public double[][] getNormBetaMultinomial(int idx) {
double[][] res = new double[nclasses()][];
Submodel sm = _submodels[idx];
int N = _dinfo.fullN() + 1;
double[] beta = sm.beta;
if (sm.idxs != null)
beta = ArrayUtils.expandAndScatter(beta, nclasses() * (_dinfo.fullN() + 1), sm.idxs);
for (int i = 0; i < res.length; ++i) res[i] = Arrays.copyOfRange(beta, i * N, (i + 1) * N);
return res;
}
public double[] getNormBeta() {
return _submodels[_best_lambda_idx].getBeta(MemoryManager.malloc8d(_dinfo.fullN() + 1));
}
/**
* Main constructor
*
* @param params Model parameters
* @param dinfo Data Info
* @param nClasses number of classes (1 for regression, 0 for autoencoder)
* @param train User-given training data frame, prepared by AdaptTestTrain
* @param valid User-specified validation data frame, prepared by AdaptTestTrain
*/
public DeepLearningModelInfo(
final DeepLearningParameters params,
final DataInfo dinfo,
int nClasses,
Frame train,
Frame valid) {
_classification = nClasses > 1;
_train = train;
_valid = valid;
data_info = dinfo;
parameters =
(DeepLearningParameters) params.clone(); // make a copy, don't change model's parameters
DeepLearningParameters.Sanity.modifyParms(
parameters, parameters, nClasses); // sanitize the model_info's parameters
final int num_input = dinfo.fullN();
final int num_output =
get_params()._autoencoder
? num_input
: (_classification ? train.lastVec().cardinality() : 1);
if (!get_params()._autoencoder) assert (num_output == nClasses);
_saw_missing_cats = dinfo._cats > 0 ? new boolean[data_info._cats] : null;
assert (num_input > 0);
assert (num_output > 0);
if (has_momenta() && adaDelta())
throw new IllegalArgumentException(
"Cannot have non-zero momentum and adaptive rate at the same time.");
final int layers = get_params()._hidden.length;
// units (# neurons for each layer)
units = new int[layers + 2];
if (get_params()._max_categorical_features <= Integer.MAX_VALUE - dinfo._nums)
units[0] = Math.min(dinfo._nums + get_params()._max_categorical_features, num_input);
else units[0] = num_input;
System.arraycopy(get_params()._hidden, 0, units, 1, layers);
units[layers + 1] = num_output;
boolean printLevels = units[0] > 1000L;
boolean warn = units[0] > 100000L;
if (printLevels) {
final String[][] domains = dinfo._adaptedFrame.domains();
int[] levels = new int[domains.length];
for (int i = 0; i < levels.length; ++i) {
levels[i] = domains[i] != null ? domains[i].length : 0;
}
Arrays.sort(levels);
if (warn) {
Log.warn(
"===================================================================================================================================");
Log.warn(
num_input
+ " input features"
+ (dinfo._cats > 0 ? " (after categorical one-hot encoding)" : "")
+ ". Can be slow and require a lot of memory.");
}
if (levels[levels.length - 1] > 0) {
int levelcutoff = levels[levels.length - 1 - Math.min(10, levels.length - 1)];
int count = 0;
for (int i = 0;
i < dinfo._adaptedFrame.numCols() - (get_params()._autoencoder ? 0 : 1) && count < 10;
++i) {
if (dinfo._adaptedFrame.domains()[i] != null
&& dinfo._adaptedFrame.domains()[i].length >= levelcutoff) {
if (warn) {
Log.warn(
"Categorical feature '"
+ dinfo._adaptedFrame._names[i]
+ "' has cardinality "
+ dinfo._adaptedFrame.domains()[i].length
+ ".");
} else {
Log.info(
"Categorical feature '"
+ dinfo._adaptedFrame._names[i]
+ "' has cardinality "
+ dinfo._adaptedFrame.domains()[i].length
+ ".");
}
}
count++;
}
}
if (warn) {
Log.warn("Suggestions:");
Log.warn(" *) Limit the size of the first hidden layer");
if (dinfo._cats > 0) {
Log.warn(
" *) Limit the total number of one-hot encoded features with the parameter 'max_categorical_features'");
Log.warn(
" *) Run h2o.interaction(...,pairwise=F) on high-cardinality categorical columns to limit the factor count, see http://learn.h2o.ai");
}
Log.warn(
"===================================================================================================================================");
}
}
// weights (to connect layers)
dense_row_weights = new Storage.DenseRowMatrix[layers + 1];
dense_col_weights = new Storage.DenseColMatrix[layers + 1];
// decide format of weight matrices row-major or col-major
if (get_params()._col_major)
dense_col_weights[0] = new Storage.DenseColMatrix(units[1], units[0]);
else dense_row_weights[0] = new Storage.DenseRowMatrix(units[1], units[0]);
for (int i = 1; i <= layers; ++i)
dense_row_weights[i] = new Storage.DenseRowMatrix(units[i + 1] /*rows*/, units[i] /*cols*/);
// biases (only for hidden layers and output layer)
biases = new Storage.DenseVector[layers + 1];
for (int i = 0; i <= layers; ++i) biases[i] = new Storage.DenseVector(units[i + 1]);
// average activation (only for hidden layers)
if (get_params()._autoencoder && get_params()._sparsity_beta > 0) {
avg_activations = new Storage.DenseVector[layers];
