Be aware: Like a number of prior ones, this publish is an excerpt from the forthcoming e book, Deep Studying and Scientific Computing with R torch. And like many excerpts, it’s a product of arduous trade-offs. For added depth and extra examples, I’ve to ask you to please seek the advice of the e book.

Wavelets and the Wavelet Remodel

What are wavelets? Just like the Fourier foundation, they’re capabilities; however they don’t prolong infinitely. As a substitute, they’re localized in time: Away from the middle, they shortly decay to zero. Along with a location parameter, in addition they have a scale: At totally different scales, they seem squished or stretched. Squished, they’ll do higher at detecting excessive frequencies; the converse applies after they’re stretched out in time.

The essential operation concerned within the Wavelet Remodel is convolution – have the (flipped) wavelet slide over the information, computing a sequence of dot merchandise. This manner, the wavelet is mainly in search of similarity.

As to the wavelet capabilities themselves, there are a lot of of them. In a sensible software, we’d need to experiment and choose the one which works finest for the given knowledge. In comparison with the DFT and spectrograms, extra experimentation tends to be concerned in wavelet evaluation.

The subject of wavelets may be very totally different from that of Fourier transforms in different respects, as nicely. Notably, there’s a lot much less standardization in terminology, use of symbols, and precise practices. On this introduction, I’m leaning closely on one particular exposition, the one in Arnt Vistnes’ very good e book on waves (Vistnes 2018). In different phrases, each terminology and examples replicate the alternatives made in that e book.

Introducing the Morlet wavelet

The Morlet, also referred to as Gabor, wavelet is outlined like so:

[
Psi_{omega_{a},K,t_{k}}(t_n) = (e^{-i omega_{a} (t_n – t_k)} – e^{-K^2}) e^{- omega_a^2 (t_n – t_k )^2 /(2K )^2}
]

This formulation pertains to discretized knowledge, the varieties of knowledge we work with in apply. Thus, (t_k) and (t_n) designate time limits, or equivalently, particular person time-series samples.

This equation appears daunting at first, however we are able to “tame” it a bit by analyzing its construction, and pointing to the primary actors. For concreteness, although, we first take a look at an instance wavelet.

We begin by implementing the above equation:

Evaluating code and mathematical formulation, we discover a distinction. The operate itself takes one argument, (t_n); its realization, 4 (omega, Okay, t_k, and t). It is because the torch code is vectorized: On the one hand, omega, Okay, and t_k, which, within the method, correspond to (omega_{a}), (Okay), and (t_k) , are scalars. (Within the equation, they’re assumed to be fastened.) t, however, is a vector; it’s going to maintain the measurement occasions of the sequence to be analyzed.

We choose instance values for omega, Okay, and t_k, in addition to a variety of occasions to judge the wavelet on, and plot its values:

omega <- 6 * pi
Okay <- 6
t_k <- 5
 
sample_time <- torch_arange(3, 7, 0.0001)

create_wavelet_plot <- operate(omega, Okay, t_k, sample_time) {
  morlet <- morlet(omega, Okay, t_k, sample_time)
  df <- knowledge.body(
    x = as.numeric(sample_time),
    actual = as.numeric(morlet$actual),
    imag = as.numeric(morlet$imag)
  ) %>%
    pivot_longer(-x, names_to = "half", values_to = "worth")
  ggplot(df, aes(x = x, y = worth, coloration = half)) +
    geom_line() +
    scale_colour_grey(begin = 0.8, finish = 0.4) +
    xlab("time") +
    ylab("wavelet worth") +
    ggtitle("Morlet wavelet",
      subtitle = paste0("ω_a = ", omega / pi, "π , Okay = ", Okay)
    ) +
    theme_minimal()
}

create_wavelet_plot(omega, Okay, t_k, sample_time)

What we see here’s a advanced sine curve – be aware the true and imaginary elements, separated by a part shift of (pi/2) – that decays on each side of the middle. Wanting again on the equation, we are able to determine the components liable for each options. The primary time period within the equation, (e^{-i omega_{a} (t_n – t_k)}), generates the oscillation; the third, (e^{- omega_a^2 (t_n – t_k )^2 /(2K )^2}), causes the exponential decay away from the middle. (In case you’re questioning in regards to the second time period, (e^{-Okay^2}): For given (Okay), it’s only a fixed.)

