public StringBuilder toString(StringBuilder sb, String[] fs, long idx) {
Vec vecs[] = vecs();
for (int c = 0; c < fs.length; c++) {
Vec vec = vecs[c];
if (vec.isEnum()) {
String s = "----------";
if (!vec.isNA(idx)) {
int x = (int) vec.at8(idx);
if (x >= 0 && x < vec._domain.length) s = vec._domain[x];
}
sb.append(String.format(fs[c], s));
} else if (vec.isInt()) {
if (vec.isNA(idx)) {
Chunk C = vec.elem2BV(0); // 1st Chunk
int len = C.pformat_len0(); // Printable width
for (int i = 0; i < len; i++) sb.append('-');
} else {
try {
sb.append(String.format(fs[c], vec.at8(idx)));
} catch (IllegalFormatException ife) {
System.out.println("Format: " + fs[c] + " col=" + c + " not for ints");
ife.printStackTrace();
}
}
} else {
sb.append(String.format(fs[c], vec.at(idx)));
if (vec.isNA(idx)) sb.append(' ');
}
sb.append(' '); // Column seperator
}
sb.append('\n');
return sb;
}
@Override
public void map(Chunk[] ix, NewChunk[] ncs) {
final Vec[] vecs = new Vec[_cols.length];
final Vec anyv = _base.anyVec();
final long nrow = anyv.length();
long r = ix[0].at80(0);
int last_ci = anyv.elem2ChunkIdx(r < nrow ? r : 0); // memoize the last chunk index
long last_c0 = anyv._espc[last_ci]; // ... last chunk start
long last_c1 = anyv._espc[last_ci + 1]; // ... last chunk end
Chunk[] last_cs = new Chunk[vecs.length]; // ... last chunks
for (int c = 0; c < _cols.length; c++) {
vecs[c] = _base.vecs()[_cols[c]];
last_cs[c] = vecs[c].elem2BV(last_ci);
}
for (int i = 0; i < ix[0]._len; i++) {
// select one row
r = ix[0].at80(i) - 1; // next row to select
if (r < 0) continue;
if (r >= nrow) {
for (int c = 0; c < vecs.length; c++) ncs[c].addNum(Double.NaN);
} else {
if (r < last_c0 || r >= last_c1) {
last_ci = anyv.elem2ChunkIdx(r);
last_c0 = anyv._espc[last_ci];
last_c1 = anyv._espc[last_ci + 1];
for (int c = 0; c < vecs.length; c++) last_cs[c] = vecs[c].elem2BV(last_ci);
}
for (int c = 0; c < vecs.length; c++) ncs[c].addNum(last_cs[c].at(r));
}
}
}
public void dropInteractions() { // only called to cleanup the InteractionWrappedVecs!
if (_interactions != null) {
Vec[] vecs = _adaptedFrame.remove(_interactionVecs);
for (Vec v : vecs) v.remove();
_interactions = null;
}
}
public Vec replace(int col, Vec nv) {
assert col < _names.length;
Vec rv = vecs()[col];
assert rv.group().equals(nv.group());
_vecs[col] = nv;
_keys[col] = nv._key;
if (DKV.get(nv._key) == null) // If not already in KV, put it there
DKV.put(nv._key, nv);
return rv;
}
/**
* Check that the vectors are all compatible. All Vecs have their content sharded using same
* number of rows per chunk.
