MixParser.java

package edu.stanford.nlp.sempre;

import java.util.*;
import java.util.regex.Pattern;

import fig.basic.LogInfo;
import fig.basic.Option;
import fig.basic.Pair;

/**
 * A parser that mixes the derivation lists from other parsers.
 *
 * @author ppasupat
 */
public class MixParser extends Parser {
  public static class Options {
    @Option(gloss = "verbosity")
    public int verbose = 1;

    /**
     * Syntax: [className]:[options]
     * - className also includes the package name (BeamParser, FloatingParser, tables.dpd.DPDParser, ...)
     * - options is a comma-separated list of [group] or [group]-[iter1]-[iter2]-...
     *   [iter1], [iter2], ... are iteration indices (0-based), "xc" (computing expected counts),
     *   "nxc" (not computing expected counts), index + "xc", or index + "nxc"
     *
     * Example: Using
     *      -MixParser.parsers FloatingParser tables.serialize.SerializedParser:train-0-2-3,dev
     *   will cause derivations from SerializedParser to be mixed in during all dev iterations
     *   and training iterations 0, 2, and 3.
     */
    @Option(gloss = "list of parsers to use along with options")
    public List<String> parsers = new ArrayList<>();
  }
  public static Options opts = new Options();

  final List<Pair<Parser, MixParserOption>> parsers;
  int iter, numIters;
  String group;

  public MixParser(Spec spec) {
    super(spec);
    parsers = new ArrayList<>();
    for (String parserAndOptions : opts.parsers) {
      if (opts.verbose >= 1)
        LogInfo.logs("Adding parser %s", parserAndOptions);
      String[] tokens = parserAndOptions.split(":");
      if (tokens.length > 2)
        throw new RuntimeException("Invalid parser options: " + parserAndOptions);
      String parserName = tokens[0];
      Parser parser;
      try {
        Class<?> parserClass = Class.forName(SempreUtils.resolveClassName(parserName));
        parser = ((Parser) parserClass.getConstructor(spec.getClass()).newInstance(spec));
      } catch (ClassNotFoundException e1) {
        throw new RuntimeException("Illegal parser: " + parserName);
      } catch (Exception e) {
        e.printStackTrace();
        throw new RuntimeException("Error while instantiating parser: " + parserName + "\n" + e);
      }
      if (tokens.length > 1)
        parsers.add(new Pair<>(parser, new MixParserOption(this, tokens[1])));
      else
        parsers.add(new Pair<>(parser, new MixParserOption(this)));
    }
  }

  // Don't do it.
  @Override protected void computeCatUnaryRules() {
    catUnaryRules = Collections.emptyList();
  };

  @Override
  public ParserState newParserState(Params params, Example ex, boolean computeExpectedCounts) {
    return new MixParserState(this, params, ex, computeExpectedCounts);
  }

  @Override
  public void onBeginDataGroup(int iter, int numIters, String group) {
    this.iter = iter;
    this.numIters = numIters;
    this.group = group;
  }
}

class MixParserOption {
  private final MixParser mixParser;
  private boolean allowedAll = false;
  private List<Pair<String, String>> allowedGroupsAndIter = new ArrayList<>();

  public MixParserOption(MixParser mixParser) {
    this.mixParser = mixParser;
    // Allow in all groups
    allowedAll = true;
  }

  public MixParserOption(MixParser mixParser, String optionString) {
    this.mixParser = mixParser;
    String[] tokens = optionString.split(",");
    for (String option : tokens) {
      String[] subtokens = option.split("-");
      if (subtokens.length == 1)
        allowedGroupsAndIter.add(new Pair<>(subtokens[0], "all"));
      else
        for (int i = 1; i < subtokens.length; i++) {
          if (!Pattern.matches("^([0-9]*(n?xc)?)$", subtokens[i]))
            throw new RuntimeException("Invalid iteration options: " + subtokens[i]);
          allowedGroupsAndIter.add(new Pair<>(subtokens[0], subtokens[i]));
        }
    }
  }

  public boolean isAllowed(boolean computeExpectedCounts) {
    if (allowedAll) return true;
    String xcFlag = computeExpectedCounts ? "xc" : "nxc";
    return allowedGroupsAndIter.contains(new Pair<>(mixParser.group, "all"))
        || allowedGroupsAndIter.contains(new Pair<>(mixParser.group, "" + mixParser.iter))
        || allowedGroupsAndIter.contains(new Pair<>(mixParser.group, xcFlag))
        || allowedGroupsAndIter.contains(new Pair<>(mixParser.group, "" + mixParser.iter + xcFlag));
  }
}

class MixParserState extends ParserState {

  public MixParserState(Parser parser, Params params, Example ex, boolean computeExpectedCounts) {
    super(parser, params, ex, computeExpectedCounts);
  }

  @Override
  public void infer() {
    for (Pair<Parser, MixParserOption> pair : ((MixParser) parser).parsers) {
      if (!pair.getSecond().isAllowed(computeExpectedCounts)) {
        if (MixParser.opts.verbose >= 1)
          LogInfo.logs("Skipping %s", pair.getFirst().getClass().getSimpleName());
        continue;
      }
      if (MixParser.opts.verbose >= 1)
        LogInfo.begin_track("Using %s", pair.getFirst().getClass().getSimpleName());
      ParserState parserState = pair.getFirst().newParserState(params, ex, false);
      parserState.infer();
      predDerivations.addAll(parserState.predDerivations);
      if (MixParser.opts.verbose >= 1) {
        LogInfo.logs("Number of derivations: %d", parserState.predDerivations.size());
        LogInfo.end_track();
      }

    }
    ensureExecuted();
    if (computeExpectedCounts) {
      expectedCounts = new HashMap<>();
      ParserState.computeExpectedCounts(predDerivations, expectedCounts);
    }
  }

}