The third time period truly is a Gaussian, with location parameter (t_k) and scale (Okay). We’ll speak about (Okay) in nice element quickly, however what’s with (t_k)? (t_k) is the middle of the wavelet; for the Morlet wavelet, that is additionally the placement of most amplitude. As distance from the middle will increase, values shortly method zero. That is what is supposed by wavelets being localized: They’re “lively” solely on a brief vary of time.

The roles of (Okay) and (omega_a)

Now, we already stated that (Okay) is the size of the Gaussian; it thus determines how far the curve spreads out in time. However there may be additionally (omega_a). Wanting again on the Gaussian time period, it, too, will influence the unfold.

First although, what’s (omega_a)? The subscript (a) stands for “evaluation”; thus, (omega_a) denotes a single frequency being probed.

Now, let’s first examine visually the respective impacts of (omega_a) and (Okay).

p1 <- create_wavelet_plot(6 * pi, 4, 5, sample_time)
p2 <- create_wavelet_plot(6 * pi, 6, 5, sample_time)
p3 <- create_wavelet_plot(6 * pi, 8, 5, sample_time)
p4 <- create_wavelet_plot(4 * pi, 6, 5, sample_time)
p5 <- create_wavelet_plot(6 * pi, 6, 5, sample_time)
p6 <- create_wavelet_plot(8 * pi, 6, 5, sample_time)

(p1 | p4) /
  (p2 | p5) /
  (p3 | p6)

Within the left column, we preserve (omega_a) fixed, and differ (Okay). On the suitable, (omega_a) modifications, and (Okay) stays the identical.

Firstly, we observe that the upper (Okay), the extra the curve will get unfold out. In a wavelet evaluation, which means that extra time limits will contribute to the rework’s output, leading to excessive precision as to frequency content material, however lack of decision in time. (We’ll return to this – central – trade-off quickly.)

As to (omega_a), its influence is twofold. On the one hand, within the Gaussian time period, it counteracts – precisely, even – the size parameter, (Okay). On the opposite, it determines the frequency, or equivalently, the interval, of the wave. To see this, check out the suitable column. Equivalent to the totally different frequencies, we’ve, within the interval between 4 and 6, 4, six, or eight peaks, respectively.

This double position of (omega_a) is the rationale why, all-in-all, it does make a distinction whether or not we shrink (Okay), protecting (omega_a) fixed, or enhance (omega_a), holding (Okay) fastened.

This state of issues sounds difficult, however is much less problematic than it might sound. In apply, understanding the position of (Okay) is vital, since we have to choose smart (Okay) values to attempt. As to the (omega_a), however, there might be a mess of them, comparable to the vary of frequencies we analyze.

So we are able to perceive the influence of (Okay) in additional element, we have to take a primary take a look at the Wavelet Remodel.

Wavelet Remodel: An easy implementation

Whereas general, the subject of wavelets is extra multifaceted, and thus, could appear extra enigmatic than Fourier evaluation, the rework itself is simpler to understand. It’s a sequence of native convolutions between wavelet and sign. Right here is the method for particular scale parameter (Okay), evaluation frequency (omega_a), and wavelet location (t_k):

[
W_{K, omega_a, t_k} = sum_n x_n Psi_{omega_{a},K,t_{k}}^*(t_n)
]

That is only a dot product, computed between sign and complex-conjugated wavelet. (Right here advanced conjugation flips the wavelet in time, making this convolution, not correlation – a undeniable fact that issues rather a lot, as you’ll see quickly.)