*/
public void checkCompatible() {
try {
Vec v0 = anyVec();
int nchunks = v0.nChunks();
for (Vec vec : vecs()) {
if (vec instanceof AppendableVec) continue; // New Vectors are endlessly compatible
if (vec.nChunks() != nchunks)
throw new IllegalArgumentException(
"Vectors different numbers of chunks, " + nchunks + " and " + vec.nChunks());
}
// Also check each chunk has same rows
for (int i = 0; i < nchunks; i++) {
long es = v0.chunk2StartElem(i);
for (Vec vec : vecs())
if (!(vec instanceof AppendableVec) && vec.chunk2StartElem(i) != es)
throw new IllegalArgumentException(
"Vector chunks different numbers of rows, "
+ es
+ " and "
+ vec.chunk2StartElem(i));
}
} catch (Throwable ex) {
Throwables.propagate(ex);
}
}
@Override
protected void setupLocal() {
// Precompute the first input chunk index and start row inside that chunk for this partition
Vec anyInVec = _srcVecs[0];
long[] partSizes = Utils.partitione(anyInVec.length(), _ratios);
long pnrows = 0;
for (int p = 0; p < _partIdx; p++) pnrows += partSizes[p];
long[] espc = anyInVec._espc;
while (_pcidx < espc.length - 1 && (pnrows -= (espc[_pcidx + 1] - espc[_pcidx])) > 0)
_pcidx++;
assert pnrows <= 0;
_psrow = (int) (pnrows + espc[_pcidx + 1] - espc[_pcidx]);
}
// Make vector templates for all output frame vectors
private Vec[][] makeTemplates(Frame dataset, float[] ratios) {
Vec anyVec = dataset.anyVec();
final long[][] espcPerSplit = computeEspcPerSplit(anyVec._espc, anyVec.length(), ratios);
final int num = dataset.numCols(); // number of columns in input frame
final int nsplits = espcPerSplit.length; // number of splits
final String[][] domains = dataset.domains(); // domains
Vec[][] t = new Vec[nsplits][ /*num*/]; // resulting vectors for all
for (int i = 0; i < nsplits; i++) {
// vectors for j-th split
t[i] = new Vec(Vec.newKey(), espcPerSplit[i /*-th split*/]).makeZeros(num, domains);
}
return t;
}
public Frame(String[] names, Vec[] vecs) {
// assert names==null || names.length == vecs.length : "Number of columns does not match to
// number of cols' names.";
_names = names;
_vecs = vecs;
_keys = new Key[vecs.length];
for (int i = 0; i < vecs.length; i++) {
Key k = _keys[i] = vecs[i]._key;
if (DKV.get(k) == null) // If not already in KV, put it there
DKV.put(k, vecs[i]);
}
Vec v0 = anyVec();
if (v0 == null) return;
VectorGroup grp = v0.group();
for (int i = 0; i < vecs.length; i++) assert grp.equals(vecs[i].group());
}
@Override
Val apply(Env env, Env.StackHelp stk, AST asts[]) {
// Compute the variable args. Find the common row count
Val vals[] = new Val[asts.length];
Vec vec = null;
for (int i = 1; i < asts.length; i++) {
vals[i] = stk.track(asts[i].exec(env));
if (vals[i].isFrame()) {
Vec anyvec = vals[i].getFrame().anyVec();
if (anyvec == null) continue; // Ignore the empty frame
if (vec == null) vec = anyvec;
else if (vec.length() != anyvec.length())
throw new IllegalArgumentException(
"cbind frames must have all the same rows, found "
+ vec.length()
+ " and "
+ anyvec.length()
+ " rows.");
}
}
boolean clean = false;
if (vec == null) {
vec = Vec.makeZero(1);
clean = true;
} // Default to length 1
// Populate the new Frame
Frame fr = new Frame();
for (int i = 1; i < asts.length; i++) {
switch (vals[i].type()) {
case Val.FRM:
fr.add(fr.makeCompatible(vals[i].getFrame()));
break;
case Val.FUN:
throw H2O.unimpl();
case Val.STR:
throw H2O.unimpl();
case Val.NUM:
// Auto-expand scalars to fill every row
double d = vals[i].getNum();
fr.add(Double.toString(d), vec.makeCon(d));
break;
default:
throw H2O.unimpl();
}
}
if (clean) vec.remove();
return new ValFrame(fr);
}
/**
* Global redistribution of a Frame (balancing of chunks), done by calling process (all-to-one +
* one-to-all)
*
* @param fr Input frame
* @param seed RNG seed
* @param shuffle whether to shuffle the data globally
* @return Shuffled frame
*/
public static Frame shuffleAndBalance(
final Frame fr, int splits, long seed, final boolean local, final boolean shuffle) {
if ((fr.vecs()[0].nChunks() < splits || shuffle) && fr.numRows() > splits) {
Vec[] vecs = fr.vecs().clone();
Log.info("Load balancing dataset, splitting it into up to " + splits + " chunks.");
long[] idx = null;
if (shuffle) {
idx = new long[splits];
for (int r = 0; r < idx.length; ++r) idx[r] = r;
Utils.shuffleArray(idx, seed);
}
Key keys[] = new Vec.VectorGroup().addVecs(vecs.length);
final long rows_per_new_chunk = (long) (Math.ceil((double) fr.numRows() / splits));
