Examples with Chunk water.fvec.Chunk used on opensource projects

Example 61 with Chunk

use of water.fvec.Chunk in project h2o-2 by h2oai.

the class MRUtils method div.

public static Frame div(Frame fr, final double d) {
    Frame r = new MRTask2() {

        @Override
        public void map(Chunk[] cs, NewChunk[] ncs) {
            for (int i = 0; i < ncs.length; i++) {
                NewChunk nc = ncs[i];
                Chunk c = cs[i];
                for (int r = 0; r < c._len; r++) if (d != 0)
                    nc.addNum(c.at0(r) * d);
                else
                    nc.addNA();
            }
        }
    }.doAll(fr.numCols(), fr).outputFrame(fr.names(), fr.domains());
    return r;
}

Example 62 with Chunk

use of water.fvec.Chunk in project h2o-3 by h2oai.

the class GBMTest method highCardinality.

public void highCardinality(int nbins_cats) {
    GBMModel gbm = null;
    GBMModel.GBMParameters parms = new GBMModel.GBMParameters();
    Frame train = null, test = null, train_preds = null, test_preds = null;
    Scope.enter();
    try {
        {
            CreateFrame cf = new CreateFrame();
            cf.rows = 10000;
            cf.cols = 10;
            cf.integer_range = 1000;
            cf.categorical_fraction = 1.0;
            cf.integer_fraction = 0.0;
            cf.binary_fraction = 0.0;
            cf.time_fraction = 0.0;
            cf.string_fraction = 0.0;
            cf.binary_ones_fraction = 0.0;
            cf.missing_fraction = 0.2;
            cf.factors = 3000;
            cf.response_factors = 2;
            cf.positive_response = false;
            cf.has_response = true;
            cf.seed = 1235;
            cf.seed_for_column_types = 1234;
            train = cf.execImpl().get();
        }
        {
            CreateFrame cf = new CreateFrame();
            cf.rows = 10000;
            cf.cols = 10;
            cf.integer_range = 1000;
            cf.categorical_fraction = 1.0;
            cf.integer_fraction = 0.0;
            cf.binary_fraction = 0.0;
            cf.time_fraction = 0.0;
            cf.string_fraction = 0.0;
            cf.binary_ones_fraction = 0.0;
            cf.missing_fraction = 0.2;
            cf.factors = 5000;
            cf.response_factors = 2;
            cf.positive_response = false;
            cf.has_response = true;
            cf.seed = 5321;
            cf.seed_for_column_types = 1234;
            test = cf.execImpl().get();
        }
        parms._train = train._key;
        // Train on the outcome
        parms._response_column = "response";
        //allow it to overfit
        parms._max_depth = 20;
        parms._min_rows = 1;
        parms._ntrees = 1;
        parms._nbins_cats = nbins_cats;
        parms._seed = 0x2834234;
        GBM job = new GBM(parms);
        gbm = job.trainModel().get();
        train_preds = gbm.score(train);
        test_preds = gbm.score(test);
        new MRTask() {

            public void map(Chunk c) {
                for (int i = 0; i < c._len; ++i) if (c.isNA(i))
                    c.set(i, 0.5);
            }
        }.doAll(train.vec("response"));
        new MRTask() {

            public void map(Chunk c) {
                for (int i = 0; i < c._len; ++i) if (c.isNA(i))
                    c.set(i, 0.5);
            }
        }.doAll(test.vec("response"));
        Log.info("Train AUC: " + ModelMetricsBinomial.make(train_preds.vec(2), train.vec("response")).auc());
        Log.info("Test AUC: " + ModelMetricsBinomial.make(test_preds.vec(2), test.vec("response")).auc());
        // Build a POJO, validate same results
        Assert.assertTrue(gbm.testJavaScoring(train, train_preds, 1e-15));
        Key old = gbm._key;
        gbm._key = Key.make(gbm._key + "ha");
        Assert.assertTrue(gbm.testJavaScoring(test, test_preds, 1e-15));
        DKV.remove(old);
    } finally {
        if (gbm != null)
            gbm.delete();
        if (train != null)
            train.remove();
        if (test != null)
            test.remove();
        if (train_preds != null)
            train_preds.remove();
        if (test_preds != null)
            test_preds.remove();
        Scope.exit();
    }
}

Example 63 with Chunk

use of water.fvec.Chunk in project h2o-3 by h2oai.

the class AstImpute method apply.

