예제 #1
0
파일: linsolve.go 프로젝트: reggo/reggo
// Creates the features from the inputs. Features must be nSamples x nFeatures or nil
func FeaturizeTrainable(t Trainable, inputs common.RowMatrix, featurizedInputs *mat64.Dense) *mat64.Dense {
	nSamples, nDim := inputs.Dims()
	if featurizedInputs == nil {
		nFeatures := t.NumFeatures()
		featurizedInputs = mat64.NewDense(nSamples, nFeatures, nil)
	}

	rowViewer, isRowViewer := inputs.(mat64.RawRowViewer)
	var f func(start, end int)
	if isRowViewer {
		f = func(start, end int) {
			featurizer := t.NewFeaturizer()
			for i := start; i < end; i++ {
				featurizer.Featurize(rowViewer.RawRowView(i), featurizedInputs.RawRowView(i))
			}
		}
	} else {
		f = func(start, end int) {
			featurizer := t.NewFeaturizer()
			input := make([]float64, nDim)
			for i := start; i < end; i++ {
				inputs.Row(input, i)
				featurizer.Featurize(input, featurizedInputs.RawRowView(i))
			}
		}
	}

	common.ParallelFor(nSamples, common.GetGrainSize(nSamples, minGrain, maxGrain), f)
	return featurizedInputs
}
예제 #2
0
파일: kmeans.go 프로젝트: reggo/reggo
func (k *KMeans) Train(inputs *common.RowMatrix) {
	nSamples, inputDim := inputs.Dims()
	centroids := make([][]float64, k.Clusters)

	// Assign the centroids to random data point to start
	perm := rand.Perm(nSamples)
	for i := 0; i < k.Clusters; i++ {
		data := make([]float64, inputDim)
		idx := perm[i]
		inputs.Row(data, idx)
		centroids[i] = data
	}

	distances := mat64.NewDense(nSamples, k.Clusters, nil)
	row := make([]float64, inputDim)
	for i := 0; i < nSamples; i++ {
		inputs.Row(row, i)
		for j := 0; j < k.Clusters; j++ {
			d := floats.Distance(row, centroids[j])
			distances.Set(i, j, d)
		}
	}

}
예제 #3
0
파일: regtest.go 프로젝트: reggo/reggo
// TestLinearsolveAndDeriv compares the optimal weights found from gradient-based optimization with those found
// from computing a linear solve
func TestLinearsolveAndDeriv(t *testing.T, linear train.LinearTrainable, inputs, trueOutputs common.RowMatrix, name string) {
	// Compare with no weights

	rows, cols := trueOutputs.Dims()
	predOutLinear := mat64.NewDense(rows, cols, nil)
	parametersLinearSolve := train.LinearSolve(linear, nil, inputs, trueOutputs, nil, nil)

	linear.SetParameters(parametersLinearSolve)
	linear.Predictor().PredictBatch(inputs, predOutLinear)

	//fmt.Println("Pred out linear", predOutLinear)

	linear.RandomizeParameters()
	parameters := linear.Parameters(nil)

	batch := train.NewBatchGradBased(linear, true, inputs, trueOutputs, nil, loss.SquaredDistance{}, regularize.None{})
	problem := batch
	settings := multivariate.DefaultSettings()
	settings.GradAbsTol = 1e-11
	//settings. = 0

	result, err := multivariate.OptimizeGrad(problem, parameters, settings, nil)
	if err != nil {
		t.Errorf("Error training: %v", err)
	}

	parametersDeriv := result.Loc

	deriv := make([]float64, linear.NumParameters())

	loss1 := batch.ObjGrad(parametersDeriv, deriv)

	linear.SetParameters(parametersDeriv)
	predOutDeriv := mat64.NewDense(rows, cols, nil)
	linear.Predictor().PredictBatch(inputs, predOutDeriv)

	linear.RandomizeParameters()
	init2 := linear.Parameters(nil)
	batch2 := train.NewBatchGradBased(linear, true, inputs, trueOutputs, nil, loss.SquaredDistance{}, regularize.None{})
	problem2 := batch2
	result2, err := multivariate.OptimizeGrad(problem2, init2, settings, nil)
	parametersDeriv2 := result2.Loc