// loop over cols (same indexing for each column)
Futures fs = new Futures();
for (int col = 0; col < vecs.length; col++) {
AppendableVec vec = new AppendableVec(keys[col]);
// create outgoing chunks for this col
NewChunk[] outCkg = new NewChunk[splits];
for (int i = 0; i < splits; ++i) outCkg[i] = new NewChunk(vec, i);
// loop over all incoming chunks
for (int ckg = 0; ckg < vecs[col].nChunks(); ckg++) {
final Chunk inCkg = vecs[col].chunkForChunkIdx(ckg);
// loop over local rows of incoming chunks (fast path)
for (int row = 0; row < inCkg._len; ++row) {
int outCkgIdx =
(int) ((inCkg._start + row) / rows_per_new_chunk); // destination chunk idx
if (shuffle)
outCkgIdx = (int) (idx[outCkgIdx]); // shuffle: choose a different output chunk
assert (outCkgIdx >= 0 && outCkgIdx < splits);
outCkg[outCkgIdx].addNum(inCkg.at0(row));
}
}
for (int i = 0; i < outCkg.length; ++i) outCkg[i].close(i, fs);
Vec t = vec.close(fs);
t._domain = vecs[col]._domain;
vecs[col] = t;
}
fs.blockForPending();
Log.info("Load balancing done.");
return new Frame(fr.names(), vecs);
}
return fr;
}
@Override
public String toString() {
// Across
String s = "{" + _names[0];
long bs = _vecs[0].byteSize();
for (int i = 1; i < _names.length; i++) {
s += "," + _names[i];
bs += _vecs[i].byteSize();
}
s += "}, " + PrettyPrint.bytes(bs) + "\n";
// Down
Vec v0 = firstReadable();
if (v0 == null) return s;
int nc = v0.nChunks();
s += "Chunk starts: {";
for (int i = 0; i < nc; i++) s += v0.elem2BV(i)._start + ",";
s += "}";
return s;
}
/** Appends a named column, keeping the last Vec as the response */
public void add(String name, Vec vec) {
assert _vecs.length == 0 || anyVec().group().equals(vec.group());
final int len = _names.length;
_names = Arrays.copyOf(_names, len + 1);
_vecs = Arrays.copyOf(_vecs, len + 1);
_keys = Arrays.copyOf(_keys, len + 1);
_names[len] = name;
_vecs[len] = vec;
_keys[len] = vec._key;
}
@Override
ValFrame apply(Env env, Env.StackHelp stk, AST asts[]) {
Frame fr = stk.track(asts[1].exec(env)).getFrame();
double frac = asts[2].exec(env).getNum();
double nrow = fr.numRows() * frac;
Vec vecs[] = fr.vecs();
long[] idxs = new long[fr.numCols()];
int j = 0;
for (int i = 0; i < idxs.length; i++) if (vecs[i].naCnt() < nrow) idxs[j++] = i;
Vec vec = Vec.makeVec(Arrays.copyOf(idxs, j), null, Vec.VectorGroup.VG_LEN1.addVec());
return new ValFrame(new Frame(vec));
}
/**
* Compute the L2 norm for each row of the frame
*
* @param fr Input frame
* @return Vec containing L2 values for each row, is in K-V store
*/
public static Vec getL2(final Frame fr, final double[] scale) {
// add workspace vec at end
final int idx = fr.numCols();
assert (scale.length == idx) : "Mismatch for number of columns";
fr.add("L2", fr.anyVec().makeZero());
Vec res;
try {
new MRTask2() {
@Override
public void map(Chunk[] cs) {
for (int r = 0; r < cs[0]._len; r++) {
double norm2 = 0;
for (int i = 0; i < idx; i++) norm2 += Math.pow(cs[i].at0(r) * scale[i], 2);
cs[idx].set0(r, Math.sqrt(norm2));
}
}
}.doAll(fr);
} finally {
res = fr.remove(idx);
}
res.rollupStats();
return res;
}
@Override
public void map(Chunk cs) {
int idx = _chunkOffset + cs.cidx();
Key ckey = Vec.chunkKey(_v._key, idx);
if (_cmap != null) {
assert !cs.hasFloat()
: "Input chunk (" + cs.getClass() + ") has float, but is expected to be categorical";
NewChunk nc = new NewChunk(_v, idx);
// loop over rows and update ints for new domain mapping according to vecs[c].domain()
for (int r = 0; r < cs._len; ++r) {
if (cs.isNA(r)) nc.addNA();
else nc.addNum(_cmap[(int) cs.at8(r)], 0);
}
nc.close(_fs);
} else {
DKV.put(ckey, cs.deepCopy(), _fs, true);
}
}
public static Key makeByteVec(Key k, String... data) {
byte[][] chunks = new byte[data.length][];
long[] espc = new long[data.length + 1];
for (int i = 0; i < chunks.length; ++i) {
chunks[i] = data[i].getBytes();
espc[i + 1] = espc[i] + data[i].length();
}
Futures fs = new Futures();
Key key = Vec.newKey();
ByteVec bv = new ByteVec(key, Vec.ESPC.rowLayout(key, espc));
for (int i = 0; i < chunks.length; ++i) {
Key chunkKey = bv.chunkKey(i);
DKV.put(
chunkKey,
new Value(chunkKey, chunks[i].length, chunks[i], TypeMap.C1NCHUNK, Value.ICE),
fs);
}
DKV.put(bv._key, bv, fs);
Frame fr = new Frame(k, new String[] {"makeByteVec"}, new Vec[] {bv});
DKV.put(k, fr, fs);
fs.blockForPending();
return k;
}
@Override
Val apply(Env env, Env.StackHelp stk, AST asts[]) {
// Execute all args. Find a canonical frame; all Frames must look like this one.