Correspondingly, easy implementation ends in a sequence of dot merchandise, every comparable to a special alignment of wavelet and sign. Beneath, in wavelet_transform(), arguments omega and Okay are scalars, whereas x, the sign, is a vector. The result’s the wavelet-transformed sign, for some particular Okay and omega of curiosity.

wavelet_transform <- operate(x, omega, Okay) {
  n_samples <- dim(x)[1]
  W <- torch_complex(
    torch_zeros(n_samples), torch_zeros(n_samples)
  )
  for (i in 1:n_samples) {
    # transfer heart of wavelet
    t_k <- x[i, 1]
    m <- morlet(omega, Okay, t_k, x[, 1])
    # compute native dot product
    # be aware wavelet is conjugated
    dot <- torch_matmul(
      m$conj()$unsqueeze(1),
      x[, 2]$to(dtype = torch_cfloat())
    )
    W[i] <- dot
  }
  W
}

To check this, we generate a easy sine wave that has a frequency of 100 Hertz in its first half, and double that within the second.

gencos <- operate(amp, freq, part, fs, period) {
  x <- torch_arange(0, period, 1 / fs)[1:-2]$unsqueeze(2)
  y <- amp * torch_cos(2 * pi * freq * x + part)
  torch_cat(checklist(x, y), dim = 2)
}

# sampling frequency
fs <- 8000

f1 <- 100
f2 <- 200
part <- 0
period <- 0.25

s1 <- gencos(1, f1, part, fs, period)
s2 <- gencos(1, f2, part, fs, period)

s3 <- torch_cat(checklist(s1, s2), dim = 1)
s3[(dim(s1)[1] + 1):(dim(s1)[1] * 2), 1] <-
  s3[(dim(s1)[1] + 1):(dim(s1)[1] * 2), 1] + period

df <- knowledge.body(
  x = as.numeric(s3[, 1]),
  y = as.numeric(s3[, 2])
)
ggplot(df, aes(x = x, y = y)) +
  geom_line() +
  xlab("time") +
  ylab("amplitude") +
  theme_minimal()

Now, we run the Wavelet Remodel on this sign, for an evaluation frequency of 100 Hertz, and with a Okay parameter of two, discovered by way of fast experimentation:

Okay <- 2
omega <- 2 * pi * f1

res <- wavelet_transform(x = s3, omega, Okay)
df <- knowledge.body(
  x = as.numeric(s3[, 1]),
  y = as.numeric(res$abs())
)

ggplot(df, aes(x = x, y = y)) +
  geom_line() +
  xlab("time") +
  ylab("Wavelet Remodel") +
  theme_minimal()

The rework accurately picks out the a part of the sign that matches the evaluation frequency. For those who really feel like, you may need to double-check what occurs for an evaluation frequency of 200 Hertz.

Now, in actuality we’ll need to run this evaluation not for a single frequency, however a variety of frequencies we’re focused on. And we’ll need to attempt totally different scales Okay. Now, in case you executed the code above, you could be apprehensive that this might take a lot of time.

Effectively, it by necessity takes longer to compute than its Fourier analogue, the spectrogram. For one, that’s as a result of with spectrograms, the evaluation is “simply” two-dimensional, the axes being time and frequency. With wavelets there are, as well as, totally different scales to be explored. And secondly, spectrograms function on complete home windows (with configurable overlap); a wavelet, however, slides over the sign in unit steps.

Nonetheless, the scenario is just not as grave because it sounds. The Wavelet Remodel being a convolution, we are able to implement it within the Fourier area as an alternative. We’ll do this very quickly, however first, as promised, let’s revisit the subject of various Okay.

Decision in time versus in frequency

We already noticed that the upper Okay, the extra spread-out the wavelet. We are able to use our first, maximally easy, instance, to analyze one speedy consequence. What, for instance, occurs for Okay set to twenty?

Okay <- 20

res <- wavelet_transform(x = s3, omega, Okay)
df <- knowledge.body(
  x = as.numeric(s3[, 1]),
  y = as.numeric(res$abs())
)

ggplot(df, aes(x = x, y = y)) +
  geom_line() +
  xlab("time") +
  ylab("Wavelet Remodel") +
  theme_minimal()

The Wavelet Remodel nonetheless picks out the right area of the sign – however now, as an alternative of a rectangle-like end result, we get a considerably smoothed model that doesn’t sharply separate the 2 areas.

Notably, the primary 0.05 seconds, too, present appreciable smoothing. The bigger a wavelet, the extra element-wise merchandise might be misplaced on the finish and the start. It is because transforms are computed aligning the wavelet in any respect sign positions, from the very first to the final. Concretely, once we compute the dot product at location t_k = 1, only a single pattern of the sign is taken into account.