	//fmt.Println("starting deriv2 loss")
	deriv2 := make([]float64, linear.NumParameters())
	loss2 := batch2.ObjGrad(parametersDeriv2, deriv2)

	//fmt.Println("starting derivlin loss")
	derivlinear := make([]float64, linear.NumParameters())
	lossLin := batch2.ObjGrad(parametersLinearSolve, derivlinear)

	_ = loss1
	_ = loss2
	_ = lossLin

	/*

		fmt.Println("param deriv 1 =", parametersDeriv)
		fmt.Println("param deriv2  =", parametersDeriv2)
		fmt.Println("linear params =", parametersLinearSolve)

		fmt.Println("deriv1 loss =", loss1)
		fmt.Println("deriv2 loss =", loss2)
		fmt.Println("lin loss    =", lossLin)

		fmt.Println("deriv    =", deriv)
		fmt.Println("deriv2   =", deriv2)
		fmt.Println("linderiv =", derivlinear)

		//fmt.Println("Pred out deriv", predOutDeriv)

	*/

	/*
		for i := 0; i < rows; i++ {
			fmt.Println(predOutLinear.RowView(i), predOutBatch.RowView(i))
		}
	*/

	if !floats.EqualApprox(parametersLinearSolve, parametersDeriv, 1e-8) {
		t.Errorf("Parameters don't match for gradient based and linear solve.")
		//for i := range parametersDeriv {
		//	fmt.Printf("index %v: Deriv = %v, linsolve = %v, diff = %v\n", i, parametersDeriv[i], parametersLinearSolve[i], parametersDeriv[i]-parametersLinearSolve[i])
		//}
	}

}
예제 #4
0
파일: regtest.go 프로젝트: reggo/reggo
// TestPredict tests that predict returns the expected value, and that calling predict in parallel
// also works
func TestPredictAndBatch(t *testing.T, p Predictor, inputs, trueOutputs common.RowMatrix, name string) {
	nSamples, inputDim := inputs.Dims()
	if inputDim != p.InputDim() {
		panic("input Dim doesn't match predictor input dim")
	}
	nOutSamples, outputDim := trueOutputs.Dims()
	if outputDim != p.OutputDim() {
		panic("outpuDim doesn't match predictor outputDim")
	}
	if nOutSamples != nSamples {
		panic("inputs and outputs have different number of rows")
	}

	// First, test sequentially
	for i := 0; i < nSamples; i++ {
		trueOut := make([]float64, outputDim)
		for j := 0; j < outputDim; j++ {
			trueOut[j] = trueOutputs.At(i, j)
		}
		// Predict with nil
		input := make([]float64, inputDim)
		inputCpy := make([]float64, inputDim)
		for j := 0; j < inputDim; j++ {
			input[j] = inputs.At(i, j)
			inputCpy[j] = inputs.At(i, j)
		}

		out1, err := p.Predict(input, nil)
		if err != nil {
			t.Errorf(name + ": Error predicting with nil output")
			return
		}
		if !floats.Equal(input, inputCpy) {
			t.Errorf("%v: input changed with nil input for row %v", name, i)
			break
		}
		out2 := make([]float64, outputDim)
		for j := 0; j < outputDim; j++ {
			out2[j] = rand.NormFloat64()
		}

		_, err = p.Predict(input, out2)
		if err != nil {
			t.Errorf("%v: error predicting with non-nil input for row %v", name, i)
			break
		}
		if !floats.Equal(input, inputCpy) {
			t.Errorf("%v: input changed with non-nil input for row %v", name, i)
			break
		}

		if !floats.Equal(out1, out2) {
			t.Errorf(name + ": different answers with nil and non-nil predict ")
			break
		}
		if !floats.EqualApprox(out1, trueOut, 1e-14) {
			t.Errorf("%v: predicted output doesn't match for row %v. Expected %v, found %v", name, i, trueOut, out1)
			break
		}
	}

