private void processExamples(String group, List<Example> examples) {
Evaluation evaluation = new Evaluation();
if (examples.isEmpty()) return;
final String prefix = "iter=0." + group;
Execution.putOutput("group", group);
LogInfo.begin_track_printAll("Processing %s: %s examples", prefix, examples.size());
LogInfo.begin_track("Dumping metadata");
dumpMetadata(group, examples);
LogInfo.end_track();
LogInfo.begin_track("Examples");
for (int e = 0; e < examples.size(); e++) {
Example ex = examples.get(e);
LogInfo.begin_track_printAll("%s: example %s/%s: %s", prefix, e, examples.size(), ex.id);
ex.log();
Execution.putOutput("example", e);
StopWatchSet.begin("Parser.parse");
ParserState state = builder.parser.parse(params, ex, false);
StopWatchSet.end();
out.printf("########## Example %s ##########\n", ex.id);
dumpExample(exampleToLispTree(state));
LogInfo.logs("Current: %s", ex.evaluation.summary());
evaluation.add(ex.evaluation);
LogInfo.logs("Cumulative(%s): %s", prefix, evaluation.summary());
LogInfo.end_track();
ex.predDerivations.clear(); // To save memory
}
LogInfo.end_track();
LogInfo.logs("Stats for %s: %s", prefix, evaluation.summary());
evaluation.logStats(prefix);
evaluation.putOutput(prefix);
LogInfo.end_track();
}
@Before
public void setUp() {
String args =
"-modelType dmv "
+ "-Options.stage1.numIters 60 "
+ "-numThreads 1 "
+ "-inputLists "
// + "../wsj/train/ "
// + "wsjLists/unsupervised-train-wsj "
+ "../wsj/3.0/conll/wsj-10-noP-dev.deps "
// + "-inputFileExt mrg "
// + "-inputFormat mrg "
+ "-examplesInSingleFile "
+ "-inputFormat conll "
+ "-dontOutputParams "
// + "-maxExampleLength 10 "
// + "-removePunctuation "
+ "-connlTagPos 3 "
+ "-connlHeadPos 6 "
+ "-useTagsAsWords";
/*initialisation procedure from Induction class*/
Options opts = new Options();
Execution.init(args.split(" "), new Object[] {opts}); // parse input params
model = new GenerativeDMVModel(opts);
model.readExamples();
model.logStats();
model.preInit();
model.init(InitType.uniformz, null, "stage1");
opts.outputIterFreq = opts.stage1.numIters;
lopts = opts.stage1;
name = "stage1";
}
@Override
public void run() {
builder = new Builder();
builder.build();
dataset = new Dataset();
dataset.read();
params = new Params();
for (String group : dataset.groups()) {
String filename = Execution.getFile("dumped-" + group + ".gz");
out = IOUtils.openOutHard(filename);
processExamples(group, dataset.examples(group));
out.close();
LogInfo.logs("Finished dumping to %s", filename);
StopWatchSet.logStats();
}
}
class Eval {
// Predictions text file: context, actual, predicted, rank
private static final PrintWriter predOut = IOUtils.openOutEasy(Execution.getFile("predictions"));
// Machine-readable entropies: file, location, isIdent, entropy, reciprocal
// rank
private static final PrintWriter entOut = IOUtils.openOutEasy(Execution.getFile("entropies"));
public static Evaluation eval(InferState state) {
// Print out information about how well we're doing.