// (h2o.impute data col method combine_method groupby groupByFrame values)
@Override
public Val apply(Env env, Env.StackHelp stk, AstRoot[] asts) {
    // Argument parsing and sanity checking
    // Whole frame being imputed
    Frame fr = stk.track(asts[1].exec(env)).getFrame();
    // Column within frame being imputed
    final int col = (int) asts[2].exec(env).getNum();
    if (col >= fr.numCols())
        throw new IllegalArgumentException("Column not -1 or in range 0 to " + fr.numCols());
    final boolean doAllVecs = col == -1;
    final Vec vec = doAllVecs ? null : fr.vec(col);
    // Technique used for imputation
    AstRoot method = null;
    boolean ffill0 = false, bfill0 = false;
    switch(asts[3].exec(env).getStr().toUpperCase()) {
        case "MEAN":
            method = new AstMean();
            break;
        case "MEDIAN":
            method = new AstMedian();
            break;
        case "MODE":
            method = new AstMode();
            break;
        case "FFILL":
            ffill0 = true;
            break;
        case "BFILL":
            bfill0 = true;
            break;
        default:
            throw new IllegalArgumentException("Method must be one of mean, median or mode");
    }
    // Only for median, how is the median computed on even sample sizes?
    QuantileModel.CombineMethod combine = QuantileModel.CombineMethod.valueOf(asts[4].exec(env).getStr().toUpperCase());
    // Group-by columns.  Empty is allowed, and perfectly normal.
    AstRoot ast = asts[5];
    AstNumList by2;
    if (ast instanceof AstNumList)
        by2 = (AstNumList) ast;
    else if (ast instanceof AstNum)
        by2 = new AstNumList(((AstNum) ast).getNum());
    else if (ast instanceof AstStrList) {
        String[] names = ((AstStrList) ast)._strs;
        double[] list = new double[names.length];
        int i = 0;
        for (String name : ((AstStrList) ast)._strs) list[i++] = fr.find(name);
        Arrays.sort(list);
        by2 = new AstNumList(list);
    } else
        throw new IllegalArgumentException("Requires a number-list, but found a " + ast.getClass());
    Frame groupByFrame = asts[6].str().equals("_") ? null : stk.track(asts[6].exec(env)).getFrame();
    AstRoot vals = asts[7];
    AstNumList values;
    if (vals instanceof AstNumList)
        values = (AstNumList) vals;
    else if (vals instanceof AstNum)
        values = new AstNumList(((AstNum) vals).getNum());
    else
        values = null;
    boolean doGrpBy = !by2.isEmpty() || groupByFrame != null;
    // Compute the imputed value per-group.  Empty groups are allowed and OK.
    IcedHashMap<AstGroup.G, Freezable[]> group_impute_map;
    if (!doGrpBy) {
        // Skip the grouping work
        if (ffill0 || bfill0) {
            // do a forward/backward fill on the NA
            // TODO: requires chk.previousNonNA and chk.nextNonNA style methods (which may go across chk boundaries)s
            final boolean ffill = ffill0;
            final boolean bfill = bfill0;
            throw H2O.unimpl("No ffill or bfill imputation supported");
        //        new MRTask() {
        //          @Override public void map(Chunk[] cs) {
        //            int len=cs[0]._len; // end of this chk
        //            long start=cs[0].start();  // absolute beginning of chk s.t. start-1 bleeds into previous chk
        //            long absEnd = start+len;   // absolute end of the chk s.t. absEnd+1 bleeds into next chk
        //            for(int c=0;c<cs.length;++c )
        //              for(int r=0;r<cs[0]._len;++r ) {
        //                if( cs[c].isNA(r) ) {
        //                  if( r > 0 && r < len-1 ) {
        //                    cs[c].set(r,ffill?)
        //                  }
        //                }
        //              }
        //          }
        //        }.doAll(doAllVecs?fr:new Frame(vec));
        //        return new ValNum(Double.NaN);
        } else {
            final double[] res = values == null ? new double[fr.numCols()] : values.expand();
            if (values == null) {
                // fill up res if no values supplied user, common case
                if (doAllVecs) {
                    for (int i = 0; i < res.length; ++i) if (fr.vec(i).isNumeric() || fr.vec(i).isCategorical())
                        res[i] = fr.vec(i).isNumeric() ? fr.vec(i).mean() : ArrayUtils.maxIndex(fr.vec(i).bins());
                } else {
                    Arrays.fill(res, Double.NaN);
                    if (method instanceof AstMean)
                        res[col] = vec.mean();
                    if (method instanceof AstMedian)
                        res[col] = AstMedian.median(new Frame(vec), combine);
                    if (method instanceof AstMode)
                        res[col] = AstMode.mode(vec);
                }
            }
            new MRTask() {