mean_a = new float[layers];
for (int i = 0; i < layers; ++i) avg_activations[i] = new Storage.DenseVector(units[i + 1]);
}
allocateHelperArrays();
// for diagnostics
mean_rate = new float[units.length];
rms_rate = new float[units.length];
mean_bias = new float[units.length];
rms_bias = new float[units.length];
mean_weight = new float[units.length];
rms_weight = new float[units.length];
}
/**
* Train a Deep Learning model, assumes that all members are populated If checkpoint == null,
* then start training a new model, otherwise continue from a checkpoint
*/
public final void buildModel() {
DeepLearningModel cp = null;
if (_parms._checkpoint == null) {
cp =
new DeepLearningModel(
dest(),
_parms,
new DeepLearningModel.DeepLearningModelOutput(DeepLearning.this),
_train,
_valid,
nclasses());
cp.model_info().initializeMembers();
} else {
final DeepLearningModel previous = DKV.getGet(_parms._checkpoint);
if (previous == null) throw new IllegalArgumentException("Checkpoint not found.");
Log.info("Resuming from checkpoint.");
_job.update(0, "Resuming from checkpoint");
if (isClassifier() != previous._output.isClassifier())
throw new H2OIllegalArgumentException(
"Response type must be the same as for the checkpointed model.");
if (isSupervised() != previous._output.isSupervised())
throw new H2OIllegalArgumentException(
"Model type must be the same as for the checkpointed model.");
// check the user-given arguments for consistency
DeepLearningParameters oldP =
previous._parms; // sanitized parameters for checkpointed model
DeepLearningParameters newP = _parms; // user-given parameters for restart
DeepLearningParameters oldP2 = (DeepLearningParameters) oldP.clone();
DeepLearningParameters newP2 = (DeepLearningParameters) newP.clone();
DeepLearningParameters.Sanity.modifyParms(
oldP, oldP2, nclasses()); // sanitize the user-given parameters
DeepLearningParameters.Sanity.modifyParms(
newP, newP2, nclasses()); // sanitize the user-given parameters
DeepLearningParameters.Sanity.checkpoint(oldP2, newP2);
DataInfo dinfo;
try {
// PUBDEV-2513: Adapt _train and _valid (in-place) to match the frames that were used for
// the previous model
// This can add or remove dummy columns (can happen if the dataset is sparse and datasets
// have different non-const columns)
for (String st : previous.adaptTestForTrain(_train, true, false)) Log.warn(st);
for (String st : previous.adaptTestForTrain(_valid, true, false)) Log.warn(st);
dinfo = makeDataInfo(_train, _valid, _parms, nclasses());
DKV.put(dinfo);
cp = new DeepLearningModel(dest(), _parms, previous, false, dinfo);
cp.write_lock(_job);
if (!Arrays.equals(cp._output._names, previous._output._names)) {
throw new H2OIllegalArgumentException(
"The columns of the training data must be the same as for the checkpointed model. Check ignored columns (or disable ignore_const_cols).");
}
if (!Arrays.deepEquals(cp._output._domains, previous._output._domains)) {
throw new H2OIllegalArgumentException(
"Categorical factor levels of the training data must be the same as for the checkpointed model.");
}
if (dinfo.fullN() != previous.model_info().data_info().fullN()) {
throw new H2OIllegalArgumentException(
"Total number of predictors is different than for the checkpointed model.");
}
if (_parms._epochs <= previous.epoch_counter) {
throw new H2OIllegalArgumentException(
"Total number of epochs must be larger than the number of epochs already trained for the checkpointed model ("
+ previous.epoch_counter
+ ").");
}
// these are the mutable parameters that are to be used by the model (stored in
// model_info._parms)
final DeepLearningParameters actualNewP =
cp.model_info()
.get_params(); // actually used parameters for model building (defaults filled in,
// etc.)
assert (actualNewP != previous.model_info().get_params());
assert (actualNewP != newP);
assert (actualNewP != oldP);
DeepLearningParameters.Sanity.update(actualNewP, newP, nclasses());
Log.info(
"Continuing training after "
+ String.format("%.3f", previous.epoch_counter)
+ " epochs from the checkpointed model.");
cp.update(_job);
} catch (H2OIllegalArgumentException ex) {
if (cp != null) {
cp.unlock(_job);
cp.delete();
cp = null;
}
throw ex;
} finally {
if (cp != null) cp.unlock(_job);
}
}
trainModel(cp);
// clean up, but don't delete weights and biases if user asked for export
List<Key> keep = new ArrayList<>();
try {
if (_parms._export_weights_and_biases
&& cp._output.weights != null
&& cp._output.biases != null) {
for (Key k : Arrays.asList(cp._output.weights)) {
keep.add(k);
for (Vec vk : ((Frame) DKV.getGet(k)).vecs()) {
keep.add(vk._key);
}
}
for (Key k : Arrays.asList(cp._output.biases)) {
keep.add(k);
for (Vec vk : ((Frame) DKV.getGet(k)).vecs()) {
keep.add(vk._key);
}
}
}
} finally {
Scope.exit(keep.toArray(new Key[keep.size()]));
}
}
@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();
}