	// Check that predict errors with bad sized arguments
	badOuput := make([]float64, outputDim+1)
	input := make([]float64, inputDim)
	for i := 0; i < inputDim; i++ {
		input[i] = inputs.At(0, i)
	}
	output := make([]float64, outputDim)
	for i := 0; i < outputDim; i++ {
		output[i] = trueOutputs.At(0, i)
	}

	_, err := p.Predict(input, badOuput)
	if err == nil {
		t.Errorf("Predict did not throw an error with an output too large")
	}
	if outputDim > 1 {
		badOuput := make([]float64, outputDim-1)
		_, err := p.Predict(input, badOuput)
		if err == nil {
			t.Errorf("Predict did not throw an error with an output too small")
		}
	}

	badInput := make([]float64, inputDim+1)
	_, err = p.Predict(badInput, output)
	if err == nil {
		t.Errorf("Predict did not err when input is too large")
	}
	if inputDim > 1 {
		badInput := make([]float64, inputDim-1)
		_, err = p.Predict(badInput, output)
		if err == nil {
			t.Errorf("Predict did not err when input is too small")
		}
	}

	// Now, test batch
	// With non-nil
	inputCpy := &mat64.Dense{}
	inputCpy.Clone(inputs)
	predOutput, err := p.PredictBatch(inputs, nil)
	if err != nil {
		t.Errorf("Error batch predicting: %v", err)
	}
	if !inputCpy.Equals(inputs) {
		t.Errorf("Inputs changed during call to PredictBatch")
	}
	predOutputRows, predOutputCols := predOutput.Dims()
	if predOutputRows != nSamples || predOutputCols != outputDim {
		t.Errorf("Dimension mismatch after predictbatch with nil input")
	}

	outputs := mat64.NewDense(nSamples, outputDim, nil)
	_, err = p.PredictBatch(inputs, outputs)

	pd := predOutput.(*mat64.Dense)
	if !pd.Equals(outputs) {
		t.Errorf("Different outputs from predict batch with nil and non-nil")
	}

	badInputs := mat64.NewDense(nSamples, inputDim+1, nil)
	_, err = p.PredictBatch(badInputs, outputs)
	if err == nil {
		t.Error("PredictBatch did not err when input dim too large")
	}
	badInputs = mat64.NewDense(nSamples+1, inputDim, nil)
	_, err = p.PredictBatch(badInputs, outputs)
	if err == nil {
		t.Errorf("PredictBatch did not err with row mismatch")
	}
	badOuputs := mat64.NewDense(nSamples, outputDim+1, nil)
	_, err = p.PredictBatch(inputs, badOuputs)
	if err == nil {
		t.Errorf("PredictBatch did not err with output dim too large")
	}
}
예제 #5
0
파일: gradbased.go 프로젝트: reggo/reggo
// NewBatchGradBased creates a new batch grad based with the given inputs
func NewBatchGradBased(trainable Trainable, cacheFeatures bool, inputs, outputs common.RowMatrix, weights []float64, losser loss.DerivLosser, regularizer regularize.Regularizer) *BatchGradBased {
	var features *mat64.Dense
	if cacheFeatures {
		features = FeaturizeTrainable(trainable, inputs, nil)
	}

	// TODO: Add in error checking

	if losser == nil {
		losser = loss.SquaredDistance{}
	}
	if regularizer == nil {
		regularizer = regularize.None{}
	}

	if weights != nil {
		// TODO: Fix weights
		panic("non-nil weights")
	}

	nTrain, outputDim := outputs.Dims()
	_, inputDim := inputs.Dims()
	g := &BatchGradBased{
		t:           trainable,
		inputs:      inputs,
		outputs:     outputs,
		features:    features,
		losser:      losser,
		regularizer: regularizer,
		nTrain:      nTrain,
		outputDim:   outputDim,
		inputDim:    inputDim,
		nParameters: trainable.NumParameters(),
		grainSize:   trainable.GrainSize(),
	}