Other than presumably introducing unreliability on the boundaries, how does wavelet scale have an effect on the evaluation? Effectively, since we’re correlating (convolving, technically; however on this case, the impact, ultimately, is similar) the wavelet with the sign, point-wise similarity is what issues. Concretely, assume the sign is a pure sine wave, the wavelet we’re utilizing is a windowed sinusoid just like the Morlet, and that we’ve discovered an optimum Okay that properly captures the sign’s frequency. Then another Okay, be it bigger or smaller, will lead to much less point-wise overlap.

Performing the Wavelet Remodel within the Fourier area

Quickly, we’ll run the Wavelet Remodel on an extended sign. Thus, it’s time to pace up computation. We already stated that right here, we profit from time-domain convolution being equal to multiplication within the Fourier area. The general course of then is that this: First, compute the DFT of each sign and wavelet; second, multiply the outcomes; third, inverse-transform again to the time area.

The DFT of the sign is shortly computed:

F <- torch_fft_fft(s3[ , 2])

With the Morlet wavelet, we don’t even need to run the FFT: Its Fourier-domain illustration could be said in closed type. We’ll simply make use of that formulation from the outset. Right here it’s:

morlet_fourier <- operate(Okay, omega_a, omega) {
  2 * (torch_exp(-torch_square(
    Okay * (omega - omega_a) / omega_a
  )) -
    torch_exp(-torch_square(Okay)) *
      torch_exp(-torch_square(Okay * omega / omega_a)))
}

Evaluating this assertion of the wavelet to the time-domain one, we see that – as anticipated – as an alternative of parameters t and t_k it now takes omega and omega_a. The latter, omega_a, is the evaluation frequency, the one we’re probing for, a scalar; the previous, omega, the vary of frequencies that seem within the DFT of the sign.

In instantiating the wavelet, there may be one factor we have to pay particular consideration to. In FFT-think, the frequencies are bins; their quantity is set by the size of the sign (a size that, for its half, instantly will depend on sampling frequency). Our wavelet, however, works with frequencies in Hertz (properly, from a person’s perspective; since this unit is significant to us). What this implies is that to morlet_fourier, as omega_a we have to go not the worth in Hertz, however the corresponding FFT bin. Conversion is completed relating the variety of bins, dim(x)[1], to the sampling frequency of the sign, fs:

# once more search for 100Hz elements
omega <- 2 * pi * f1

# want the bin comparable to some frequency in Hz
omega_bin <- f1/fs * dim(s3)[1]

We instantiate the wavelet, carry out the Fourier-domain multiplication, and inverse-transform the end result:

Okay <- 3

m <- morlet_fourier(Okay, omega_bin, 1:dim(s3)[1])
prod <- F * m
reworked <- torch_fft_ifft(prod)

Placing collectively wavelet instantiation and the steps concerned within the evaluation, we’ve the next. (Be aware learn how to wavelet_transform_fourier, we now, conveniently, go within the frequency worth in Hertz.)

wavelet_transform_fourier <- operate(x, omega_a, Okay, fs) {
  N <- dim(x)[1]
  omega_bin <- omega_a / fs * N
  m <- morlet_fourier(Okay, omega_bin, 1:N)
  x_fft <- torch_fft_fft(x)
  prod <- x_fft * m
  w <- torch_fft_ifft(prod)
  w
}

We’ve already made vital progress. We’re prepared for the ultimate step: automating evaluation over a variety of frequencies of curiosity. This may lead to a three-dimensional illustration, the wavelet diagram.

Creating the wavelet diagram

Within the Fourier Remodel, the variety of coefficients we acquire will depend on sign size, and successfully reduces to half the sampling frequency. With its wavelet analogue, since anyway we’re doing a loop over frequencies, we’d as nicely resolve which frequencies to investigate.

Firstly, the vary of frequencies of curiosity could be decided working the DFT. The subsequent query, then, is about granularity. Right here, I’ll be following the advice given in Vistnes’ e book, which relies on the relation between present frequency worth and wavelet scale, Okay.

