// DCT returns DCT-II coefficients given an input signal.
func DCT(x []float64) []float64 {
n := len(x)
nh := n / 2
evenVector := make([]float64, n)
for i := 0; i < nh; i++ {
evenVector[i], evenVector[n-1-i] = x[2*i], x[2*i+1]
}
array := fft.FFTReal(evenVector)
theta := math.Pi / (2.0 * float64(n))
// 1/4 Shift
for k := 1; k < nh; k++ {
w := cmplx.Exp(complex(0, -float64(k)*theta))
wCont := -complex(imag(w), real(w))
array[k] *= w
array[n-k] *= wCont
}
array[nh] *= complex(math.Cos(theta*float64(nh)), 0.0)
dctCoef := make([]float64, n)
for i := range dctCoef {
dctCoef[i] = real(array[i])
}
return dctCoef
}
// Compute the spectrogram of a signal with window width nfft and overlap
// percentage overlap.
func Compute(signal []float64, nfft int, overlap float64) Spectrogram {
spectra := make([][]complex128, 0)
start := 0
end := start + nfft
off := int((1 - overlap) * float64(nfft))
// pre-compute the window function
window := window.Hamming(nfft)
// pre-allocate buffer to hold window * signal
windowed := make([]float64, nfft)
for end < len(signal) {
for i, s := range signal[start:end] {
windowed[i] = s * window[i]
}
spectrum := fft.FFTReal(windowed)
// FIXME: memory is still used for the entire spectrum since the GC doesn't
// understand and can't free internal and partial pointers, to free it must
// do a copy. If this spectrum persists for a while then we should do a
// copy().
spectra = append(spectra, spectrum[0:len(spectrum)/2])
start += off
end += off
}
spectrogram := Spectrogram{spectra, nil}
return spectrogram
}
func SamplesToFrequencies(samples *TMSnippet) Frequencies {
var freqs Frequencies
f64samples := make([]float64, len(samples.Samples))
for i, v := range samples.Samples {
f64samples[i] = float64(v)
}
freqs.Freqs = cplxsToReals(fft.FFTReal(f64samples))
freqs.DeltaHz = 1 / samples.SampleHz // TODO: check if this even makes sense
return freqs
}
func buildFFT(data tsData) [][]float64 {
x := make([]float64, len(data.Values))
for i, d := range data.Values {
x[i] = d.Value
}
X := fft.FFTReal(x)
Xout := make([][]float64, len(X))
for i, Xi := range X {
Xout[i] = make([]float64, 2)
Xout[i][0] = real(Xi)
Xout[i][1] = imag(Xi)
}
return Xout
}
// NewSpectrogram constructs a spectrogram
// windowLen should be a power of 2
func NewSpectrogram(r io.Reader, windowLen, overlap int) (*Spectrogram, error) {
data, err := wav.ReadWav(r)
if err != nil {
return &Spectrogram{}, err
}
if expected, actual := uint16(16), data.BitsPerSample; expected != actual {
return &Spectrogram{}, fmt.Errorf("rate %d, not %d", actual, expected)
}
if expected, actual := uint16(1), data.NumChannels; expected != actual {
return &Spectrogram{}, fmt.Errorf("%d chans, not %d", actual, expected)
}
// find maximum energy
max := -math.MaxFloat64
for _, value := range data.Data {
if max < float64(value[0]) {
max = float64(value[0])
}
}
sampleData := make([]float64, data.NumSamples)
for i := range data.Data {
sampleData[i] = float64(data.Data[i][0]) / max
}
// create hammed windows for the fft
windows := spectral.Segment(sampleData, windowLen, overlap)
nWindow := len(windows) // (len(sampleData) - windowLen) / (windowLen - overlap) + 1
ApplyHamming(windows)
// do the fft, get positive fq components, and discard phase
s := Spectrogram{
data: make([][]float64, nWindow),
sampleRate: data.SampleRate,
windowLength: windowLen,
overlap: overlap,
numSamples: data.NumSamples,
}
for i, w := range windows {
s.data[i] = make([]float64, int(windowLen/2))
for j, v := range fft.FFTReal(w)[0:int(windowLen/2)] {
s.data[i][j] = 2 * cmplx.Abs(v)
}
}
return &s, nil
}
// STFT returns complex spectrogram given an input signal.
