private void InitializeTrainingState(int fieldCount, int featureCount, FieldAwareFactorizationMachinePredictor predictor, out float[] linearWeights,
out AlignedArray latentWeightsAligned, out float[] linearAccumulatedSquaredGrads, out AlignedArray latentAccumulatedSquaredGradsAligned)
{
linearWeights = new float[featureCount];
latentWeightsAligned = new AlignedArray(featureCount * fieldCount * _latentDimAligned, 16);
linearAccumulatedSquaredGrads = new float[featureCount];
latentAccumulatedSquaredGradsAligned = new AlignedArray(featureCount * fieldCount * _latentDimAligned, 16);
if (predictor == null)
{
var rng = Host.Rand;
for (int j = 0; j < featureCount; j++)
{
linearWeights[j] = 0;
linearAccumulatedSquaredGrads[j] = 1;
for (int f = 0; f < fieldCount; f++)
{
int vBias = j * fieldCount * _latentDimAligned + f * _latentDimAligned;
for (int k = 0; k < _latentDimAligned; k++)
{
if (k < _latentDim)
{
latentWeightsAligned[vBias + k] = _radius * (float)rng.NextDouble();
}
else
{
latentWeightsAligned[vBias + k] = 0;
}
latentAccumulatedSquaredGradsAligned[vBias + k] = 1;
}
}
}
}
else
{
predictor.CopyLinearWeightsTo(linearWeights);
predictor.CopyLatentWeightsTo(latentWeightsAligned);
for (int j = 0; j < featureCount; j++)
{
linearAccumulatedSquaredGrads[j] = 1;
for (int f = 0; f < fieldCount; f++)
{
int vBias = j * fieldCount * _latentDimAligned + f * _latentDimAligned;
for (int k = 0; k < _latentDimAligned; k++)
{
latentAccumulatedSquaredGradsAligned[vBias + k] = 1;
}
}
}
}
}
private static double CalculateAvgLoss(IChannel ch, RoleMappedData data, bool norm, float[] linearWeights, AlignedArray latentWeightsAligned,
int latentDimAligned, AlignedArray latentSum, int[] featureFieldBuffer, int[] featureIndexBuffer, float[] featureValueBuffer, VBuffer <float> buffer, ref long badExampleCount)
{
var featureColumns = data.Schema.GetColumns(RoleMappedSchema.ColumnRole.Feature);
Func <int, bool> pred = c => featureColumns.Select(ci => ci.Index).Contains(c) || c == data.Schema.Label.Value.Index || c == data.Schema.Weight?.Index;
var getters = new ValueGetter <VBuffer <float> > [featureColumns.Count];
float label = 0;
float weight = 1;
double loss = 0;
float modelResponse = 0;
long exampleCount = 0;
badExampleCount = 0;
int count = 0;
using (var cursor = data.Data.GetRowCursor(pred))
{
var labelGetter = RowCursorUtils.GetLabelGetter(cursor, data.Schema.Label.Value.Index);
var weightGetter = data.Schema.Weight?.Index is int weightIdx?cursor.GetGetter <float>(weightIdx) : null;
for (int f = 0; f < featureColumns.Count; f++)
{
getters[f] = cursor.GetGetter <VBuffer <float> >(featureColumns[f].Index);
}
while (cursor.MoveNext())
{
labelGetter(ref label);
weightGetter?.Invoke(ref weight);
float annihilation = label - label + weight - weight;
if (!FloatUtils.IsFinite(annihilation))
{
badExampleCount++;
continue;
}
if (!FieldAwareFactorizationMachineUtils.LoadOneExampleIntoBuffer(getters, buffer, norm, ref count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer))
{
badExampleCount++;
continue;
}
FieldAwareFactorizationMachineInterface.CalculateIntermediateVariables(featureColumns.Count, latentDimAligned, count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer, linearWeights, latentWeightsAligned, latentSum, ref modelResponse);
