Not logged in.  Login/Logout/Register | List snippets | | Create snippet | Upload image | Upload data

262
LINES

< > BotCompany Repo | #1001011 - Token prediction, multiple predictors (adding zip support)

JavaX source code [tags: use-pretranspiled] - run with: x30.jar

Libraryless. Click here for Pure Java version (2382L/19K/57K).

!747

abstract class P {
  abstract S read(L<S> tok);
  abstract P derive(); // clone & reset counter for actual use
}

m {
  static S corpusID = "#1001010";
  static int numSnippets = 3000;
  static boolean showGUI = true;
  static int maxCharsGUI = 500000;
  static boolean allTokens = true;
  
  static Collector collector;
  static L<S> tok;
  static Set<int> predicted;
  
  static class Chain extends P {
    new L<P> list;
    
    *() {}
    *(L<P> *list) {}
    *(P... a) { list = asList(a); }
    
    void add(P p) { list.add(p); }
    
    S read(L<S> tok) {
      for (P p : list) {
        S s = p.read(tok);
        if (s != null) return s;
      }
      return null;
    }
    
    P derive() {
      new Chain c;
      for (P p : list)
        c.add(p.derive());
      return c;
    }
  }
    
  static class Tuples extends P {
    Map<L<S>,S> map = new HashMap<L<S>,S>();
    int n, seen;

    *(int *n) {
    }
    
    S read(L<S> tok) {
      while (tok.size() > seen) {
        ++seen;
        if (seen > n)
          map.put(new ArrayList<S>(tok.subList(seen-n-1, seen-1)), tok.get(seen-1));
      }
      
      if (tok.size() >= n)
        return map.get(new ArrayList<S>(tok.subList(tok.size()-n, tok.size())));
        
      return null;
    }
    
    // slow...
    P oldDerive() {
      Tuples t = new Tuples(n);
      t.map.putAll(map);
      // t.seen == 0 which is ok
      return t;
    }
    
    // fast!
    P derive() {
      Tuples t = new Tuples(n);
      t.map = new DerivedHashMap<L<S>,S>(map);
      return t;
    }
  }
  
  static class DerivedHashMap<A, B> extends AbstractMap<A, B> {
    Map<A, B> base;
    new HashMap<A, B> additions;
    
    *(Map<A, B> *base) {}
    
    public B get(Object key) {
      B b = additions.get(key);
      if (b != null) return b;
      return base.get(key);
    }
    
    public B put(A key, B value) {
      return additions.put(key, value);
    }
    
    public Set<Map.Entry<A,B>> entrySet() {
      throw fail();
    }
  }
  
  // TODO: Put NewX back in
  
  p {
    tok = makeCorpusJavaTok();
    print("Tokens in corpus: " + tok.size());
    
    print("Learning...");
    collector = new Collector;
    /*test(new Tuples(1));
    test(new Tuples(2));
    test(new Tuples(3));
    test(new Tuples(4));
    test(new Chain(new Tuples(2), new Tuples(1)));*/
    test(new Chain(new Tuples(4), new Tuples(3), new Tuples(2), new Tuples(1)));

    print("Learning done.");
    if (collector.winner != null && showGUI) {
      predicted = collector.predicted;
      showColoredText();
    }
  }
  
  // test a predictor
  static void test(P p) {
    predicted = new TreeSet<int>();
    int points = 0, total = 0, lastPercent = 0;
    new L<S> history;
    for (int i = allTokens ? 0 : 1; i < tok.size(); i += allTokens ? 1 : 2) {
      S t = tok.get(i);
      S x = p.read(history);
      boolean correct = t.equals(x);
      total += t.length();
      if (correct) {
        predicted.add(i);
        points += t.length();
      }
      history.add(t);
      int percent = roundUpTo(10, (int) (i*100L/tok.size()));
      if (percent > lastPercent) {
        print("Learning " + percent + "% done.");
        lastPercent = percent;
      }
    }
    double score = points*100.0/total;
    collector.add(p, score);
  }
  
  static void showColoredText() ctex {
    JFrame jf = new JFrame("Predicted = green");
    Container cp = jf.getContentPane();

    JTextPane pane = new JTextPane();
    //pane.setFont(loadFont("#1000993", 24));
    Document doc = pane.getStyledDocument();

    int i = tok.size(), len = 0;
    while (len <= maxCharsGUI && i > 0) {
      --i;
      len += tok.get(i).length();
    }
    
    for (; i < tok.size(); i++) {
      if (tok.get(i).length() == 0) continue;
      boolean green = predicted.contains(i);
      SimpleAttributeSet set = new SimpleAttributeSet();
      StyleConstants.setForeground(set, green ? Color.green : Color.gray);
      doc.insertString(doc.getLength(), tok.get(i), set);
    }
    
