private void ProcessRowFirstPass()
{
AssertValid(assertGetters: true);
Single label = 0;
_labelGetter(ref label);
if (Single.IsNaN(label))
{
NumUnlabeledInstances++;
label = 0;
}
var intLabel = (int)label;
if (intLabel != label || intLabel < 0)
{
throw Host.Except("Invalid label: {0}", label);
}
_scoreGetter(ref _scores);
Host.Check(_scores.Length == _scoresArr.Length);
if (VBufferUtils.HasNaNs(ref _scores) || VBufferUtils.HasNonFinite(ref _scores))
{
NumBadScores++;
return;
}
_scores.CopyTo(_scoresArr);
Single weight = 1;
if (_weightGetter != null)
{
_weightGetter(ref weight);
if (!FloatUtils.IsFinite(weight))
{
NumBadWeights++;
weight = 1;
}
}
int j = 0;
foreach (var index in Enumerable.Range(0, _scoresArr.Length).OrderBy(i => _scoresArr[i]))
{
_indicesArr[j++] = index;
}
UnweightedCounters.UpdateFirstPass(intLabel, _scoresArr, 1, _indicesArr);
if (WeightedCounters != null)
{
WeightedCounters.UpdateFirstPass(intLabel, _scoresArr, weight, _indicesArr);
}
if (_clusterCentroids != null)
{
_featGetter(ref _features);
VectorUtils.Add(ref _features, ref _clusterCentroids[_indicesArr[0]]);
}
}
public PathFlows <double, TLabel> Aggregate(PathFlows <double, TLabel>[] paths)
{
var prices = new double[paths.First().Flows.Length];
foreach (PathFlows <double, TLabel> pathFlows in paths)
{
VectorUtils.Add(ref prices, pathFlows.Flows);
}
VectorUtils.Mult(ref prices, 1.0 / paths.Length);
return(new PathFlows <double, TLabel>(prices, paths.First().Labels));
}
protected void CombineCore(ref VBuffer <Single> dst, VBuffer <Single>[] src, Single[] weights = null)
{
Host.AssertNonEmpty(src);
Host.Assert(weights == null || Utils.Size(weights) == Utils.Size(src));
// REVIEW: Should this be tolerant of NaNs?
int len = GetClassCount(src);
if (!TryNormalize(src))
{
GetNaNOutput(ref dst, len);
return;
}
var values = dst.Values;
if (Utils.Size(values) < len)
{
values = new Single[len];
}
else
{
Array.Clear(values, 0, len);
}
// Set the output to values.
dst = new VBuffer <Single>(len, values, dst.Indices);
Single weightTotal;
if (weights == null)
{
weightTotal = src.Length;
for (int i = 0; i < src.Length; i++)
{
VectorUtils.Add(ref src[i], ref dst);
}
}
else
{
weightTotal = 0;
for (int i = 0; i < src.Length; i++)
{
var w = weights[i];
weightTotal += w;
VectorUtils.AddMult(ref src[i], w, ref dst);
}
}
VectorUtils.ScaleBy(ref dst, 1 / weightTotal);
}
/// <summary>
/// Batch-parallel optimizer
/// </summary>
/// <remarks>
/// REVIEW: consider getting rid of multithread-targeted members
/// Using TPL, the distinction between Multithreaded and Sequential implementations is unnecessary
/// </remarks>
protected virtual float DifferentiableFunctionMultithreaded(ref VBuffer <float> xDense, ref VBuffer <float> gradient, IProgressChannel pch)
{
Contracts.Assert(_data == null);
Contracts.Assert(_cursorFactory == null);
Contracts.Assert(_numChunks > 0);
Contracts.Assert(Utils.Size(_ranges) == _numChunks + 1);
Contracts.Assert(Utils.Size(_localLosses) == _numChunks);
Contracts.Assert(Utils.Size(_localGradients) + 1 == _numChunks);
Contracts.AssertValueOrNull(pch);
// Declare a local variable, since the lambda cannot capture the xDense. The gradient
// calculation will modify the local gradients, but not this xx value.
var xx = xDense;
var gg = gradient;
Parallel.For(0, _numChunks,
ichk =>
{
if (ichk == 0)
{
_localLosses[ichk] = DifferentiableFunctionComputeChunk(ichk, ref xx, ref gg, pch);
}
else
{
_localLosses[ichk] = DifferentiableFunctionComputeChunk(ichk, ref xx, ref _localGradients[ichk - 1], null);
}
});
gradient = gg;
float loss = _localLosses[0];
for (int i = 1; i < _numChunks; i++)
{
VectorUtils.Add(ref _localGradients[i - 1], ref gradient);
loss += _localLosses[i];
}
return(loss);
}
/// <summary>
/// Combine a bunch of models into one by averaging parameters
/// </summary>
protected void CombineParameters(IList <IParameterMixer <Float> > models, out VBuffer <Float> weights, out Float bias)
{
Type type = GetType();
Contracts.Check(type == models[0].GetType(), "Submodel for parameter mixer has the wrong type");
var first = (LinearPredictor)models[0];
weights = default(VBuffer <Float>);
first.Weight.CopyTo(ref weights);
bias = first.Bias;
for (int i = 1; i < models.Count; i++)
{
var m = models[i];
Contracts.Check(type == m.GetType(), "Submodel for parameter mixer has the wrong type");
var sub = (LinearPredictor)m;
var subweights = sub.Weight;
VectorUtils.Add(ref subweights, ref weights);
bias += sub.Bias;
}
VectorUtils.ScaleBy(ref weights, (Float)1 / models.Count);
bias /= models.Count;
}