loss += weight * CalculateLoss(label, modelResponse);
exampleCount++;
}
}
return(loss / exampleCount);
}
internal FieldAwareFactorizationMachinePredictor(IHostEnvironment env, bool norm, int fieldCount, int featureCount, int latentDim,
float[] linearWeights, AlignedArray latentWeightsAligned) : base(env, LoaderSignature)
{
Host.Assert(fieldCount > 0);
Host.Assert(featureCount > 0);
Host.Assert(latentDim > 0);
Host.Assert(Utils.Size(linearWeights) == featureCount);
LatentDimAligned = FieldAwareFactorizationMachineUtils.GetAlignedVectorLength(latentDim);
Host.Assert(latentWeightsAligned.Size == checked (featureCount * fieldCount * LatentDimAligned));
_norm = norm;
FieldCount = fieldCount;
FeatureCount = featureCount;
LatentDim = latentDim;
_linearWeights = linearWeights;
_latentWeightsAligned = latentWeightsAligned;
}
public Row GetRow(Row input, Func <int, bool> predicate, out Action action)
{
var latentSum = new AlignedArray(_pred.FieldCount * _pred.FieldCount * _pred.LatentDimAligned, 16);
var featureBuffer = new VBuffer <float>();
var featureFieldBuffer = new int[_pred.FeatureCount];
var featureIndexBuffer = new int[_pred.FeatureCount];
var featureValueBuffer = new float[_pred.FeatureCount];
var inputGetters = new ValueGetter <VBuffer <float> > [_pred.FieldCount];
if (predicate(0) || predicate(1))
{
for (int f = 0; f < _pred.FieldCount; f++)
{
inputGetters[f] = input.GetGetter <VBuffer <float> >(_inputColumnIndexes[f]);
}
}
action = null;
var getters = new Delegate[2];
if (predicate(0))
{
ValueGetter <float> responseGetter = (ref float value) =>
{
value = _pred.CalculateResponse(inputGetters, featureBuffer, featureFieldBuffer, featureIndexBuffer, featureValueBuffer, latentSum);
};
getters[0] = responseGetter;
}
if (predicate(1))
{
ValueGetter <float> probGetter = (ref float value) =>
{
value = _pred.CalculateResponse(inputGetters, featureBuffer, featureFieldBuffer, featureIndexBuffer, featureValueBuffer, latentSum);
value = MathUtils.SigmoidSlow(value);
};
getters[1] = probGetter;
}
return(new SimpleRow(OutputSchema, input, getters));
}
public static void CalculateIntermediateVariables(int fieldCount, int latentDim, int count, int[] fieldIndices, int[] featureIndices, float[] featureValues,
float[] linearWeights, AlignedArray latentWeights, AlignedArray latentSum, ref float response)
{
Contracts.AssertNonEmpty(fieldIndices);
Contracts.AssertNonEmpty(featureValues);
Contracts.AssertNonEmpty(featureIndices);
Contracts.AssertNonEmpty(linearWeights);
Contracts.Assert(Compat(latentWeights));
Contracts.Assert(Compat(latentSum));
unsafe
{
fixed(int *pf = &fieldIndices[0])
fixed(int *pi = &featureIndices[0])
fixed(float *px = &featureValues[0])
fixed(float *pw = &linearWeights[0])
fixed(float *pv = &latentWeights.Items[0])
fixed(float *pq = &latentSum.Items[0])
fixed(float *pr = &response)
CalculateIntermediateVariablesNative(fieldCount, latentDim, count, pf, pi, px, pw, Ptr(latentWeights, pv), Ptr(latentSum, pq), pr);
}
}
/// <summary>
/// Initialize model parameters with a trained model.
