Esempio n. 1
0
 public JsonObject toJson() {
   JsonObject res = new JsonObject();
   JsonArray rows = new JsonArray();
   for (int i = 0; i < _rows.length; ++i) rows.add(new JsonPrimitive(_rows[i]));
   JsonArray dist = new JsonArray();
   for (int i = 0; i < _dist.length; ++i) dist.add(new JsonPrimitive(_dist[i]));
   res.add("rows_per_cluster", rows);
   res.add("sqr_error_per_cluster", dist);
   return res;
 }
Esempio n. 2
0
 @Override
 public JsonObject toJson() {
   JsonObject res = new JsonObject();
   res.addProperty(Constants.VERSION, H2O.VERSION);
   res.addProperty(Constants.TYPE, KMeansModel.class.getName());
   res.addProperty(Constants.ERROR, _error);
   JsonArray ary = new JsonArray();
   for (double[] dd : clusters()) {
     JsonArray ary2 = new JsonArray();
     for (double d : dd) ary2.add(new JsonPrimitive(d));
     ary.add(ary2);
   }
   res.add(Constants.CLUSTERS, ary);
   return res;
 }
Esempio n. 3
0
  @Override
  protected Response serve() {
    int tasks = 0;
    int finished = 0;
    RFModel model = _modelKey.value();
    double[] weights = _weights.value();
    // Finish refresh after rf model is done and confusion matrix for all trees is computed
    boolean done = false;
    int classCol = _classCol.specified() ? _classCol.value() : findResponseIdx(model);

    tasks = model._totalTrees;
    finished = model.size();

    // Handle cancelled/aborted jobs
    if (_job.value() != null) {
      Job jjob = Job.findJob(_job.value());
      if (jjob != null && jjob.isCancelled())
        return Response.error(
            jjob.exception == null ? "Job was cancelled by user!" : jjob.exception);
    }

    JsonObject response = defaultJsonResponse();
    // CM return and possible computation is requested
    if (!_noCM.value() && (finished == tasks || _iterativeCM.value()) && finished > 0) {
      // Compute the highest number of trees which is less then a threshold
      int modelSize = tasks * _refreshThresholdCM.value() / 100;
      modelSize =
          modelSize == 0 || finished == tasks ? finished : modelSize * (finished / modelSize);

      // Get the computing the matrix - if no job is computing, then start a new job
      Job cmJob =
          ConfusionTask.make(
              model, modelSize, _dataKey.value()._key, classCol, weights, _oobee.value());
      // Here the the job is running - it saved a CM which can be already finished or in invalid
      // state.
      CMFinal confusion = UKV.get(cmJob.dest());
      // if the matrix is valid, report it in the JSON
      if (confusion != null && confusion.valid() && modelSize > 0) {
        // finished += 1;
        JsonObject cm = new JsonObject();
        JsonArray cmHeader = new JsonArray();
        JsonArray matrix = new JsonArray();
        cm.addProperty(JSON_CM_TYPE, _oobee.value() ? "OOB error estimate" : "full scoring");
        cm.addProperty(JSON_CM_CLASS_ERR, confusion.classError());
        cm.addProperty(JSON_CM_ROWS_SKIPPED, confusion.skippedRows());
        cm.addProperty(JSON_CM_ROWS, confusion.rows());
        // create the header
        for (String s : cfDomain(confusion, 1024)) cmHeader.add(new JsonPrimitive(s));
        cm.add(JSON_CM_HEADER, cmHeader);
        // add the matrix
        final int nclasses = confusion.dimension();
        JsonArray classErrors = new JsonArray();
        for (int crow = 0; crow < nclasses; ++crow) {
          JsonArray row = new JsonArray();
          int classHitScore = 0;
          for (int ccol = 0; ccol < nclasses; ++ccol) {
            row.add(new JsonPrimitive(confusion.matrix(crow, ccol)));
            if (crow != ccol) classHitScore += confusion.matrix(crow, ccol);
          }
          // produce infinity members in case of 0.f/0
          classErrors.add(
              new JsonPrimitive(
                  (float) classHitScore / (classHitScore + confusion.matrix(crow, crow))));
          matrix.add(row);
        }
        cm.add(JSON_CM_CLASSES_ERRORS, classErrors);
        cm.add(JSON_CM_MATRIX, matrix);
        cm.addProperty(JSON_CM_TREES, modelSize);
        response.add(JSON_CM, cm);
        // Signal end only and only if all trees were generated and confusion matrix is valid
        done = finished == tasks;
      }
    } else if (_noCM.value() && finished == tasks) done = true;

    // Trees
    JsonObject trees = new JsonObject();
    trees.addProperty(Constants.TREE_COUNT, model.size());
    if (model.size() > 0) {
      trees.add(Constants.TREE_DEPTH, model.depth().toJson());
      trees.add(Constants.TREE_LEAVES, model.leaves().toJson());
    }
    response.add(Constants.TREES, trees);

    // Build a response
    Response r;
    if (done) {
      r = jobDone(response);
      r.addHeader(
          "<div class='alert'>"
              + /*RFScore.link(MODEL_KEY, model._key, "Use this model for scoring.") */ GeneratePredictionsPage
                  .link(model._key, "Predict!")
              + " </div>");
    } else {
      r = Response.poll(response, finished, tasks);
    }
    r.setBuilder(JSON_CM, new ConfusionMatrixBuilder());
    r.setBuilder(TREES, new TreeListBuilder());
    return r;
  }