (options.adjustSignalToFloorValue) { statsModeArgs.add("--floor"); statsModeArgs.add(Double.toString(options.signalFloorValue)); } // extract survival options if any // TODO: clean this up - we should not be checking for "%survival%" final String survivalFileName = sequenceMode.getValue("survival"); if (StringUtils.isNotBlank(survivalFileName) && !"%survival%".equals(survivalFileName)) { statsModeArgs.add("--survival"); statsModeArgs.add(survivalFileName); } LOGGER.debug("Estimating statistics: " + statsModeArgs); // we create a new DAVMode here since we want to use the old StatsMode code // which is no longer exposed by DiscoverAndValidate (BDVal main method) final DAVMode statsMode = new DAVMode(); statsMode.registerMode("stats", StatsMode.class); final DAVOptions statsModeOptions = new DAVOptions(); statsMode.process( statsModeArgs.toArray(new String[statsModeArgs.size()]), statsModeOptions); } catch (Exception e) { LOGGER.error("Error executing --mode stats for all splits", e); } } } } }
/** * For a specific sub-set of blocks (child nodes), find a 'base' subset of parents for which the * block's logLikelihood is not -Infinity * * @param candidateParentsPerNode * @param chosenArcsPerNode * @param setOfBlocks * @return */ protected double getOutOfMinusInfinity( Int2ObjectOpenHashMap<IntOpenHashSet> candidateParentsPerNode, Int2ObjectOpenHashMap<ObjectOpenHashSet<Arc>> chosenArcsPerNode, IntOpenHashSet setOfBlocks, TIntDoubleHashMap logLPerNode) { double totalLogL = 0; ProgressLogger pl = new ProgressLogger(LOGGER, ProgressLogger.TEN_SECONDS, "blocks"); pl.start("Begin initializing, to avoid zero likelihood, using set-cover heuristic"); pl.expectedUpdates = setOfBlocks.size(); int nArcs = 0; for (int v : setOfBlocks) { pl.update(); IntOpenHashSet vParents = candidateParentsPerNode.get(v); Int2ObjectOpenHashMap<IntOpenHashSet> parentActions = new Int2ObjectOpenHashMap<IntOpenHashSet>(); Int2ObjectOpenHashMap<IntArrayList> cPlusV = auxiliary.getCplusOnline(v); Int2ObjectOpenHashMap<IntArrayList> cMinusV = auxiliary.getCminusOnline(v); if (cPlusV != null) { IntSet actions = cPlusV.keySet(); // Heuristic: first add the parents that participate in A+ for // most actions for (int action : actions) { for (int u : cPlusV.get(action)) { if (!parentActions.containsKey(u)) { parentActions.put(u, new IntOpenHashSet()); } parentActions.get(u).add(action); } } } KeepMaximum km = new KeepMaximum(); km.addAllKey2Listsize(parentActions); IntOpenHashSet baseSetOfParents = new IntOpenHashSet(); double logL = Double.NEGATIVE_INFINITY; while (logL == Double.NEGATIVE_INFINITY && (km.getMaximumKey() != -1)) { int u = km.getMaximumKey(); if (baseSetOfParents.contains(u)) { throw new IllegalStateException("Attempted to add twice the same parent"); } baseSetOfParents.add(u); logL = blockLogLikelihood(v, cPlusV, cMinusV, baseSetOfParents); IntOpenHashSet uActions = parentActions.get(u); for (int parent : vParents) { parentActions.get(parent).removeAll(uActions); } vParents.remove(u); parentActions.remove(u); km.reset(); km.addAllKey2Listsize(parentActions); } // keep track of the likelihood totalLogL += logL; if (logLPerNode != null) { logLPerNode.put(v, logL); } chosenArcsPerNode.put(v, new ObjectOpenHashSet<Arc>()); for (int u : baseSetOfParents) { nArcs++; chosenArcsPerNode.get(v).add(new Arc(u, v)); } } pl.stop("Done initialization. Added " + nArcs + " arcs, logLikelihood=" + totalLogL); return totalLogL; }