/// </summary>
/// <param name="env">The host environment</param>
/// <param name="norm">True if user wants to normalize feature vector to unit length.</param>
/// <param name="fieldCount">The number of fileds, which is the symbol `m` in the doc: https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf </param>
/// <param name="featureCount">The number of features, which is the symbol `n` in the doc: https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf </param>
/// <param name="latentDim">The latent dimensions, which is the length of `v_{j, f}` in the doc: https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf </param>
/// <param name="linearWeights">The linear coefficients of the features, which is the symbol `w` in the doc: https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf </param>
/// <param name="latentWeights">Latent representation of each feature. Note that one feature may have <see cref="FieldCount"/> latent vectors
/// and each latent vector contains <see cref="LatentDim"/> values. In the f-th field, the j-th feature's latent vector, `v_{j, f}` in the doc
/// https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf, starts at latentWeights[j * fieldCount * latentDim + f * latentDim].
/// The k-th element in v_{j, f} is latentWeights[j * fieldCount * latentDim + f * latentDim + k]. The size of the array must be featureCount x fieldCount x latentDim.</param>
public FieldAwareFactorizationMachineModelParameters(IHostEnvironment env, bool norm, int fieldCount, int featureCount, int latentDim,
float[] linearWeights, float[] latentWeights) : base(env, LoaderSignature)
{
Host.Assert(fieldCount > 0);
Host.Assert(featureCount > 0);
Host.Assert(latentDim > 0);
Host.Assert(Utils.Size(linearWeights) == featureCount);
LatentDimAligned = FieldAwareFactorizationMachineUtils.GetAlignedVectorLength(latentDim);
Host.Assert(Utils.Size(latentWeights) == checked (featureCount * fieldCount * LatentDimAligned));
_norm = norm;
FieldCount = fieldCount;
FeatureCount = featureCount;
LatentDim = latentDim;
_linearWeights = linearWeights;
_latentWeightsAligned = new AlignedArray(FeatureCount * FieldCount * LatentDimAligned, 16);
for (int j = 0; j < FeatureCount; j++)
{
for (int f = 0; f < FieldCount; f++)
{
int index = j * FieldCount * LatentDim + f * LatentDim;
int indexAligned = j * FieldCount * LatentDimAligned + f * LatentDimAligned;
for (int k = 0; k < LatentDimAligned; k++)
{
if (k < LatentDim)
{
_latentWeightsAligned[indexAligned + k] = latentWeights[index + k];
}
else
{
_latentWeightsAligned[indexAligned + k] = 0;
}
}
}
}
}
private void TrainCore(IChannel ch, IProgressChannel pch, RoleMappedData data, RoleMappedData validData, FieldAwareFactorizationMachinePredictor predictor)
{
Host.AssertValue(ch);
Host.AssertValue(pch);
data.CheckBinaryLabel();
var featureColumns = data.Schema.GetColumns(RoleMappedSchema.ColumnRole.Feature);
int fieldCount = featureColumns.Count;
int totalFeatureCount = 0;
int[] fieldColumnIndexes = new int[fieldCount];
for (int f = 0; f < fieldCount; f++)
{
var col = featureColumns[f];
Host.Assert(col.Type.AsVector.VectorSize > 0);
if (col == null)
throw ch.ExceptParam(nameof(data), "Empty feature column not allowed");
Host.Assert(!data.Schema.Schema.IsHidden(col.Index));
if (!col.Type.IsKnownSizeVector || col.Type.ItemType != NumberType.Float)
throw ch.ExceptParam(nameof(data), "Training feature column '{0}' must be a known-size vector of R4, but has type: {1}.", col.Name, col.Type);
fieldColumnIndexes[f] = col.Index;
totalFeatureCount += col.Type.AsVector.VectorSize;
