@Override
protected Frame predictScoreImpl(Frame orig, Frame adaptedFr, String destination_key) {
return reconstruct(
orig,
adaptedFr,
(null == destination_key ? Key.make() : Key.make(destination_key)),
true,
_parms._impute_original);
}
/**
* Project each archetype into original feature space
*
* @param frame Original training data with m rows and n columns
* @param destination_key Frame Id for output
* @return Frame containing k rows and n columns, where each row corresponds to an archetype
*/
public Frame scoreArchetypes(Frame frame, Key destination_key, boolean reverse_transform) {
final int ncols = _output._names.length;
Frame adaptedFr = new Frame(frame);
adaptTestForTrain(adaptedFr, true, false);
assert ncols == adaptedFr.numCols();
String[][] adaptedDomme = adaptedFr.domains();
double[][] proj = new double[_parms._k][_output._nnums + _output._ncats];
// Categorical columns
for (int d = 0; d < _output._ncats; d++) {
double[][] block = _output._archetypes_raw.getCatBlock(d);
for (int k = 0; k < _parms._k; k++)
proj[k][_output._permutation[d]] = _parms.mimpute(block[k], _output._lossFunc[d]);
}
// Numeric columns
for (int d = _output._ncats; d < (_output._ncats + _output._nnums); d++) {
int ds = d - _output._ncats;
for (int k = 0; k < _parms._k; k++) {
double num = _output._archetypes_raw.getNum(ds, k);
proj[k][_output._permutation[d]] = _parms.impute(num, _output._lossFunc[d]);
if (reverse_transform)
proj[k][_output._permutation[d]] =
proj[k][_output._permutation[d]] / _output._normMul[ds] + _output._normSub[ds];
}
}
// Convert projection of archetypes into a frame with correct domains
Frame f =
ArrayUtils.frame(
(null == destination_key ? Key.make() : destination_key), adaptedFr.names(), proj);
for (int i = 0; i < ncols; i++) f.vec(i).setDomain(adaptedDomme[i]);
return f;
}
@Override
protected Frame rebalance(final Frame original_fr, boolean local, final String name) {
if (original_fr == null) return null;
if (_parms._force_load_balance) {
int original_chunks = original_fr.anyVec().nChunks();
_job.update(0, "Load balancing " + name.substring(name.length() - 5) + " data...");
int chunks = desiredChunks(original_fr, local);
if (!_parms._reproducible) {
if (original_chunks >= chunks) {
if (!_parms._quiet_mode)
Log.info(
"Dataset already contains " + original_chunks + " chunks. No need to rebalance.");
return original_fr;
}
} else { // reproducible, set chunks to 1
assert chunks == 1;
if (!_parms._quiet_mode)
Log.warn("Reproducibility enforced - using only 1 thread - can be slow.");
if (original_chunks == 1) return original_fr;
}
if (!_parms._quiet_mode)
Log.info(
"Rebalancing "
+ name.substring(name.length() - 5)
+ " dataset into "
+ chunks
+ " chunks.");
Key newKey = Key.make(name + ".chks" + chunks);
RebalanceDataSet rb = new RebalanceDataSet(original_fr, newKey, chunks);
H2O.submitTask(rb).join();
Frame rebalanced_fr = DKV.get(newKey).get();
Scope.track(rebalanced_fr);
return rebalanced_fr;
}
return original_fr;
}
// GLRM scoring is data imputation based on feature domains using reconstructed XY (see Udell
// (2015), Section 5.3)
private Frame reconstruct(
Frame orig,
Frame adaptedFr,
Key destination_key,
boolean save_imputed,
boolean reverse_transform) {
final int ncols = _output._names.length;
assert ncols == adaptedFr.numCols();
String prefix = "reconstr_";
// Need [A,X,P] where A = adaptedFr, X = loading frame, P = imputed frame
// Note: A is adapted to original training frame, P has columns shuffled so cats come before
// nums!