func (s *STFT) STFT(input []float64) [][]complex128 {
numFrames := s.NumFrames(input)
spectrogram := make([][]complex128, numFrames)
frames := s.DivideFrames(input)
for i, frame := range frames {
// Windowing
windowed := window.Windowing(frame, s.Window)
// Complex Spectrum
spectrogram[i] = fft.FFTReal(windowed)
}
return spectrogram
}
func (cq *ConstantQ) processOctaveBlock(octave int) [][]complex128 {
apf := cq.kernel.Properties.atomsPerFrame
bpo := cq.kernel.Properties.binsPerOctave
fftHop := cq.kernel.Properties.fftHop
fftSize := cq.kernel.Properties.fftSize
cv := fft.FFTReal(cq.buffers[octave][:fftSize])
cq.buffers[octave] = cq.buffers[octave][fftHop:]
cqrowvec := cq.kernel.processForward(cv)
// Reform into a column matrix
cqblock := make([][]complex128, apf, apf)
for j := 0; j < apf; j++ {
cqblock[j] = make([]complex128, bpo, bpo)
for i := 0; i < bpo; i++ {
cqblock[j][i] = cqrowvec[i*apf+j]
}
}
return cqblock
}
func fftanalyzer(values chan []int32) {
for {
in := <-values
var buf [2048]float64
for k := 0; k <= (len(in)/2)-2; k++ {
v1 := float64(in[k*2]) * 0.5 * (1.0 - math.Cos(2.0*math.Pi*float64(k*2)/(float64(len(in))-1.0)))
v2 := float64(in[k*2+1]) * 0.5 * (1.0 - math.Cos(2.0*math.Pi*float64(k*2)/(float64(len(in))-1.0)))
buf[k] = v1 + v2
}
buffer := fft.FFTReal(buf[0:2048])
for i, val := range buffer {
buf[i] = cmplx.Abs(val)
}
var sendBuf [64]float64
sendBuf[0] = fftAvg(buf[0:len(buf)-1], 0, 30)
sendBuf[1] = fftAvg(buf[0:len(buf)-1], 31, 60)
sendBuf[2] = fftAvg(buf[0:len(buf)-1], 61, 100)
sendBuf[3] = fftAvg(buf[0:len(buf)-1], 101, 150)
sendBuf[4] = fftAvg(buf[0:len(buf)-1], 151, 200)
sendBuf[5] = fftAvg(buf[0:len(buf)-1], 201, 250)
sendBuf[6] = fftAvg(buf[0:len(buf)-1], 251, 300)
sendBuf[7] = fftAvg(buf[0:len(buf)-1], 301, 350)
sendBuf[8] = fftAvg(buf[0:len(buf)-1], 351, 400)
for n := 9; n < 64; n++ {
sendBuf[n] = fftAvg(buf[0:len(buf)-1], (351 + (250 * (n - 9))), (500 + (250 * (n - 9))))
}
send := make([]int64, 64)
x := 8.0
y := 3.0
for h, v := range sendBuf {
q := v * (math.Log(x) / y)
x = x + (x)
send[h] = int64(q) / math.MaxInt32
}
fft_values <- send
}
}
func KCMmeans(rawData [][]float64, k int, distanceFunction DistanceFunction, threshold int) ([]int, []Observation, error) {
var minClusteredData []ClusteredObservation
var means []Observation
var err error
max, trace := int64(0), make([]float64, k)
for clusters := 1; clusters <= k; clusters++ {
data := make([]ClusteredObservation, len(rawData))
for ii, jj := range rawData {
data[ii].Observation = jj
}
seeds := seed(data, clusters, distanceFunction)
clusteredData, _ := kmeans(data, seeds, distanceFunction, threshold)
counts := make([]int, clusters)
for _, jj := range clusteredData {
counts[jj.ClusterNumber]++
}
input := &bytes.Buffer{}
for c := 0; c < clusters; c++ {
/*err := binary.Write(input, binary.LittleEndian, rand.Float64())
if err != nil {
panic(err)
}*/
err := binary.Write(input, binary.LittleEndian, int64(counts[c]))
