// Stopping criteria
boolean isDone(KMeansModel model, double[][] newCenters, double[][] oldCenters) {
if (!isRunning()) return true; // Stopped/cancelled
// Stopped for running out iterations
if (model._output._iterations >= _parms._max_iterations) return true;
// Compute average change in standardized cluster centers
if (oldCenters == null) return false; // No prior iteration, not stopping
double average_change = 0;
for (int clu = 0; clu < _parms._k; clu++)
average_change +=
hex.genmodel.GenModel.KMeans_distance(
oldCenters[clu], newCenters[clu], _isCats, null, null);
average_change /= _parms._k; // Average change per cluster
model._output._avg_centroids_chg =
ArrayUtils.copyAndFillOf(
model._output._avg_centroids_chg,
model._output._avg_centroids_chg.length + 1,
average_change);
model._output._training_time_ms =
ArrayUtils.copyAndFillOf(
model._output._training_time_ms,
model._output._training_time_ms.length + 1,
System.currentTimeMillis());
return average_change < TOLERANCE;
}
/**
* 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;
}
protected double doScoringAndSaveModel(
boolean finalScoring, boolean oob, boolean build_tree_one_node) {
double training_r2 = Double.NaN; // Training R^2 value, if computed
long now = System.currentTimeMillis();
if (_firstScore == 0) _firstScore = now;
long sinceLastScore = now - _timeLastScoreStart;
boolean updated = false;
new ProgressUpdate(
"Built " + _model._output._ntrees + " trees so far (out of " + _parms._ntrees + ").")
.fork(_progressKey);
// Now model already contains tid-trees in serialized form
if (_parms._score_each_iteration
|| finalScoring
|| (now - _firstScore < 4000)
|| // Score every time for 4 secs
// Throttle scoring to keep the cost sane; limit to a 10% duty cycle & every 4 secs
(sinceLastScore > 4000
&& // Limit scoring updates to every 4sec
(double) (_timeLastScoreEnd - _timeLastScoreStart) / sinceLastScore
< 0.1)) { // 10% duty cycle
checkMemoryFootPrint();
// If validation is specified we use a model for scoring, so we need to
// update it! First we save model with trees (i.e., make them available
// for scoring) and then update it with resulting error
_model.update(_key);
updated = true;
Log.info("============================================================== ");
SharedTreeModel.SharedTreeOutput out = _model._output;
_timeLastScoreStart = now;
// Score on training data
new ProgressUpdate("Scoring the model.").fork(_progressKey);
Score sc =
new Score(this, true, oob, _model._output.getModelCategory())
.doAll(train(), build_tree_one_node);
ModelMetrics mm = sc.makeModelMetrics(_model, _parms.train());
out._training_metrics = mm;
if (oob)
out._training_metrics._description = "Metrics reported on Out-Of-Bag training samples";
out._scored_train[out._ntrees].fillFrom(mm);
if (out._ntrees > 0) Log.info("Training " + out._scored_train[out._ntrees].toString());
// Score again on validation data
if (_parms._valid != null) {
Score scv =
new Score(this, false, false, _model._output.getModelCategory())
.doAll(valid(), build_tree_one_node);
ModelMetrics mmv = scv.makeModelMetrics(_model, _parms.valid());
out._validation_metrics = mmv;
out._scored_valid[out._ntrees].fillFrom(mmv);
if (out._ntrees > 0) Log.info("Validation " + out._scored_valid[out._ntrees].toString());
}
if (out._ntrees > 0) { // Compute variable importances
out._model_summary = createModelSummaryTable(out);
out._scoring_history = createScoringHistoryTable(out);
out._varimp = new hex.VarImp(_improvPerVar, out._names);
out._variable_importances = hex.ModelMetrics.calcVarImp(out._varimp);
Log.info(out._model_summary.toString());
// For Debugging:
// Log.info(out._scoring_history.toString());
// Log.info(out._variable_importances.toString());
}
ConfusionMatrix cm = mm.cm();
if (cm != null) {
if (cm._cm.length <= _parms._max_confusion_matrix_size) {
Log.info(cm.toASCII());
} else {
Log.info(
"Confusion Matrix is too large (max_confusion_matrix_size="
+ _parms._max_confusion_matrix_size
+ "): "
+ _nclass
+ " classes.");
}
}
_timeLastScoreEnd = System.currentTimeMillis();
}
// Double update - after either scoring or variable importance
if (updated) _model.update(_key);
return training_r2;
}