Iteration over frequencies is then applied as a loop:

wavelet_grid <- operate(x, Okay, f_start, f_end, fs) {
  # downsample evaluation frequency vary
  # as per Vistnes, eq. 14.17
  num_freqs <- 1 + log(f_end / f_start)/ log(1 + 1/(8 * Okay))
  freqs <- seq(f_start, f_end, size.out = flooring(num_freqs))
  
  reworked <- torch_zeros(
    num_freqs, dim(x)[1],
    dtype = torch_cfloat()
    )
  for(i in 1:num_freqs) {
    w <- wavelet_transform_fourier(x, freqs[i], Okay, fs)
    reworked[i, ] <- w
  }
  checklist(reworked, freqs)
}

Calling wavelet_grid() will give us the evaluation frequencies used, along with the respective outputs from the Wavelet Remodel.

Subsequent, we create a utility operate that visualizes the end result. By default, plot_wavelet_diagram() shows the magnitude of the wavelet-transformed sequence; it will possibly, nonetheless, plot the squared magnitudes, too, in addition to their sq. root, a way a lot really useful by Vistnes whose effectiveness we’ll quickly have alternative to witness.

The operate deserves a couple of additional feedback.

Firstly, identical as we did with the evaluation frequencies, we down-sample the sign itself, avoiding to recommend a decision that’s not truly current. The method, once more, is taken from Vistnes’ e book.

Then, we use interpolation to acquire a brand new time-frequency grid. This step could even be mandatory if we preserve the unique grid, since when distances between grid factors are very small, R’s picture() could refuse to just accept axes as evenly spaced.

Lastly, be aware how frequencies are organized on a log scale. This results in way more helpful visualizations.

plot_wavelet_diagram <- operate(x,
                                 freqs,
                                 grid,
                                 Okay,
                                 fs,
                                 f_end,
                                 sort = "magnitude") {
  grid <- change(sort,
    magnitude = grid$abs(),
    magnitude_squared = torch_square(grid$abs()),
    magnitude_sqrt = torch_sqrt(grid$abs())
  )

  # downsample time sequence
  # as per Vistnes, eq. 14.9
  new_x_take_every <- max(Okay / 24 * fs / f_end, 1)
  new_x_length <- flooring(dim(grid)[2] / new_x_take_every)
  new_x <- torch_arange(
    x[1],
    x[dim(x)[1]],
    step = x[dim(x)[1]] / new_x_length
  )
  
  # interpolate grid
  new_grid <- nnf_interpolate(
    grid$view(c(1, 1, dim(grid)[1], dim(grid)[2])),
    c(dim(grid)[1], new_x_length)
  )$squeeze()
  out <- as.matrix(new_grid)

  # plot log frequencies
  freqs <- log10(freqs)
  
  picture(
    x = as.numeric(new_x),
    y = freqs,
    z = t(out),
    ylab = "log frequency [Hz]",
    xlab = "time [s]",
    col = hcl.colours(12, palette = "Mild grays")
  )
  primary <- paste0("Wavelet Remodel, Okay = ", Okay)
  sub <- change(sort,
    magnitude = "Magnitude",
    magnitude_squared = "Magnitude squared",
    magnitude_sqrt = "Magnitude (sq. root)"
  )

  mtext(aspect = 3, line = 2, at = 0, adj = 0, cex = 1.3, primary)
  mtext(aspect = 3, line = 1, at = 0, adj = 0, cex = 1, sub)
}

Let’s use this on a real-world instance.

An actual-world instance: Chaffinch’s track

For the case research, I’ve chosen what, to me, was essentially the most spectacular wavelet evaluation proven in Vistnes’ e book. It’s a pattern of a chaffinch’s singing, and it’s out there on Vistnes’ web site.

url <- "http://www.physics.uio.no/pow/wavbirds/chaffinch.wav"

obtain.file(
 file.path(url),
 destfile = "/tmp/chaffinch.wav"
)

We use torchaudio to load the file, and convert from stereo to mono utilizing tuneR’s appropriately named mono(). (For the sort of evaluation we’re doing, there is no such thing as a level in protecting two channels round.)