if err != nil {
panic(err)
}
for _, jj := range seeds[c] {
err = binary.Write(input, binary.LittleEndian, jj)
if err != nil {
panic(err)
}
}
for _, j := range clusteredData {
if j.ClusterNumber == c {
for ii, jj := range j.Observation {
err = binary.Write(input, binary.LittleEndian, jj-seeds[c][ii])
if err != nil {
panic(err)
}
}
}
}
}
in, output := make(chan []byte, 1), &bytes.Buffer{}
in <- input.Bytes()
close(in)
compress.BijectiveBurrowsWheelerCoder(in).MoveToFrontRunLengthCoder().AdaptiveCoder().Code(output)
/*output := &bytes.Buffer{}
writer := lzma.NewWriterLevel(output, lzma.BestCompression)
writer.Write(input.Bytes())
writer.Close()*/
complexity := int64(output.Len())
trace[clusters-1] = float64(complexity)
fmt.Printf("%v %v\n", clusters, complexity)
if complexity > max {
max, minClusteredData, means = complexity, clusteredData, make([]Observation, len(seeds))
for ii := range seeds {
means[ii] = make([]float64, len(seeds[ii]))
for jj := range seeds[ii] {
means[ii][jj] = seeds[ii][jj]
}
}
}
}
f := fft.FFTReal(trace)
points, phase, complex := make(plotter.XYs, len(f)-1), make(plotter.XYs, len(f)-1), make(plotter.XYs, len(f))
for i, j := range f[1:] {
points[i].X, points[i].Y = float64(i), cmplx.Abs(j)
phase[i].X, phase[i].Y = float64(i), cmplx.Phase(j)
complex[i].X, complex[i].Y = real(j), imag(j)
}
p, err := plot.New()
if err != nil {
panic(err)
}
p.Title.Text = "FFT Real"
p.X.Label.Text = "X"
p.Y.Label.Text = "Y"
scatter, err := plotter.NewScatter(points)
if err != nil {
panic(err)
}
scatter.Shape = draw.CircleGlyph{}
scatter.Radius = vg.Points(1)
p.Add(scatter)
if err := p.Save(8, 8, "fft_real.png"); err != nil {
panic(err)
}
p, err = plot.New()
if err != nil {
panic(err)
}
p.Title.Text = "FFT Phase"
p.X.Label.Text = "X"
p.Y.Label.Text = "Y"
scatter, err = plotter.NewScatter(phase)
if err != nil {
panic(err)
}
scatter.Shape = draw.CircleGlyph{}
scatter.Radius = vg.Points(1)
scatter.Color = color.RGBA{0, 0, 255, 255}
p.Add(scatter)
if err := p.Save(8, 8, "fft_phase.png"); err != nil {
panic(err)
}
p, err = plot.New()
if err != nil {
panic(err)
}
p.Title.Text = "FFT Complex"
p.X.Label.Text = "X"
p.Y.Label.Text = "Y"
scatter, err = plotter.NewScatter(complex)
if err != nil {
panic(err)
}
scatter.Shape = draw.CircleGlyph{}
scatter.Radius = vg.Points(1)
scatter.Color = color.RGBA{0, 0, 255, 255}
p.Add(scatter)
if err := p.Save(8, 8, "fft_complex.png"); err != nil {
panic(err)
}
labels := make([]int, len(minClusteredData))
for ii, jj := range minClusteredData {
labels[ii] = jj.ClusterNumber
}
return labels, means, err
}
// Pwelch estimates the power spectral density of x using Welch's method.
// Fs is the sampling frequency (samples per time unit) of x. Fs is used
// to calculate freqs.
// Returns the power spectral density Pxx and corresponding frequencies freqs.
// Designed to be similar to the matplotlib implementation below.