}
ch.Check(checked(totalFeatureCount * fieldCount * _latentDimAligned) <= Utils.ArrayMaxSize, "Latent dimension or the number of fields too large");
if (predictor != null)
{
ch.Check(predictor.FeatureCount == totalFeatureCount, "Input model's feature count mismatches training feature count");
ch.Check(predictor.LatentDim == _latentDim, "Input model's latent dimension mismatches trainer's");
}
if (validData != null)
{
validData.CheckBinaryLabel();
var validFeatureColumns = data.Schema.GetColumns(RoleMappedSchema.ColumnRole.Feature);
Host.Assert(fieldCount == validFeatureColumns.Count);
for (int f = 0; f < fieldCount; f++)
Host.Assert(featureColumns[f] == validFeatureColumns[f]);
}
bool shuffle = _shuffle;
if (shuffle && !data.Data.CanShuffle)
{
ch.Warning("Training data does not support shuffling, so ignoring request to shuffle");
shuffle = false;
}
var rng = shuffle ? Host.Rand : null;
var featureGetters = new ValueGetter<VBuffer<float>>[fieldCount];
var featureBuffer = new VBuffer<float>();
var featureValueBuffer = new float[totalFeatureCount];
var featureIndexBuffer = new int[totalFeatureCount];
var featureFieldBuffer = new int[totalFeatureCount];
var latentSum = new AlignedArray(fieldCount * fieldCount * _latentDimAligned, 16);
var metricNames = new List<string>() { "Training-loss" };
if (validData != null)
metricNames.Add("Validation-loss");
int iter = 0;
long exampleCount = 0;
long badExampleCount = 0;
long validBadExampleCount = 0;
double loss = 0;
double validLoss = 0;
pch.SetHeader(new ProgressHeader(metricNames.ToArray(), new string[] { "iterations", "examples" }), entry =>
{
entry.SetProgress(0, iter, _numIterations);
entry.SetProgress(1, exampleCount);
});
Func<int, bool> pred = c => fieldColumnIndexes.Contains(c) || c == data.Schema.Label.Index || (data.Schema.Weight != null && c == data.Schema.Weight.Index);
InitializeTrainingState(fieldCount, totalFeatureCount, predictor, out float[] linearWeights,
out AlignedArray latentWeightsAligned, out float[] linearAccSqGrads, out AlignedArray latentAccSqGradsAligned);
// refer to Algorithm 3 in https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf
while (iter++ < _numIterations)
{
using (var cursor = data.Data.GetRowCursor(pred, rng))
{
var labelGetter = RowCursorUtils.GetLabelGetter(cursor, data.Schema.Label.Index);
var weightGetter = data.Schema.Weight == null ? null : RowCursorUtils.GetGetterAs<float>(NumberType.R4, cursor, data.Schema.Weight.Index);
for (int i = 0; i < fieldCount; i++)
featureGetters[i] = cursor.GetGetter<VBuffer<float>>(fieldColumnIndexes[i]);
loss = 0;
exampleCount = 0;
badExampleCount = 0;
while (cursor.MoveNext())
{
float label = 0;
float weight = 1;
int count = 0;
float modelResponse = 0;
labelGetter(ref label);
weightGetter?.Invoke(ref weight);
float annihilation = label - label + weight - weight;
if (!FloatUtils.IsFinite(annihilation))
{
badExampleCount++;
continue;
}
if (!FieldAwareFactorizationMachineUtils.LoadOneExampleIntoBuffer(featureGetters, featureBuffer, _norm, ref count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer))
{
badExampleCount++;
continue;
}
// refer to Algorithm 1 in [3] https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf
FieldAwareFactorizationMachineInterface.CalculateIntermediateVariables(fieldCount, _latentDimAligned, count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer, linearWeights, latentWeightsAligned, latentSum, ref modelResponse);
var slope = CalculateLossSlope(label, modelResponse);
// refer to Algorithm 2 in [3] https://github.com/wschin/fast-ffm/blob/master/fast-ffm.pdf