// Each argument turns into either a Frame (whose rows are entirely
// inlined) or a scalar (which is replicated across as a single row).
Frame fr = null; // Canonical Frame; all frames have the same column count, types and names
int nchks = 0; // Total chunks
Val vals[] = new Val[asts.length]; // Computed AST results
for (int i = 1; i < asts.length; i++) {
vals[i] = stk.track(asts[i].exec(env));
if (vals[i].isFrame()) {
fr = vals[i].getFrame();
nchks += fr.anyVec().nChunks(); // Total chunks
} else nchks++; // One chunk per scalar
}
// No Frame, just a pile-o-scalars?
Vec zz = null; // The zero-length vec for the zero-frame frame
if (fr == null) { // Zero-length, 1-column, default name
fr = new Frame(new String[] {Frame.defaultColName(0)}, new Vec[] {zz = Vec.makeZero(0)});
if (asts.length == 1) return new ValFrame(fr);
}
// Verify all Frames are the same columns, names, and types. Domains can vary, and will be the
// union
final Frame frs[] = new Frame[asts.length]; // Input frame
final byte[] types = fr.types(); // Column types
final int ncols = fr.numCols();
final long[] espc = new long[nchks + 1]; // Compute a new layout!
int coffset = 0;
for (int i = 1; i < asts.length; i++) {
Val val = vals[i]; // Save values computed for pass 2
Frame fr0 =
val.isFrame()
? val.getFrame()
// Scalar: auto-expand into a 1-row frame
: stk.track(new Frame(fr._names, Vec.makeCons(val.getNum(), 1L, fr.numCols())));
// Check that all frames are compatible
if (fr.numCols() != fr0.numCols())
throw new IllegalArgumentException(
"rbind frames must have all the same columns, found "
+ fr.numCols()
+ " and "
+ fr0.numCols()
+ " columns.");
if (!Arrays.deepEquals(fr._names, fr0._names))
throw new IllegalArgumentException(
"rbind frames must have all the same column names, found "
+ Arrays.toString(fr._names)
+ " and "
+ Arrays.toString(fr0._names));
if (!Arrays.equals(types, fr0.types()))
throw new IllegalArgumentException(
"rbind frames must have all the same column types, found "
+ Arrays.toString(types)
+ " and "
+ Arrays.toString(fr0.types()));
frs[i] = fr0; // Save frame
// Roll up the ESPC row counts
long roffset = espc[coffset];
long[] espc2 = fr0.anyVec().espc();
for (int j = 1; j < espc2.length; j++) // Roll up the row counts
espc[coffset + j] = (roffset + espc2[j]);
coffset += espc2.length - 1; // Chunk offset
}
if (zz != null) zz.remove();
// build up the new domains for each vec
HashMap<String, Integer>[] dmap = new HashMap[types.length];
String[][] domains = new String[types.length][];
int[][][] cmaps = new int[types.length][][];
for (int k = 0; k < types.length; ++k) {
dmap[k] = new HashMap<>();
int c = 0;
byte t = types[k];
if (t == Vec.T_CAT) {
int[][] maps = new int[frs.length][];
for (int i = 1; i < frs.length; i++) {
maps[i] = new int[frs[i].vec(k).domain().length];
for (int j = 0; j < maps[i].length; j++) {
String s = frs[i].vec(k).domain()[j];
if (!dmap[k].containsKey(s)) dmap[k].put(s, maps[i][j] = c++);
else maps[i][j] = dmap[k].get(s);
}
}
cmaps[k] = maps;
} else {
cmaps[k] = new int[frs.length][];
}
domains[k] = c == 0 ? null : new String[c];
for (Map.Entry<String, Integer> e : dmap[k].entrySet()) domains[k][e.getValue()] = e.getKey();
}
// Now make Keys for the new Vecs
Key<Vec>[] keys = fr.anyVec().group().addVecs(fr.numCols());
Vec[] vecs = new Vec[fr.numCols()];
int rowLayout = Vec.ESPC.rowLayout(keys[0], espc);
for (int i = 0; i < vecs.length; i++)
vecs[i] = new Vec(keys[i], rowLayout, domains[i], types[i]);
// Do the row-binds column-by-column.
// Switch to F/J thread for continuations
ParallelRbinds t;
H2O.submitTask(t = new ParallelRbinds(frs, espc, vecs, cmaps)).join();
return new ValFrame(new Frame(fr.names(), t._vecs));
}
/**
* The train/valid Frame instances are sorted by categorical (themselves sorted by cardinality
* greatest to least) with all numerical columns following. The response column(s) are placed at
* the end.