Frame fullFrm = new Frame(adaptedFr);
Frame loadingFrm = DKV.get(_output._representation_key).get();
fullFrm.add(loadingFrm);
String[][] adaptedDomme = adaptedFr.domains();
for (int i = 0; i < ncols; i++) {
Vec v = fullFrm.anyVec().makeZero();
v.setDomain(adaptedDomme[i]);
fullFrm.add(prefix + _output._names[i], v);
}
GLRMScore gs = new GLRMScore(ncols, _parms._k, save_imputed, reverse_transform).doAll(fullFrm);
// Return the imputed training frame
int x = ncols + _parms._k, y = fullFrm.numCols();
Frame f =
fullFrm.extractFrame(
x, y); // this will call vec_impl() and we cannot call the delete() below just yet
f = new Frame((null == destination_key ? Key.make() : destination_key), f.names(), f.vecs());
DKV.put(f);
gs._mb.makeModelMetrics(
GLRMModel.this, orig, null, null); // save error metrics based on imputed data
return f;
}
/**
* Train a Deep Learning neural net model
*
* @param model Input model (e.g., from initModel(), or from a previous training run)
* @return Trained model
*/
public final DeepLearningModel trainModel(DeepLearningModel model) {
Frame validScoreFrame = null;
Frame train, trainScoreFrame;
try {
// if (checkpoint == null && !quiet_mode) logStart(); //if checkpoint is given, some
// Job's params might be uninitialized (but the restarted model's parameters are correct)
if (model == null) {
model = DKV.get(dest()).get();
}
Log.info(
"Model category: "
+ (_parms._autoencoder
? "Auto-Encoder"
: isClassifier() ? "Classification" : "Regression"));
final long model_size = model.model_info().size();
Log.info(
"Number of model parameters (weights/biases): " + String.format("%,d", model_size));
model.write_lock(_job);
_job.update(0, "Setting up training data...");
final DeepLearningParameters mp = model.model_info().get_params();
// temporary frames of the same "name" as the orig _train/_valid (asking the parameter's
// Key, not the actual frame)
// Note: don't put into DKV or they would overwrite the _train/_valid frames!
Frame tra_fr = new Frame(mp._train, _train.names(), _train.vecs());
Frame val_fr = _valid != null ? new Frame(mp._valid, _valid.names(), _valid.vecs()) : null;
train = tra_fr;
if (model._output.isClassifier() && mp._balance_classes) {
_job.update(0, "Balancing class distribution of training data...");
float[] trainSamplingFactors =
new float
[train
.lastVec()
.domain()
.length]; // leave initialized to 0 -> will be filled up below
if (mp._class_sampling_factors != null) {
if (mp._class_sampling_factors.length != train.lastVec().domain().length)
throw new IllegalArgumentException(
"class_sampling_factors must have "
+ train.lastVec().domain().length
+ " elements");
trainSamplingFactors =
mp._class_sampling_factors.clone(); // clone: don't modify the original
}
train =
sampleFrameStratified(
train,
train.lastVec(),
train.vec(model._output.weightsName()),
trainSamplingFactors,
(long) (mp._max_after_balance_size * train.numRows()),
mp._seed,
true,
false);
Vec l = train.lastVec();
Vec w = train.vec(model._output.weightsName());
MRUtils.ClassDist cd = new MRUtils.ClassDist(l);
model._output._modelClassDist =
_weights != null ? cd.doAll(l, w).rel_dist() : cd.doAll(l).rel_dist();
}
model.training_rows = train.numRows();
if (_weights != null && _weights.min() == 0 && _weights.max() == 1 && _weights.isInt()) {
model.training_rows = Math.round(train.numRows() * _weights.mean());
Log.warn(
"Not counting "
+ (train.numRows() - model.training_rows)
+ " rows with weight=0 towards an epoch.");
}
Log.info("One epoch corresponds to " + model.training_rows + " training data rows.");
trainScoreFrame =
sampleFrame(
train,
mp._score_training_samples,
mp._seed); // training scoring dataset is always sampled uniformly from the training
// dataset
if (trainScoreFrame != train) Scope.track(trainScoreFrame);
if (!_parms._quiet_mode)
Log.info("Number of chunks of the training data: " + train.anyVec().nChunks());
if (val_fr != null) {
model.validation_rows = val_fr.numRows();
// validation scoring dataset can be sampled in multiple ways from the given validation
// dataset
if (model._output.isClassifier()
&& mp._balance_classes