                @Override
                public void map(Chunk[] cs) {
                    int len = cs[0]._len;
                    // run down each chk
                    for (int c = 0; c < cs.length; ++c) if (!Double.isNaN(res[c]))
                        for (int row = 0; row < len; ++row) if (cs[c].isNA(row))
                            cs[c].set(row, res[c]);
                }
            }.doAll(fr);
            return new ValNums(res);
        }
    } else {
        if (col >= fr.numCols())
            throw new IllegalArgumentException("Column not -1 or in range 0 to " + fr.numCols());
        Frame imputes = groupByFrame;
        if (imputes == null) {
            // Build and run a GroupBy command
            AstGroup ast_grp = new AstGroup();
            // simple case where user specified a column... col == -1 means do all columns
            if (doAllVecs) {
                AstRoot[] aggs = new AstRoot[(int) (3 + 3 * (fr.numCols() - by2.cnt()))];
                aggs[0] = ast_grp;
                aggs[1] = new AstFrame(fr);
                aggs[2] = by2;
                int c = 3;
                for (int i = 0; i < fr.numCols(); ++i) {
                    if (!by2.has(i) && (fr.vec(i).isCategorical() || fr.vec(i).isNumeric())) {
                        aggs[c] = fr.vec(i).isNumeric() ? new AstMean() : new AstMode();
                        aggs[c + 1] = new AstNumList(i, i + 1);
                        aggs[c + 2] = new AstStr("rm");
                        c += 3;
                    }
                }
                imputes = ast_grp.apply(env, stk, aggs).getFrame();
            } else
                imputes = ast_grp.apply(env, stk, new AstRoot[] { ast_grp, new AstFrame(fr), by2, /**/
                method, new AstNumList(col, col + 1), new AstStr("rm") }).getFrame();
        }
        if (// >2 makes it ambiguous which columns are groupby cols and which are aggs, throw IAE
        by2.isEmpty() && imputes.numCols() > 2)
            throw new IllegalArgumentException("Ambiguous group-by frame. Supply the `by` columns to proceed.");
        final int[] bycols0 = ArrayUtils.seq(0, Math.max((int) by2.cnt(), 1));
        group_impute_map = new Gather(by2.expand4(), bycols0, fr.numCols(), col).doAll(imputes)._group_impute_map;
        // Now walk over the data, replace NAs with the imputed results
        final IcedHashMap<AstGroup.G, Freezable[]> final_group_impute_map = group_impute_map;
        if (by2.isEmpty()) {
            int[] byCols = new int[imputes.numCols() - 1];
            for (int i = 0; i < byCols.length; ++i) byCols[i] = fr.find(imputes.name(i));
            by2 = new AstNumList(byCols);
        }
        final int[] bycols = by2.expand4();
        new MRTask() {