*
* <p>Interactions: 1. Num-Num (Note: N(0,1) * N(0,1) ~ N(0,1) ) 2. Num-Enum 3. Enum-Enum
*
* <p>Interactions are produced on the fly and are dense (in all 3 cases). Consumers of DataInfo
* should not have to care how these interactions are generated. Any heuristic using the fullN
* value should continue functioning the same.
*
* <p>Interactions are specified in two ways: A. As a list of pairs of column indices. B. As a
* list of pairs of column indices with limited enums.
*/
public DataInfo(
Frame train,
Frame valid,
int nResponses,
boolean useAllFactorLevels,
TransformType predictor_transform,
TransformType response_transform,
boolean skipMissing,
boolean imputeMissing,
boolean missingBucket,
boolean weight,
boolean offset,
boolean fold,
Model.InteractionPair[] interactions) {
super(Key.<DataInfo>make());
_valid = valid != null;
assert predictor_transform != null;
assert response_transform != null;
_offset = offset;
_weights = weight;
_fold = fold;
assert !(skipMissing && imputeMissing) : "skipMissing and imputeMissing cannot both be true";
_skipMissing = skipMissing;
_imputeMissing = imputeMissing;
_predictor_transform = predictor_transform;
_response_transform = response_transform;
_responses = nResponses;
_useAllFactorLevels = useAllFactorLevels;
_interactions = interactions;
// create dummy InteractionWrappedVecs and shove them onto the front
if (_interactions != null) {
_interactionVecs = new int[_interactions.length];
train =
Model.makeInteractions(
train,
false,
_interactions,
_useAllFactorLevels,
_skipMissing,
predictor_transform == TransformType.STANDARDIZE)
.add(train);
if (valid != null)
valid =
Model.makeInteractions(
valid,
true,
_interactions,
_useAllFactorLevels,
_skipMissing,
predictor_transform == TransformType.STANDARDIZE)
.add(valid); // FIXME: should be using the training subs/muls!
}
_permutation = new int[train.numCols()];
final Vec[] tvecs = train.vecs();
// Count categorical-vs-numerical
final int n = tvecs.length - _responses - (offset ? 1 : 0) - (weight ? 1 : 0) - (fold ? 1 : 0);
int[] nums = MemoryManager.malloc4(n);
int[] cats = MemoryManager.malloc4(n);
int nnums = 0, ncats = 0;
for (int i = 0; i < n; ++i)
if (tvecs[i].isCategorical()) cats[ncats++] = i;
else nums[nnums++] = i;
_nums = nnums;
_cats = ncats;
_catLvls = new int[ncats][];
// sort the cats in the decreasing order according to their size
for (int i = 0; i < ncats; ++i)
for (int j = i + 1; j < ncats; ++j)
if (tvecs[cats[i]].domain().length < tvecs[cats[j]].domain().length) {
int x = cats[i];
cats[i] = cats[j];
cats[j] = x;
}
String[] names = new String[train.numCols()];
Vec[] tvecs2 = new Vec[train.numCols()];
// Compute the cardinality of each cat
_catModes = new int[ncats];
_catOffsets = MemoryManager.malloc4(ncats + 1);
_catMissing = new boolean[ncats];
int len = _catOffsets[0] = 0;
int interactionIdx = 0; // simple index into the _interactionVecs array
ArrayList<Integer> interactionIds;
if (_interactions == null) {
interactionIds = new ArrayList<>();
for (int i = 0; i < tvecs.length; ++i)
if (tvecs[i] instanceof InteractionWrappedVec) {
interactionIds.add(i);
}
_interactionVecs = new int[interactionIds.size()];
for (int i = 0; i < _interactionVecs.length; ++i) _interactionVecs[i] = interactionIds.get(i);
}
for (int i = 0; i < ncats; ++i) {
names[i] = train._names[cats[i]];
Vec v = (tvecs2[i] = tvecs[cats[i]]);
_catMissing[i] = missingBucket; // needed for test time
if (v instanceof InteractionWrappedVec) {
if (_interactions != null) _interactions[interactionIdx].vecIdx = i;
_interactionVecs[interactionIdx++] =
i; // i (and not cats[i]) because this is the index in _adaptedFrame
_catOffsets[i + 1] = (len += v.domain().length + (missingBucket ? 1 : 0));
} else
_catOffsets[i + 1] =
(len +=
v.domain().length
- (useAllFactorLevels ? 0 : 1)
+ (missingBucket ? 1 : 0)); // missing values turn into a new factor level
_catModes[i] =
imputeMissing ? imputeCat(train.vec(cats[i])) : _catMissing[i] ? v.domain().length : -100;
_permutation[i] = cats[i];
}
_numMeans = new double[nnums];
_numOffsets = MemoryManager.malloc4(nnums + 1);
_numOffsets[0] = len;
boolean isIWV; // is InteractionWrappedVec?