	// TODO: Add in row viewer stuff
	// TODO: Create a different function for computing just the loss
	//inputRowViewer, ok := inputs.(mat64.RowViewer)
	//outputRowViewer, ok := outputs.(mat64.RowViewer)

	// TODO: Move this to its own function
	var f func(start, end int, c chan lossDerivStruct, parameters []float64)

	switch {
	default:
		panic("Shouldn't be here")
	case cacheFeatures:
		f = func(start, end int, c chan lossDerivStruct, parameters []float64) {
			lossDeriver := g.t.NewLossDeriver()
			prediction := make([]float64, g.outputDim)
			dLossDPred := make([]float64, g.outputDim)
			dLossDWeight := make([]float64, g.nParameters)
			totalDLossDWeight := make([]float64, g.nParameters)
			var loss float64
			output := make([]float64, g.outputDim)
			for i := start; i < end; i++ {
				// Compute the prediction
				lossDeriver.Predict(parameters, g.features.RawRowView(i), prediction)
				// Compute the loss

				g.outputs.Row(output, i)
				loss += g.losser.LossDeriv(prediction, output, dLossDPred)
				// Compute the derivative
				lossDeriver.Deriv(parameters, g.features.RawRowView(i), prediction, dLossDPred, dLossDWeight)

				floats.Add(totalDLossDWeight, dLossDWeight)
			}
			// Send the value back on the channel
			c <- lossDerivStruct{
				loss:  loss,
				deriv: totalDLossDWeight,
			}
		}
	case !cacheFeatures:
		f = func(start, end int, c chan lossDerivStruct, parameters []float64) {
			lossDeriver := g.t.NewLossDeriver()
			prediction := make([]float64, g.outputDim)
			dLossDPred := make([]float64, g.outputDim)
			dLossDWeight := make([]float64, g.nParameters)
			totalDLossDWeight := make([]float64, g.nParameters)
			var loss float64
			output := make([]float64, g.outputDim)

			input := make([]float64, g.inputDim)

			features := make([]float64, g.t.NumFeatures())

			featurizer := g.t.NewFeaturizer()
			for i := start; i < end; i++ {

				g.inputs.Row(input, i)

				featurizer.Featurize(input, features)

				// Compute the prediction
				lossDeriver.Predict(parameters, features, prediction)
				// Compute the loss
				g.outputs.Row(output, i)

				loss += g.losser.LossDeriv(prediction, output, dLossDPred)

				// Compute the derivative
				lossDeriver.Deriv(parameters, features, prediction, dLossDPred, dLossDWeight)

				// Add to the total derivative
				floats.Add(totalDLossDWeight, dLossDWeight)

				// Send the value back on the channel
				c <- lossDerivStruct{
					loss:  loss,
					deriv: totalDLossDWeight,
				}
			}
		}
	}

	g.lossDerivFunc = f

	return g
}
예제 #6
0
파일: predict.go 프로젝트: reggo/reggo
func BatchPredict(batch BatchPredictor, inputs common.RowMatrix, outputs common.MutableRowMatrix, inputDim, outputDim int, grainSize int) (common.MutableRowMatrix, error) {

	// TODO: Add in something about error

	// Check that the inputs and outputs are the right sizes
	nSamples, dimInputs := inputs.Dims()
	if inputDim != dimInputs {
		return outputs, errors.New("predict batch: input dimension mismatch")
	}

	if outputs == nil {
		outputs = mat64.NewDense(nSamples, outputDim, nil)
	} else {
		nOutputSamples, dimOutputs := outputs.Dims()
		if dimOutputs != outputDim {
			return outputs, errors.New("predict batch: output dimension mismatch")
		}
		if nSamples != nOutputSamples {
			return outputs, errors.New("predict batch: rows mismatch")
		}
	}

	// Perform predictions in parallel. For each parallel call, form a new predictor so that
	// memory allocations are saved and no race condition happens.