library(torchaudio)
library(tuneR)

wav <- tuneR_loader("/tmp/chaffinch.wav")
wav <- mono(wav, "each")
wav

Wave Object
    Variety of Samples:      1864548
    Period (seconds):     42.28
    Samplingrate (Hertz):   44100
    Channels (Mono/Stereo): Mono
    PCM (integer format):   TRUE
    Bit (8/16/24/32/64):    16 

For evaluation, we don’t want the whole sequence. Helpfully, Vistnes additionally revealed a advice as to which vary of samples to investigate.

waveform_and_sample_rate <- transform_to_tensor(wav)
x <- waveform_and_sample_rate[[1]]$squeeze()
fs <- waveform_and_sample_rate[[2]]

# http://www.physics.uio.no/pow/wavbirds/chaffinchInfo.txt
begin <- 34000
N <- 1024 * 128
finish <- begin + N - 1
x <- x[start:end]

dim(x)

[1] 131072

How does this look within the time area? (Don’t miss out on the event to really hear to it, in your laptop computer.)

df <- knowledge.body(x = 1:dim(x)[1], y = as.numeric(x))
ggplot(df, aes(x = x, y = y)) +
  geom_line() +
  xlab("pattern") +
  ylab("amplitude") +
  theme_minimal()

Now, we have to decide an affordable vary of research frequencies. To that finish, we run the FFT:

On the x-axis, we plot frequencies, not pattern numbers, and for higher visibility, we zoom in a bit.

bins <- 1:dim(F)[1]
freqs <- bins / N * fs

# the bin, not the frequency
cutoff <- N/4

df <- knowledge.body(
  x = freqs[1:cutoff],
  y = as.numeric(F$abs())[1:cutoff]
)
ggplot(df, aes(x = x, y = y)) +
  geom_col() +
  xlab("frequency (Hz)") +
  ylab("magnitude") +
  theme_minimal()

Primarily based on this distribution, we are able to safely limit the vary of research frequencies to between, roughly, 1800 and 8500 Hertz. (That is additionally the vary really useful by Vistnes.)

First, although, let’s anchor expectations by making a spectrogram for this sign. Appropriate values for FFT dimension and window dimension had been discovered experimentally. And although, in spectrograms, you don’t see this achieved usually, I discovered that displaying sq. roots of coefficient magnitudes yielded essentially the most informative output.

fft_size <- 1024
window_size <- 1024
energy <- 0.5

spectrogram <- transform_spectrogram(
  n_fft = fft_size,
  win_length = window_size,
  normalized = TRUE,
  energy = energy
)

spec <- spectrogram(x)
dim(spec)

[1] 513 257

Like we do with wavelet diagrams, we plot frequencies on a log scale.

bins <- 1:dim(spec)[1]
freqs <- bins * fs / fft_size
log_freqs <- log10(freqs)

frames <- 1:(dim(spec)[2])
seconds <- (frames / dim(spec)[2])  * (dim(x)[1] / fs)

picture(x = seconds,
      y = log_freqs,
      z = t(as.matrix(spec)),
      ylab = 'log frequency [Hz]',
      xlab = 'time [s]',
      col = hcl.colours(12, palette = "Mild grays")
)
primary <- paste0("Spectrogram, window dimension = ", window_size)
sub <- "Magnitude (sq. root)"
mtext(aspect = 3, line = 2, at = 0, adj = 0, cex = 1.3, primary)
mtext(aspect = 3, line = 1, at = 0, adj = 0, cex = 1, sub)

The spectrogram already exhibits a particular sample. Let’s see what could be achieved with wavelet evaluation. Having experimented with a couple of totally different Okay, I agree with Vistnes that Okay = 48 makes for a wonderful alternative:

f_start <- 1800
f_end <- 8500

Okay <- 48
c(grid, freqs) %<-% wavelet_grid(x, Okay, f_start, f_end, fs)
plot_wavelet_diagram(
  torch_tensor(1:dim(grid)[2]),
  freqs, grid, Okay, fs, f_end,
  sort = "magnitude_sqrt"
)

The acquire in decision, on each the time and the frequency axis, is totally spectacular.

Thanks for studying!

Photograph by Vlad Panov on Unsplash