for (int i = 0; i < nnums; ++i) {
names[i + ncats] = train._names[nums[i]];
Vec v = train.vec(nums[i]);
tvecs2[i + ncats] = v;
isIWV = v instanceof InteractionWrappedVec;
if (isIWV) {
if (null != _interactions) _interactions[interactionIdx].vecIdx = i + ncats;
_interactionVecs[interactionIdx++] = i + ncats;
}
_numOffsets[i + 1] = (len += (isIWV ? ((InteractionWrappedVec) v).expandedLength() : 1));
_numMeans[i] = train.vec(nums[i]).mean();
_permutation[i + ncats] = nums[i];
}
for (int i = names.length - nResponses - (weight ? 1 : 0) - (offset ? 1 : 0) - (fold ? 1 : 0);
i < names.length;
++i) {
names[i] = train._names[i];
tvecs2[i] = train.vec(i);
}
_adaptedFrame = new Frame(names, tvecs2);
train.restructure(names, tvecs2);
if (valid != null) valid.restructure(names, valid.vecs(names));
// _adaptedFrame = train;
setPredictorTransform(predictor_transform);
if (_responses > 0) setResponseTransform(response_transform);
}
public static int imputeCat(Vec v) {
if (v.isCategorical()) return v.mode();
return (int) Math.round(v.mean());
}
public Frame deepSlice(Object orows, Object ocols) {
// ocols is either a long[] or a Frame-of-1-Vec
long[] cols;
if (ocols == null) {
cols = (long[]) ocols;
assert cols == null;
} else {
if (ocols instanceof long[]) {
cols = (long[]) ocols;
} else if (ocols instanceof Frame) {
Frame fr = (Frame) ocols;
if (fr.numCols() != 1) {
throw new IllegalArgumentException(
"Columns Frame must have only one column (actually has "
+ fr.numCols()
+ " columns)");
}
long n = fr.anyVec().length();
if (n > MAX_EQ2_COLS) {
throw new IllegalArgumentException(
"Too many requested columns (requested " + n + ", max " + MAX_EQ2_COLS + ")");
}
cols = new long[(int) n];
Vec v = fr._vecs[0];
for (long i = 0; i < v.length(); i++) {
cols[(int) i] = v.at8(i);
}
} else {
throw new IllegalArgumentException(
"Columns is specified by an unsupported data type ("
+ ocols.getClass().getName()
+ ")");
}
}
// Since cols is probably short convert to a positive list.
int c2[] = null;
if (cols == null) {
c2 = new int[numCols()];
for (int i = 0; i < c2.length; i++) c2[i] = i;
} else if (cols.length == 0) {
c2 = new int[0];
} else if (cols[0] > 0) {
c2 = new int[cols.length];
for (int i = 0; i < cols.length; i++)
c2[i] = (int) cols[i] - 1; // Convert 1-based cols to zero-based
} else {
c2 = new int[numCols() - cols.length];
int j = 0;
for (int i = 0; i < numCols(); i++) {
if (j >= cols.length || i < (-cols[j] - 1)) c2[i - j] = i;
else j++;
}
}
for (int i = 0; i < c2.length; i++)
if (c2[i] >= numCols())
throw new IllegalArgumentException(
"Trying to select column " + c2[i] + " but only " + numCols() + " present.");
if (c2.length == 0)
throw new IllegalArgumentException(
"No columns selected (did you try to select column 0 instead of column 1?)");
// Do Da Slice
// orows is either a long[] or a Vec
if (orows == null)
return new DeepSlice((long[]) orows, c2)
.doAll(c2.length, this)
.outputFrame(names(c2), domains(c2));
else if (orows instanceof long[]) {
final long CHK_ROWS = 1000000;
long[] rows = (long[]) orows;
if (rows.length == 0)
return new DeepSlice(rows, c2).doAll(c2.length, this).outputFrame(names(c2), domains(c2));
if (rows[0] < 0)
return new DeepSlice(rows, c2).doAll(c2.length, this).outputFrame(names(c2), domains(c2));
// Vec'ize the index array
AppendableVec av = new AppendableVec("rownames");
int r = 0;
int c = 0;
while (r < rows.length) {
NewChunk nc = new NewChunk(av, c);
long end = Math.min(r + CHK_ROWS, rows.length);
for (; r < end; r++) {
nc.addNum(rows[r]);
}
nc.close(c++, null);
}
Vec c0 = av.close(null); // c0 is the row index vec
Frame fr2 =
new Slice(c2, this)
.doAll(c2.length, new Frame(new String[] {"rownames"}, new Vec[] {c0}))
.outputFrame(names(c2), domains(c2));
UKV.remove(c0._key); // Remove hidden vector
return fr2;
}
Frame frows = (Frame) orows;
Vec vrows = frows.anyVec();
// It's a compatible Vec; use it as boolean selector.