            @Override
            public void map(Chunk[] cs) {
                Set<Integer> _bycolz = new HashSet<>();
                for (int b : bycols) _bycolz.add(b);
                AstGroup.G g = new AstGroup.G(bycols.length, null);
                for (int row = 0; row < cs[0]._len; row++) for (int c = 0; c < cs.length; ++c) if (!_bycolz.contains(c))
                    if (cs[c].isNA(row))
                        cs[c].set(row, ((IcedDouble) final_group_impute_map.get(g.fill(row, cs, bycols))[c])._val);
            }
        }.doAll(fr);
        return new ValFrame(imputes);
    }
}

Example 64 with Chunk

use of water.fvec.Chunk in project h2o-3 by h2oai.

the class AstTable method fast_table.

// -------------------------------------------------------------------------
// Fast-path for 1 integer column
private ValFrame fast_table(Vec v1, int ncols, String colname) {
    if (ncols != 1 || !v1.isInt())
        return null;
    long spanl = (long) v1.max() - (long) v1.min() + 1;
    // Cap at decent array size, for performance
    if (spanl > 1000000)
        return null;
    // First fast-pass counting
    AstTable.FastCnt fastCnt = new AstTable.FastCnt((long) v1.min(), (int) spanl).doAll(v1);
    final long[] cnts = fastCnt._cnts;
    final long minVal = fastCnt._min;
    // Second pass to build the result frame, skipping zeros
    Vec dataLayoutVec = Vec.makeCon(0, cnts.length);
    Frame fr = new MRTask() {

        @Override
        public void map(Chunk[] cs, NewChunk nc0, NewChunk nc1) {
            final Chunk c = cs[0];
            for (int i = 0; i < c._len; ++i) {
                int idx = (int) (i + c.start());
                if (cnts[idx] > 0) {
                    nc0.addNum(idx + minVal);
                    nc1.addNum(cnts[idx]);
                }
            }
        }
    }.doAll(new byte[] { Vec.T_NUM, Vec.T_NUM }, dataLayoutVec).outputFrame(new String[] { colname, "Count" }, new String[][] { v1.domain(), null });
    dataLayoutVec.remove();
    return new ValFrame(fr);
}

Example 65 with Chunk

use of water.fvec.Chunk in project h2o-3 by h2oai.

the class DataInfoTestAdapt method checkFrame.

private void checkFrame(final DataInfo di, final Frame gold, final boolean skipMissing) {
    try {
        Vec[] vecs = new Vec[di._adaptedFrame.numCols() + gold.numCols()];
        System.arraycopy(di._adaptedFrame.vecs(), 0, vecs, 0, di._adaptedFrame.numCols());
        System.arraycopy(gold.vecs(), 0, vecs, di._adaptedFrame.numCols(), gold.numCols());
        new MRTask() {

            @Override
            public void map(Chunk[] cs) {
                int off = di._adaptedFrame.numCols();
                DataInfo.Row r = di.newDenseRow();
                //          DataInfo.Row rows[] = di.extractSparseRows(cs);
                for (int i = 0; i < cs[0]._len; ++i) {
                    //            DataInfo.Row r = rows[i];
                    di.extractDenseRow(cs, i, r);
                    if (skipMissing && r.isBad())
                        continue;
                    for (int j = 0; j < di.fullN(); ++j) {
                        double goldValue = cs[off + j].atd(i);
                        // - (di._normSub[j - di.numStart()] * di._normMul[j-di.numStart()]);
                        double thisValue = r.get(j);
                        double diff = Math.abs(goldValue - thisValue);
                        if (diff > 1e-12) {
                            if (!skipMissing && diff < 10)
                                System.out.println("row mismatch: " + i + " column= " + j + "; diff= " + diff + " but not skipping missing, so due to discrepancies in taking mean on split frames");
                            else
                                throw new RuntimeException("bonk");
                        }
                    }
                }
            }
        }.doAll(vecs);
    } finally {
        di.dropInteractions();
        di.remove();
    }
}