    JScrollPane scrollPane = new JScrollPane(pane);
    cp.add(scrollPane, BorderLayout.CENTER);

    jf.setBounds(100, 100, 600, 600);
    jf.setVisible(true);
  }
  
  !include #1000989 // SnippetDB
  
  static L<S> makeCorpusJavaTok() {
    S name = getSnippetTitle(corpusID);
    if (name.toLowerCase().indexOf(".zip") >= 0)
      return makeCorpus_zip();
    else
      return makeCorpus_mysqldump();
  }
  
  static L<S> makeCorpus_zip() ctex {
    ZipFile zipFile = new ZipFile(loadLibrary(corpusID));
    Enumeration entries = zipFile.entries();
    new L<S> tok;
    
    while (entries.hasMoreElements()) {
      ZipEntry entry = (ZipEntry) entries.nextElement(); 
      //System.out.println("File found: " + entry.getName());

      InputStream fin = zipFile.getInputStream(entry);
      // TODO: try to skip binary files?
      
      InputStreamReader reader = new InputStreamReader(fin, "UTF-8");
      new StringBuilder builder;
      BufferedReader bufferedReader = new BufferedReader(reader);
      String line;
      while ((line = bufferedReader.readLine()) != null)
        builder.append(line).append('\n');
      fin.close();
      
      new StringBuilder buf;
      buf.append("\n== File: " + entry.getName());
      buf.append("\n==\n");
      buf.append(builder.toString()).append("\n");
      if (tok.size() != 0) tok.remove(tok.size()-1);
      tok.addAll(javaTok(buf.toString()));
    }
    
    zipFile.close();
    return internAll(tok);
  }
  
  static L<S> makeCorpus_mysqldump() {
    SnippetDB db = new SnippetDB(corpusID);
    List<List<S>> rows = db.rowsOrderedBy("sn_created");
    new L<S> tok;
    for (int i = 0; i < Math.min(rows.size(), numSnippets); i++) {
      new StringBuilder buf;
      S id = db.getField(rows.get(i), "sn_id");
      S title = db.getField(rows.get(i), "sn_title");
      S text = db.getField(rows.get(i), "sn_text");
      buf.append("\n== ID: " + id);
      buf.append("\n== Title: " + title);
      buf.append("\n==\n");
      buf.append(text).append("\n");
      if (tok.size() != 0) tok.remove(tok.size()-1);
      tok.addAll(javaTok(buf.toString()));
      ++i;
    }
    return internAll(tok);
  }
  
  static L<S> internAll(L<S> tok) {
    new L<S> l;
    for (S t : tok)
      l.add(t.intern());
    return l;
  }
  
  static class Collector {
    P winner;
    double bestScore = -1;
    Set<int> predicted;

    void add(P p, double score) {
      if (winner == null || score > bestScore) {
        winner = p;
        bestScore = score;
        //S name = shorten(structure(p), 100);
        S name = p.getClass().getName();
        print("New best score: " + formatDouble(score, 2) + "% (" + name + ")");
        this.predicted = main.predicted;
      }
    }
  }
}

Author comment

Began life as a copy of #1001000

download  show line numbers  debug dex   

Travelled to 13 computer(s): aoiabmzegqzx, bhatertpkbcr, cbybwowwnfue, cfunsshuasjs, gwrvuhgaqvyk, ishqpsrjomds, lpdgvwnxivlt, mqqgnosmbjvj, onxytkatvevr, pzhvpgtvlbxg, teubizvjbppd, tslmcundralx, tvejysmllsmz

No comments. add comment

Snippet ID: #1001011
Snippet name: Token prediction, multiple predictors (adding zip support)
Eternal ID of this version: #1001011/1
Text MD5: 942f3ed24c4432b998f1e22ebdd4e9fe
Transpilation MD5: 78b9d92dea6ca60c5e4296d797492666
Author: stefan
Category:
Type: JavaX source code
Public (visible to everyone): Yes
Archived (hidden from active list): No
Created/modified: 2015-09-16 00:58:02
Source code size: 7026 bytes / 262 lines
Pitched / IR pitched: No / Yes
Views / Downloads: 404 / 594
Referenced in: [show references]

Formerly at http://tinybrain.de/1001011 & http://1001011.tinybrain.de