&& mp._score_validation_sampling
== DeepLearningParameters.ClassSamplingMethod.Stratified) {
_job.update(0, "Sampling validation data (stratified)...");
validScoreFrame =
sampleFrameStratified(
val_fr,
val_fr.lastVec(),
val_fr.vec(model._output.weightsName()),
null,
mp._score_validation_samples > 0
? mp._score_validation_samples
: val_fr.numRows(),
mp._seed + 1,
false /* no oversampling */,
false);
} else {
_job.update(0, "Sampling validation data...");
validScoreFrame = sampleFrame(val_fr, mp._score_validation_samples, mp._seed + 1);
if (validScoreFrame != val_fr) Scope.track(validScoreFrame);
}
if (!_parms._quiet_mode)
Log.info(
"Number of chunks of the validation data: " + validScoreFrame.anyVec().nChunks());
}
// Set train_samples_per_iteration size (cannot be done earlier since this depends on
// whether stratified sampling is done)
model.actual_train_samples_per_iteration =
computeTrainSamplesPerIteration(mp, model.training_rows, model);
// Determine whether shuffling is enforced
if (mp._replicate_training_data
&& (model.actual_train_samples_per_iteration
== model.training_rows * (mp._single_node_mode ? 1 : H2O.CLOUD.size()))
&& !mp._shuffle_training_data
&& H2O.CLOUD.size() > 1
&& !mp._reproducible) {
if (!mp._quiet_mode)
Log.info(
"Enabling training data shuffling, because all nodes train on the full dataset (replicated training data).");
mp._shuffle_training_data = true;
}
if (!mp._shuffle_training_data
&& model.actual_train_samples_per_iteration == model.training_rows
&& train.anyVec().nChunks() == 1) {
if (!mp._quiet_mode)
Log.info(
"Enabling training data shuffling to avoid training rows in the same order over and over (no Hogwild since there's only 1 chunk).");
mp._shuffle_training_data = true;
}
// if (!mp._quiet_mode) Log.info("Initial model:\n" + model.model_info());
long now = System.currentTimeMillis();
model._timeLastIterationEnter = now;
if (_parms._autoencoder) {
_job.update(0, "Scoring null model of autoencoder...");
if (!mp._quiet_mode) Log.info("Scoring the null model of the autoencoder.");
model.doScoring(
trainScoreFrame,
validScoreFrame,
_job._key,
0,
false); // get the null model reconstruction error
}
// put the initial version of the model into DKV
model.update(_job);
model.total_setup_time_ms += now - _job.start_time();
Log.info("Total setup time: " + PrettyPrint.msecs(model.total_setup_time_ms, true));
Log.info("Starting to train the Deep Learning model.");
_job.update(0, "Training...");
// main loop
for (; ; ) {
model.iterations++;
model.set_model_info(
mp._epochs == 0
? model.model_info()
: H2O.CLOUD.size() > 1 && mp._replicate_training_data
? (mp._single_node_mode
? new DeepLearningTask2(
_job._key,
train,
model.model_info(),
rowFraction(train, mp, model),
model.iterations)
.doAll(Key.make(H2O.SELF))
.model_info()
: // replicated data + single node mode
new DeepLearningTask2(
_job._key,
train,
model.model_info(),
rowFraction(train, mp, model),
model.iterations)
.doAllNodes()
.model_info())
: // replicated data + multi-node mode
new DeepLearningTask(
_job._key,
model.model_info(),
rowFraction(train, mp, model),
model.iterations)
.doAll(train)
.model_info()); // distributed data (always in multi-node mode)
if (stop_requested() && !timeout()) break; // cancellation
if (!model.doScoring(
trainScoreFrame, validScoreFrame, _job._key, model.iterations, false))
break; // finished training (or early stopping or convergence)
if (timeout()) break; // stop after scoring
}
// replace the model with the best model so far (if it's better)
if (!stop_requested()
&& _parms._overwrite_with_best_model
&& model.actual_best_model_key != null
&& _parms._nfolds == 0) {
DeepLearningModel best_model = DKV.getGet(model.actual_best_model_key);
if (best_model != null
&& best_model.loss() < model.loss()
&& Arrays.equals(best_model.model_info().units, model.model_info().units)) {
if (!_parms._quiet_mode)
Log.info("Setting the model to be the best model so far (based on scoring history).");
DeepLearningModelInfo mi = best_model.model_info().deep_clone();
// Don't cheat - count full amount of training samples, since that's the amount of