// Reference: http://matplotlib.org/api/mlab_api.html#matplotlib.mlab.psd
// See also: http://www.mathworks.com/help/signal/ref/pwelch.html
func Pwelch(x []float64, Fs float64, o *PwelchOptions) (Pxx, freqs []float64) {
if len(x) == 0 {
return []float64{}, []float64{}
}
nfft := o.NFFT
pad := o.Pad
noverlap := o.Noverlap
wf := o.Window
enable_scaling := !o.Scale_off
if nfft == 0 {
nfft = 256
}
if wf == nil {
wf = window.Hann
}
if pad == 0 {
pad = nfft
}
if len(x) < nfft {
x = dsputils.ZeroPadF(x, nfft)
}
lp := pad/2 + 1
var scale float64 = 2
segs := Segment(x, nfft, noverlap)
Pxx = make([]float64, lp)
for _, x := range segs {
x = dsputils.ZeroPadF(x, pad)
window.Apply(x, wf)
pgram := fft.FFTReal(x)
for j := range Pxx {
d := real(cmplx.Conj(pgram[j])*pgram[j]) / float64(len(segs))
if j > 0 && j < lp-1 {
d *= scale
}
Pxx[j] += d
}
}
w := wf(nfft)
var norm float64
for _, x := range w {
norm += math.Pow(x, 2)
}
if enable_scaling {
norm *= Fs
}
for i := range Pxx {
Pxx[i] /= norm
}
freqs = make([]float64, lp)
coef := Fs / float64(pad)
for i := range freqs {
freqs[i] = float64(i) * coef
}
return
}
// Run is the block's main loop. Here we listen on the different channels we set up.
func (b *Timeseries) Run() {
var err error
//var path, lagStr string
var path string
var tree *jee.TokenTree
//var lag time.Duration
var data tsData
var numSamples float64
// defaults
numSamples = 1
for {
select {
case ruleI := <-b.inrule:
// set a parameter of the block
rule, ok := ruleI.(map[string]interface{})
if !ok {
b.Error(errors.New("could not assert rule to map"))
}
path, err = util.ParseString(rule, "Path")
if err != nil {
b.Error(err)
continue
}
tree, err = util.BuildTokenTree(path)
if err != nil {
b.Error(err)
continue
}
/*
lagStr, err = util.ParseString(rule, "Lag")
if err != nil {
b.Error(err)
continue
}
lag, err = time.ParseDuration(lagStr)
if err != nil {
b.Error(err)
continue
}
*/
numSamples, err = util.ParseFloat(rule, "NumSamples")
if err != nil {
b.Error(err)
continue
}
data = tsData{
Values: make([]tsDataPoint, int(numSamples)),
}
case <-b.quit:
// quit * time.Second the block
return
case msg := <-b.in:
if tree == nil {
continue
}
if data.Values == nil {
continue
}
// deal with inbound data
v, err := jee.Eval(tree, msg)
if err != nil {
b.Error(err)
continue
}
var val float64
switch v := v.(type) {
case float32:
val = float64(v)
case int:
val = float64(v)
case float64:
val = v
}
//t := float64(time.Now().Add(-lag).Unix())
t := float64(time.Now().Unix())
d := tsDataPoint{
Timestamp: t,
Value: val,
}
data.Values = append(data.Values[1:], d)
case respChan := <-b.queryrule:
// deal with a query request
respChan <- map[string]interface{}{
//"Window": lagStr,
"Path": path,
"NumSamples": numSamples,
}
case respChan := <-b.querystate:
out := map[string]interface{}{
"timeseries": data,
}
respChan <- out
case respChan := <-b.queryfft:
x := make([]float64, len(data.Values))
for i, d := range data.Values {
x[i] = d.Value
}
X := fft.FFTReal(x)
Xout := make([][]float64, len(X))
for i, Xi := range X {
Xout[i] = make([]float64, 2)
Xout[i][0] = real(Xi)
Xout[i][1] = imag(Xi)
}
out := map[string]interface{}{
"fft": Xout,
}
respChan <- out
case <-b.inpoll:
outArray := make([]interface{}, len(data.Values))
for i, d := range data.Values {
di := map[string]interface{}{
"timestamp": d.Timestamp,
"value": d.Value,
}
outArray[i] = di
}
out := map[string]interface{}{
"timeseries": outArray,
}
b.out <- out
}
}
}
// Return the real and imaginary parts of the fft
func (d *DataFrame) FFT() []complex128 {
data := d.Data()
return fft.FFTReal(data)
}
// Cepstrum2Amplitude return log amplitude spectrum from cepstrum.
// Note: This function requires symmetric cepstrum as the argument.
func Cepstrum2LogAmplitude(ceps []float64) []float64 {
return gossp.ToReal(fft.FFTReal(ceps))
}