	// If the input and/or output is a RowViewer, save time by avoiding a copy
	inputRVer, inputIsRowViewer := inputs.(mat64.RawRowViewer)
	outputRVer, outputIsRowViewer := outputs.(mat64.RawRowViewer)

	var f func(start, end int)

	// wrapper function to allow parallel prediction. Uses RowView if the type has it
	switch {
	default:
		panic("Shouldn't be here")
	case inputIsRowViewer, outputIsRowViewer:
		f = func(start, end int) {
			p := batch.NewPredictor()
			for i := start; i < end; i++ {
				p.Predict(inputRVer.RawRowView(i), outputRVer.RawRowView(i))
			}
		}

	case inputIsRowViewer && !outputIsRowViewer:
		f = func(start, end int) {
			p := batch.NewPredictor()
			output := make([]float64, outputDim)
			for i := start; i < end; i++ {
				outputs.Row(output, i)
				p.Predict(inputRVer.RawRowView(i), output)
				outputs.SetRow(i, output)
			}
		}
	case !inputIsRowViewer && outputIsRowViewer:
		f = func(start, end int) {
			p := batch.NewPredictor()
			input := make([]float64, inputDim)
			for i := start; i < end; i++ {
				inputs.Row(input, i)
				p.Predict(input, outputRVer.RawRowView(i))
			}
		}
	case !inputIsRowViewer && !outputIsRowViewer:
		f = func(start, end int) {
			p := batch.NewPredictor()
			input := make([]float64, inputDim)
			output := make([]float64, outputDim)
			for i := start; i < end; i++ {
				inputs.Row(input, i)
				outputs.Row(output, i)
				p.Predict(input, output)
				outputs.SetRow(i, output)
			}
		}
	}

	common.ParallelFor(nSamples, grainSize, f)
	return outputs, nil
}
예제 #7
0
파일: linsolve.go 프로젝트: reggo/reggo
// LinearSolve trains a Linear algorithm.
// Assumes inputs and outputs are already scaled
// If features is nil will call featurize
// Will return nil if regularizer is not a linear regularizer
// Is destructive if any of the weights are zero
// Losser is always the two-norm
// Does not set the value of the parameters (in case this is called in parallel with a different routine)
func LinearSolve(linearTrainable LinearTrainable, features *mat64.Dense, inputs, trueOutputs common.RowMatrix,
	weights []float64, regularizer regularize.Regularizer) (parameters []float64) {
	// TODO: Allow tikhonov regularization
	// TODO: Add test for weights
	// TODO: Need to do something about returning a []float64

	if !IsLinearSolveRegularizer(regularizer) {
		return nil
	}

	if features == nil {
		features = FeaturizeTrainable(linearTrainable, inputs, features)
	}

	_, nFeatures := features.Dims()

	var weightedFeatures, weightedOutput *mat64.Dense

	fmt.Println("In linear solve")

	if weights != nil {
		panic("Need functionality to be better. Either banded special case in matrix or do the mulitplication by hand")
		scaledWeight := make([]float64, len(weights))
		for i, weight := range weights {
			scaledWeight[i] = math.Sqrt(weight)
		}
		diagWeight := diagonal.NewDiagonal(len(scaledWeight), scaledWeight)

		nSamples, outputDim := trueOutputs.Dims()
		weightedOutput = mat64.NewDense(nSamples, outputDim, nil)
		weightedFeatures = mat64.NewDense(nSamples, nFeatures, nil)

		weightedOutput.Copy(trueOutputs)
		weightedFeatures.Copy(features)

		// TODO: Replace this with better than mat multiply
		weightedOutput.Mul(diagWeight, weightedOutput)
		weightedFeatures.Mul(diagWeight, weightedFeatures)
	}

	switch regularizer.(type) {
	case nil:
	case regularize.None:
	default:
		panic("Shouldn't be here. Must be error in IsLinearRegularizer")
	}

	if weights == nil {
		parameterMat, err := mat64.Solve(features, trueOutputs)
		if err != nil {
			panic(err)
		}
		return parameterMat.RawMatrix().Data

	}
	parameterMat, err := mat64.Solve(weightedFeatures, weightedOutput)
	if err != nil {
		panic(err)
	}

	return parameterMat.RawMatrix().Data
}