// Build column names for the result.
Vec[] vecs = new Vec[c2.length + 1];
String[] names = new String[c2.length + 1];
for (int i = 0; i < c2.length; ++i) {
vecs[i] = _vecs[c2[i]];
names[i] = _names[c2[i]];
}
vecs[c2.length] = vrows;
names[c2.length] = "predicate";
return new DeepSelect()
.doAll(c2.length, new Frame(names, vecs))
.outputFrame(names(c2), domains(c2));
}
private void setTransform(
TransformType t, double[] normMul, double[] normSub, int vecStart, int n) {
int idx = 0; // idx!=i when interactions are in play, otherwise, it's just 'i'
for (int i = 0; i < n; ++i) {
Vec v = _adaptedFrame.vec(vecStart + i);
boolean isIWV = isInteractionVec(vecStart + i);
switch (t) {
case STANDARDIZE:
normMul[idx] = (v.sigma() != 0) ? 1.0 / v.sigma() : 1.0;
if (isIWV) for (int j = idx + 1; j < nextNumericIdx(i) + idx; j++) normMul[j] = 1;
normSub[idx] = v.mean();
break;
case NORMALIZE:
normMul[idx] = (v.max() - v.min() > 0) ? 1.0 / (v.max() - v.min()) : 1.0;
if (isIWV) for (int j = idx + 1; j < nextNumericIdx(i) + idx; j++) normMul[j] = 1;
normSub[idx] = v.mean();
break;
case DEMEAN:
normMul[idx] = 1;
if (isIWV) for (int j = idx + 1; j < nextNumericIdx(i) + idx; j++) normMul[j] = 1;
normSub[idx] = v.mean();
break;
case DESCALE:
normMul[idx] = (v.sigma() != 0) ? 1.0 / v.sigma() : 1.0;
if (isIWV) for (int j = idx + 1; j < nextNumericIdx(i) + idx; j++) normMul[j] = 1;
normSub[idx] = 0;
break;
default:
throw H2O.unimpl();
}
assert !Double.isNaN(normMul[idx]);
assert !Double.isNaN(normSub[idx]);
idx = isIWV ? (idx + nextNumericIdx(i)) : (idx + 1);
}
}
public ClassDist(final Vec label) {
super(label.domain().length);
}
/** Returns the first readable vector. */
public Vec anyVec() {
if (_col0 != null) return _col0;
for (Vec v : vecs()) if (v.readable()) return (_col0 = v);
return null;
}
/**
* Stratified sampling for classifiers
*
* @param fr Input frame
* @param label Label vector (must be enum)
* @param sampling_ratios Optional: array containing the requested sampling ratios per class (in
* order of domains), will be overwritten if it contains all 0s
* @param maxrows Maximum number of rows in the returned frame
* @param seed RNG seed for sampling
* @param allowOversampling Allow oversampling of minority classes
* @param verbose Whether to print verbose info
* @return Sampled frame, with approximately the same number of samples from each class (or given
* by the requested sampling ratios)
*/
public static Frame sampleFrameStratified(
final Frame fr,
Vec label,
float[] sampling_ratios,
long maxrows,
final long seed,
final boolean allowOversampling,
final boolean verbose) {
if (fr == null) return null;
assert (label.isEnum());
assert (maxrows >= label.domain().length);
long[] dist = new ClassDist(label).doAll(label).dist();
assert (dist.length > 0);
Log.info(
"Doing stratified sampling for data set containing "
+ fr.numRows()
+ " rows from "
+ dist.length
+ " classes. Oversampling: "
+ (allowOversampling ? "on" : "off"));
if (verbose) {
for (int i = 0; i < dist.length; ++i) {
Log.info(
"Class "
+ label.domain(i)
+ ": count: "
+ dist[i]
+ " prior: "
+ (float) dist[i] / fr.numRows());
}
}
// create sampling_ratios for class balance with max. maxrows rows (fill existing array if not
// null)
if (sampling_ratios == null
|| (Utils.minValue(sampling_ratios) == 0 && Utils.maxValue(sampling_ratios) == 0)) {
// compute sampling ratios to achieve class balance
if (sampling_ratios == null) {
sampling_ratios = new float[dist.length];
}
assert (sampling_ratios.length == dist.length);
for (int i = 0; i < dist.length; ++i) {
sampling_ratios[i] =
((float) fr.numRows() / label.domain().length) / dist[i]; // prior^-1 / num_classes
}
final float inv_scale =
Utils.minValue(
sampling_ratios); // majority class has lowest required oversampling factor to achieve
// balance
if (!Float.isNaN(inv_scale) && !Float.isInfinite(inv_scale))
Utils.div(
sampling_ratios,
inv_scale); // want sampling_ratio 1.0 for majority class (no downsampling)
}
if (!allowOversampling) {
for (int i = 0; i < sampling_ratios.length; ++i) {
