| 1 | package tuffy.infer; |
| 2 | |
| 3 | |
| 4 | import java.util.ArrayList; |
| 5 | |
| 6 | import tuffy.ground.partition.Bucket; |
| 7 | import tuffy.ground.partition.Component; |
| 8 | import tuffy.ground.partition.Partition; |
| 9 | import tuffy.infer.MRF.INIT_STRATEGY; |
| 10 | import tuffy.infer.ds.GAtom; |
| 11 | import tuffy.util.Config; |
| 12 | import tuffy.util.ExceptionMan; |
| 13 | import tuffy.util.Settings; |
| 14 | import tuffy.util.UIMan; |
| 15 | /** |
| 16 | * |
| 17 | * A bucket of inference tasks that can run in prallel. Currently, each task |
| 18 | * correspond to an MRF component so that the components can be processed in parallel. |
| 19 | */ |
| 20 | public class InferBucket{ |
| 21 | private Bucket bucket; |
| 22 | private int numThreads = 2; |
| 23 | private double cost; |
| 24 | |
| 25 | private int totalSamples = 0; |
| 26 | |
| 27 | Config.TUFFY_INFERENCE_TASK task = null; |
| 28 | Settings settings = null; |
| 29 | |
| 30 | public void infer(Settings s){ |
| 31 | settings = s; |
| 32 | task = Config.TUFFY_INFERENCE_TASK.valueOf(s.getString("task")); |
| 33 | this.runInferParallel(); |
| 34 | } |
| 35 | |
| 36 | |
| 37 | public InferBucket(Bucket bucket){ |
| 38 | this.bucket = bucket; |
| 39 | numThreads = Config.getNumThreads(); |
| 40 | } |
| 41 | |
| 42 | public void flushAtomStates(DataMover dmover, String relAtoms){ |
| 43 | ArrayList<GAtom> gatoms = new ArrayList<GAtom>(); |
| 44 | for(Component c: bucket.getComponents()){ |
| 45 | if(c.atoms == null) continue; |
| 46 | gatoms.addAll(c.atoms.values()); |
| 47 | } |
| 48 | dmover.flushAtomStates(gatoms, relAtoms); |
| 49 | } |
| 50 | |
| 51 | public void setMrfInitStrategy(INIT_STRATEGY strategy){ |
| 52 | for(Partition p : bucket.getPartitions()){ |
| 53 | if(p.mrf == null) continue; |
| 54 | p.mrf.setInitStrategy(strategy); |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | /** |
| 59 | * Get the cost after inference. |
| 60 | */ |
| 61 | public double getCost(){ |
| 62 | return cost; |
| 63 | } |
| 64 | |
| 65 | public double getSamples(){ |
| 66 | return this.totalSamples; |
| 67 | } |
| 68 | |
| 69 | private Object sentinel = new Object(); |
| 70 | |
| 71 | /** |
| 72 | * Add up the cost. |
| 73 | * |
| 74 | * @see CompWorker#run() |
| 75 | * @param c |
| 76 | */ |
| 77 | public void addCost(double c){ |
| 78 | synchronized(sentinel){ |
| 79 | cost += c; |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | /** |
| 84 | * Add up the cost. |
| 85 | * |
| 86 | * @see CompWorker#run() |
| 87 | * @param c |
| 88 | */ |
| 89 | public void addSamples(int c){ |
| 90 | synchronized(sentinel){ |
| 91 | this.totalSamples += c; |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | /** |
| 96 | * A worker thread that runs inference on one component at a time. |
| 97 | * Used for parallelization in a producer-consumer model. |
| 98 | */ |
| 99 | public static class CompWorker extends Thread{ |
| 100 | InferBucket ibucket; |
| 101 | Config.TUFFY_INFERENCE_TASK task; |
| 102 | Settings settings; |
| 103 | |
| 104 | public CompWorker(InferBucket bucket){ |
| 105 | this.ibucket = bucket; |
| 106 | settings = bucket.settings; |
| 107 | task = bucket.task; |
| 108 | } |
| 109 | |
| 110 | public void run(){ |
| 111 | while(true){ |
| 112 | Component comp = ibucket.getTask(); |
| 113 | if(comp == null) return; |
| 114 | //UIMan.println(this + " is processing " + comp + " with " + comp.numAtoms + " atoms and " + comp.numClauses); |
| 115 | InferComponent ic = new InferComponent(comp); |
| 116 | switch(task){ |
| 117 | case MAP: |
| 118 | int ntries = (Integer)(settings.get("ntries")); |
| 119 | double flipsPerAtom = (Double)(settings.get("flipsPerAtom")); |
| 120 | int nflips = (int)(flipsPerAtom * comp.numAtoms); |
| 121 | //UIMan.println("flips = " + nflips + "; tries = " + ntries); |
| 122 | ic.inferMAP(ntries, nflips); |
| 123 | ibucket.addCost(ic.getCost()); |
| 124 | break; |
| 125 | case MARGINAL: |
| 126 | int nsamples = (Integer)(settings.get("nsamples")); |
| 127 | flipsPerAtom = (Double)(settings.get("flipsPerAtom")); |
| 128 | nflips = (int)(flipsPerAtom * comp.numAtoms); |
| 129 | ibucket.addCost(ic.inferMarginal(nsamples, nflips)); |
| 130 | break; |
| 131 | } |
| 132 | } |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | /** |
| 137 | * The queue of components to be processed |
| 138 | */ |
| 139 | private ArrayList<Component> q = null; |
| 140 | |
| 141 | /** |
| 142 | * Get the next unprocessed component in the queue |
| 143 | * |
| 144 | * @see CompWorker#run() |
| 145 | */ |
| 146 | public Component getTask(){ |
| 147 | synchronized(sentinel){ |
| 148 | if(q.isEmpty()) return null; |
| 149 | Component c = q.remove(q.size()-1); |
| 150 | //UIMan.println(" got " + c + " from the queue"); |
| 151 | return c; |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | /** |
| 156 | * Solve the components in parallel. |
| 157 | * @param ntries |
| 158 | * @param nflips |
| 159 | */ |
| 160 | private void runInferParallel(){ |
| 161 | cost = 0; |
| 162 | q = new ArrayList<Component>(); |
| 163 | q.addAll(bucket.getComponents()); |
| 164 | ArrayList<CompWorker> workers = new ArrayList<CompWorker>(); |
| 165 | |
| 166 | if(numThreads > 1){ |
| 167 | UIMan.setSilent(true); |
| 168 | } |
| 169 | for(int i=0; i<numThreads; i++){ |
| 170 | CompWorker t = new CompWorker(this); |
| 171 | workers.add(t); |
| 172 | t.start(); |
| 173 | } |
| 174 | for(CompWorker t : workers){ |
| 175 | try { |
| 176 | t.join(); |
| 177 | } catch (InterruptedException e) { |
| 178 | ExceptionMan.handle(e); |
| 179 | } |
| 180 | } |
| 181 | UIMan.setSilent(false); |
| 182 | } |
| 183 | |
| 184 | |
| 185 | public void setNumThreads(int numThreads) { |
| 186 | this.numThreads = numThreads; |
| 187 | } |
| 188 | |
| 189 | |
| 190 | public int getNumThreads() { |
| 191 | return numThreads; |
| 192 | } |
| 193 | |
| 194 | |
| 195 | |
| 196 | } |