// training it took to train (without finding anything better)
mi.set_processed_global(model.model_info().get_processed_global());
mi.set_processed_local(model.model_info().get_processed_local());
model.set_model_info(mi);
model.update(_job);
model.doScoring(trainScoreFrame, validScoreFrame, _job._key, model.iterations, true);
assert (best_model.loss() == model.loss());
}
}
// store coefficient names for future use
// possibly change
model.model_info().data_info().coefNames();
if (!_parms._quiet_mode) {
Log.info(
"==============================================================================================================================================================================");
if (stop_requested()) {
Log.info("Deep Learning model training was interrupted.");
} else {
Log.info("Finished training the Deep Learning model.");
Log.info(model);
}
Log.info(
"==============================================================================================================================================================================");
}
} finally {
if (model != null) {
model.deleteElasticAverageModels();
model.unlock(_job);
if (model.actual_best_model_key != null) {
assert (model.actual_best_model_key != model._key);
DKV.remove(model.actual_best_model_key);
}
}
}
return model;
}
@Override
protected void compute2() {
CoxPHModel model = null;
try {
Scope.enter();
_parms.read_lock_frames(CoxPH.this);
init(true);
applyScoringFrameSideEffects();
// The model to be built
model = new CoxPHModel(dest(), _parms, new CoxPHModel.CoxPHOutput(CoxPH.this));
model.delete_and_lock(_key);
applyTrainingFrameSideEffects();
int nResponses = 1;
boolean useAllFactorLevels = false;
final DataInfo dinfo =
new DataInfo(
Key.make(),
_modelBuilderTrain,
null,
nResponses,
useAllFactorLevels,
DataInfo.TransformType.DEMEAN,
TransformType.NONE,
true,
false,
false,
false,
false,
false);
initStats(model, dinfo);
final int n_offsets =
(model._parms.offset_columns == null) ? 0 : model._parms.offset_columns.length;
final int n_coef = dinfo.fullN() - n_offsets;
final double[] step = MemoryManager.malloc8d(n_coef);
final double[] oldCoef = MemoryManager.malloc8d(n_coef);
final double[] newCoef = MemoryManager.malloc8d(n_coef);
Arrays.fill(step, Double.NaN);
Arrays.fill(oldCoef, Double.NaN);
for (int j = 0; j < n_coef; ++j) newCoef[j] = model._parms.init;
double oldLoglik = -Double.MAX_VALUE;
final int n_time = (int) (model._output.max_time - model._output.min_time + 1);
final boolean has_start_column = (model._parms.start_column != null);
final boolean has_weights_column = (model._parms.weights_column != null);
for (int i = 0; i <= model._parms.iter_max; ++i) {
model._output.iter = i;
final CoxPHTask coxMR =
new CoxPHTask(
self(),
dinfo,
newCoef,
model._output.min_time,
n_time,
n_offsets,
has_start_column,
has_weights_column)
.doAll(dinfo._adaptedFrame);
final double newLoglik = calcLoglik(model, coxMR);
if (newLoglik > oldLoglik) {
if (i == 0) calcCounts(model, coxMR);
calcModelStats(model, newCoef, newLoglik);
calcCumhaz_0(model, coxMR);
if (newLoglik == 0) model._output.lre = -Math.log10(Math.abs(oldLoglik - newLoglik));
else model._output.lre = -Math.log10(Math.abs((oldLoglik - newLoglik) / newLoglik));
if (model._output.lre >= model._parms.lre_min) break;
Arrays.fill(step, 0);
for (int j = 0; j < n_coef; ++j)
for (int k = 0; k < n_coef; ++k)
step[j] -= model._output.var_coef[j][k] * model._output.gradient[k];
for (int j = 0; j < n_coef; ++j)
if (Double.isNaN(step[j]) || Double.isInfinite(step[j])) break;
oldLoglik = newLoglik;
System.arraycopy(newCoef, 0, oldCoef, 0, oldCoef.length);
} else {
for (int j = 0; j < n_coef; ++j) step[j] /= 2;
}
for (int j = 0; j < n_coef; ++j) newCoef[j] = oldCoef[j] - step[j];
}
model.update(_key);
} catch (Throwable t) {
Job thisJob = DKV.getGet(_key);
if (thisJob._state == JobState.CANCELLED) {
Log.info("Job cancelled by user.");
} else {
t.printStackTrace();
failed(t);
throw t;
}
} finally {
updateModelOutput();
_parms.read_unlock_frames(CoxPH.this);
Scope.exit();
done(); // Job done!
}
tryComplete();
}