sampling_ratios[i] = Math.min(1.0f, sampling_ratios[i]);
}
}
// given these sampling ratios, and the original class distribution, this is the expected number
// of resulting rows
float numrows = 0;
for (int i = 0; i < sampling_ratios.length; ++i) {
numrows += sampling_ratios[i] * dist[i];
}
final long actualnumrows = Math.min(maxrows, Math.round(numrows)); // cap #rows at maxrows
assert (actualnumrows
>= 0); // can have no matching rows in case of sparse data where we had to fill in a
// makeZero() vector
Log.info("Stratified sampling to a total of " + String.format("%,d", actualnumrows) + " rows.");
if (actualnumrows != numrows) {
Utils.mult(
sampling_ratios,
(float) actualnumrows
/ numrows); // adjust the sampling_ratios by the global rescaling factor
if (verbose)
Log.info(
"Downsampling majority class by "
+ (float) actualnumrows / numrows
+ " to limit number of rows to "
+ String.format("%,d", maxrows));
}
Log.info(
"Majority class ("
+ label.domain()[Utils.minIndex(sampling_ratios)].toString()
+ ") sampling ratio: "
+ Utils.minValue(sampling_ratios));
Log.info(
"Minority class ("
+ label.domain()[Utils.maxIndex(sampling_ratios)].toString()
+ ") sampling ratio: "
+ Utils.maxValue(sampling_ratios));
return sampleFrameStratified(fr, label, sampling_ratios, seed, verbose);
}
public int find(Vec vec) {
for (int i = 0; i < _vecs.length; i++) if (vec.equals(_vecs[i])) return i;
return -1;
}
/** Returns the first readable vector. */
public Vec firstReadable() {
if (_col0 != null) return _col0;
for (Vec v : _vecs) if (v != null && v.readable()) return (_col0 = v);
return null;
}
// internal version with repeat counter
// currently hardcoded to do up to 10 tries to get a row from each class, which can be impossible
// for certain wrong sampling ratios
private static Frame sampleFrameStratified(
final Frame fr,
Vec label,
final float[] sampling_ratios,
final long seed,
final boolean debug,
int count) {
if (fr == null) return null;
assert (label.isEnum());
assert (sampling_ratios != null && sampling_ratios.length == label.domain().length);
final int labelidx = fr.find(label); // which column is the label?
assert (labelidx >= 0);
final boolean poisson = false; // beta feature
Frame r =
new MRTask2() {
@Override
public void map(Chunk[] cs, NewChunk[] ncs) {
final Random rng = getDeterRNG(seed + cs[0].cidx());
for (int r = 0; r < cs[0]._len; r++) {
if (cs[labelidx].isNA0(r)) continue; // skip missing labels
final int label = (int) cs[labelidx].at80(r);
assert (sampling_ratios.length > label && label >= 0);
int sampling_reps;
if (poisson) {
sampling_reps = Utils.getPoisson(sampling_ratios[label], rng);
} else {
final float remainder = sampling_ratios[label] - (int) sampling_ratios[label];
sampling_reps =
(int) sampling_ratios[label] + (rng.nextFloat() < remainder ? 1 : 0);
}
for (int i = 0; i < ncs.length; i++) {
for (int j = 0; j < sampling_reps; ++j) {
ncs[i].addNum(cs[i].at0(r));
}
}
}
}
}.doAll(fr.numCols(), fr).outputFrame(fr.names(), fr.domains());
// Confirm the validity of the distribution
long[] dist = new ClassDist(r.vecs()[labelidx]).doAll(r.vecs()[labelidx]).dist();
// if there are no training labels in the test set, then there is no point in sampling the test
// set
if (dist == null) return fr;
if (debug) {
long sumdist = Utils.sum(dist);
Log.info("After stratified sampling: " + sumdist + " rows.");
for (int i = 0; i < dist.length; ++i) {
Log.info(
"Class "
+ r.vecs()[labelidx].domain(i)
+ ": count: "
+ dist[i]
+ " sampling ratio: "
+ sampling_ratios[i]
+ " actual relative frequency: "
+ (float) dist[i] / sumdist * dist.length);
}
}
// Re-try if we didn't get at least one example from each class
if (Utils.minValue(dist) == 0 && count < 10) {
Log.info(
"Re-doing stratified sampling because not all classes were represented (unlucky draw).");
r.delete();
return sampleFrameStratified(fr, label, sampling_ratios, seed + 1, debug, ++count);
}
// shuffle intra-chunk
Frame shuffled = shuffleFramePerChunk(r, seed + 0x580FF13);
r.delete();
return shuffled;
}