Evaluation evaluation = new Evaluation();
Candidate trueCandidate = state.getTrueCandidate();
Candidate predCandidate = state.getCandidates().get(0);
PredictionContext context = state.getContext();
NgramContext ngramContext = NgramContext.get(context);
Statistics statistics = state.getStatistics();
Corpus corpus = statistics.getProjectLangCorpus(context.getPath());
DataSummary summary = statistics.getStatistic(NgramKNCounts.class, corpus).getSummary();
Params params = state.getParams();
boolean oracle = state.isOracle();
int rank = state.getRank();
double entropy = state.getEntropy();
double reciprocalRank = state.getReciprocalRank();
boolean isIdent = state.isIdent();
boolean correct = state.isCorrect();
String path = context.getPath();
String trueTokenStr = trueCandidate.token;
String predToken = predCandidate.token;
evaluation.add("accuracy", correct);
evaluation.add("oracle", oracle);
evaluation.add("rank", rank);
evaluation.add("reciprocalRank", reciprocalRank);
if (oracle) {
evaluation.add("entropy", entropy);
}
if (isIdent) {
evaluation.add("identAccuracy", correct);
evaluation.add("identOracle", oracle);
if (oracle) {
evaluation.add("identEntropy", entropy);
evaluation.add("identReciprocalRank", reciprocalRank);
for (int i = 0; i < Main.clusters; i++) {
evaluation.add("identEntropy" + i, -Math.log(trueCandidate.clusterProbs[i]));
}
}
}
String contextStr = ngramContext.contextStr();
if (Main.verbose >= 2) {
String entropyStr = oracle ? Fmt.D(entropy) : "N/A";
begin_track(
"Example %s [%s]: %s (%d candidates, rank %s, entropy %s)",
path,
correct ? "CORRECT" : "WRONG",
contextStr,
state.getCandidates().size(),
rank,
entropyStr);
logs("True (prob= %s): [%s]", Fmt.D(trueCandidate.prob), trueTokenStr);
logs("Pred (prob= %s): [%s]", Fmt.D(predCandidate.prob), predToken);
if (oracle) {
KneserNey.logKNIs(true);
KneserNey.computeProb(CandidateNgram.get(context, trueCandidate), summary);
KneserNey.logKNIs(false);
}
// begin_track("True");
FeatureVector.logFeatureWeights("True", trueCandidate.features.toMap(), params);
// for (int i = 0; i < Main.clusters; i++) {
// logs("cluster=%d, score %s, prob %s", i, Fmt.D(trueCandidate.clusterScores[i]),
// Fmt.D(trueCandidate.clusterProbs[i]));
// FeatureVector.logFeatureWeights("cluster=" + i,
// trueCandidate.clusterFeatures.toMap(),
// params, Main.clusterDecorators[i]);
// }
// end_track();
KneserNey.logKNIs(true);
KneserNey.computeProb(CandidateNgram.get(context, predCandidate), summary);
KneserNey.logKNIs(false);
FeatureVector.logFeatureWeights("Pred", predCandidate.features.toMap(), params);
// for (Candidate candidate : candidates) {
// begin_track("Candidate " + candidate.token);
// for (int i = 0; i < Main.clusters; i++) {
// logs("cluster=%d, score %s, prob %s", i, Fmt.D(candidate.clusterScores[i]),
// Fmt.D(candidate.clusterProbs[i]));
// FeatureVector.logFeatureWeights("cluster=" + i,
// candidate.clusterFeatures.toMap(),
// params, Main.clusterDecorators[i]);
// }
// end_track();
// }
FeatureVector.logFeatureDiff(
"True - Pred", trueCandidate.features, predCandidate.features, params);
end_track();
}
// Longest context that has been seen
int context_max_n = ngramContext.getMax_n() - 1;
while (context_max_n > 0
&& !summary.counts[context_max_n].containsKey(ngramContext.subContext(context_max_n + 1)))
context_max_n--;
evaluation.add("context_max_n", context_max_n);
predOut.println(
"path="
+ path
+ "\tident="
+ (isIdent ? 1 : 0)
+ "\tcontext="
+ contextStr
+ "\tcontext_max_n="
+ context_max_n
+ "\ttrue="
+ trueTokenStr
+ "\tpred="
+ predToken
+ "\trank="
+ rank
+ "\tentropy="
+ entropy);
predOut.flush();
entOut.println(
path
+ "\t"
+ state.getTrueToken().loc()
+ "\t"
+ (isIdent ? 1 : 0)
+ "\t"
+ (oracle ? entropy : (state.isOov() ? "oov" : "offBeam"))
+ "\t"
+ reciprocalRank);
entOut.flush();
return evaluation;
}
}
public static void main(String[] args) {
Execution.run(args, new TreeIterable());
}
public static void main(String[] args) {
Execution.run(args, "SerializedDumperMain", new SerializedDumper(), Master.getOptionsParser());
}