FieldAwareFactorizationMachineInterface.CalculateGradientAndUpdate(_lambdaLinear, _lambdaLatent, _learningRate, fieldCount, _latentDimAligned, weight, count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer, latentSum, slope, linearWeights, latentWeightsAligned, linearAccSqGrads, latentAccSqGradsAligned);
loss += weight * CalculateLoss(label, modelResponse);
exampleCount++;
}
loss /= exampleCount;
}
if (_verbose)
{
if (validData == null)
pch.Checkpoint(loss, iter, exampleCount);
else
{
validLoss = CalculateAvgLoss(ch, validData, _norm, linearWeights, latentWeightsAligned, _latentDimAligned, latentSum,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer, featureBuffer, ref validBadExampleCount);
pch.Checkpoint(loss, validLoss, iter, exampleCount);
}
}
}
if (badExampleCount != 0)
ch.Warning($"Skipped {badExampleCount} examples with bad label/weight/features in training set");
if (validBadExampleCount != 0)
ch.Warning($"Skipped {validBadExampleCount} examples with bad label/weight/features in validation set");
_pred = new FieldAwareFactorizationMachinePredictor(Host, _norm, fieldCount, totalFeatureCount, _latentDim, linearWeights, latentWeightsAligned);
}
internal void CopyLatentWeightsTo(AlignedArray latentWeights)
{
Host.AssertValue(_latentWeightsAligned);
Host.AssertValue(latentWeights);
latentWeights.CopyFrom(_latentWeightsAligned);
}
internal float CalculateResponse(ValueGetter <VBuffer <float> >[] getters, VBuffer <float> featureBuffer,
int[] featureFieldBuffer, int[] featureIndexBuffer, float[] featureValueBuffer, AlignedArray latentSum)
{
int count = 0;
float modelResponse = 0;
FieldAwareFactorizationMachineUtils.LoadOneExampleIntoBuffer(getters, featureBuffer, _norm, ref count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer);
FieldAwareFactorizationMachineInterface.CalculateIntermediateVariables(FieldCount, LatentDimAligned, count,
featureFieldBuffer, featureIndexBuffer, featureValueBuffer, _linearWeights, _latentWeightsAligned, latentSum, ref modelResponse);
return(modelResponse);
}
public static void CalculateGradientAndUpdate(float lambdaLinear, float lambdaLatent, float learningRate, int fieldCount, int latentDim,
float weight, int count, int[] fieldIndices, int[] featureIndices, float[] featureValues, AlignedArray latentSum, float slope,
float[] linearWeights, AlignedArray latentWeights, float[] linearAccumulatedSquaredGrads, AlignedArray latentAccumulatedSquaredGrads)
{
Contracts.AssertNonEmpty(fieldIndices);
Contracts.AssertNonEmpty(featureIndices);
Contracts.AssertNonEmpty(featureValues);
Contracts.Assert(Compat(latentSum));
Contracts.AssertNonEmpty(linearWeights);
Contracts.Assert(Compat(latentWeights));
Contracts.AssertNonEmpty(linearAccumulatedSquaredGrads);
Contracts.Assert(Compat(latentAccumulatedSquaredGrads));
unsafe
{
fixed(int *pf = &fieldIndices[0])
fixed(int *pi = &featureIndices[0])
fixed(float *px = &featureValues[0])
fixed(float *pq = &latentSum.Items[0])
fixed(float *pw = &linearWeights[0])
fixed(float *pv = &latentWeights.Items[0])
fixed(float *phw = &linearAccumulatedSquaredGrads[0])
fixed(float *phv = &latentAccumulatedSquaredGrads.Items[0])
CalculateGradientAndUpdateNative(lambdaLinear, lambdaLatent, learningRate, fieldCount, latentDim, weight, count, pf, pi, px,
Ptr(latentSum, pq), slope, pw, Ptr(latentWeights, pv), phw, Ptr(latentAccumulatedSquaredGrads, phv));
}
}
private static bool Compat(AlignedArray a)
{
Contracts.AssertValue(a);
Contracts.Assert(a.Size > 0);
return(a.CbAlign == CbAlign);
}
