aaren / wavelets Goto Github PK

@aaren hasn't done anything on Github since 2017?

There are a billion forks implementing the same factorial changes:

@joschaschmiedt has accepted a PR from another user. Are you the de facto canonical copy now?

Currently there is only Morlet and Ricker.

Extend Ricker to general Derivative of Gaussian.

Add Paul

Add Shannon.

Example code:

from wavelets import WaveletAnalysis
wa = WaveletAnalysis()

# hard coded Morlet value
print wa.C_d
print wa.wavelet.C_d

# override hard coding
wa.wavelet.C_d = None
# Now C_d is explicitly computed
print wa.C_d

The explicitly computed version is giving me a value of (0.13708 + 0j), rather than the 0.776 that we expect for a Morlet with w0 = 6 (which is the default).

Hi, Thank you so much for creating such a nice wavelet package! Did you happen to have a problem with wavelet power calculation for large time series data? it seems there is a memory leakage problem. Thanks!

Allow this behaviour:

wt.fourier_periods = array_of_periods

The scales of a transform can currently be set by

wt = WaveletTransform(data)
wt.scales = array_of_scales

The equivalent fourier period is available as a property wt.fourier_periods,
however we can't set the scales via the fourier_periods. This would be useful
as the equivalent fourier period allows us to compare wavelets.

The usage example given in README.md fails on the line

power = wa.wavelet_power

with the error

ValueError: could not broadcast input array from shape (1042) into shape (1000)

Python version is 2.7.3, numpy version is 1.7.1, scipy version is 0.10.1.

I found this project today. I tried to run test_wavelets.py at first. But it didn't work because I couldn't find function named WaveletAnalysis. If you know the solution, please teach me that. If I find the solution before you reply this comment, I'll delete this.

We can reconstruct the input data from the wavelet transform. By reconstructing using a non complete set of scales we can effectively filter the data to look at variation at the scales of interest.

Currently we get the reconstruction from the reconstruction() method of WaveletAnalysis. Add an argument to the method that specifies the scales to use.

Hi Aaren,

first of all thanks a lot for this awesome clean implementation, it's a pleasure to use!
I realized yesterday that somehow my power values where off quite a lot (using unbias=True), which I was able to fix by just scaling the powers by dt. I am not sure where one would need to fix this ideally within the library (maybe some other values are affected as well?) otherwise I would have raised a pr directly ;-).

Greetings

when i use global_wavelet_spectrum, how to plot the significance testing line?

The factorial function has been moved from scipy.misc to scipy.special in version 1.3.1

my program is as follows:

    x, _ = librosa.core.load(dataset_path + "test.wav", sr=16000, mono=True)
    x = x[26000:40000]
    wa = WaveletAnalysis(sig, dt=0.1)
    print(wa.reconstruction().shape)

An error occurred during runtime:

  File "\wavelets\transform.py", line 85, in cwt
    return cwt_time(data, wavelet, widths, dt, axis)
  File "\wavelets\transform.py", line 105, in cwt_time
    wavelet_data[slices],
IndexError: only integers, slices (`:`), ellipsis (`...`), numpy.newaxis (`None`) and integer or boolean arrays are valid indices

Because there is an error in your source code, as shown below,

def cwt_time(data, wavelet, widths, dt, axis):
    # wavelets can be complex so output is complex
    ......
        wavelet_data = norm * wavelet(t, width)
        output[ind, :] = scipy.signal.fftconvolve(data,
                                                  wavelet_data[slices],
                                                  mode='same')
    return output

it should be corrected in this way. Obviously, the sequence value operation is missing：

def cwt_time(data, wavelet, widths, dt, axis):
    # wavelets can be complex so output is complex
    ......
        wavelet_data = norm * wavelet(t, width)
        output[ind, :] = scipy.signal.fftconvolve(data,
                                                  wavelet_data[slices[axis]],
                                                  mode='same')
    return output

Hi Aaren,

I am trying to compare your code with PyWavelet. My main objective is adding cone of influence visualisation for CWT in PyWavelet. I am using your bias example with three frequencies. I am using expression for COI taken from Table 1 in Torrence and Combo and I am computing my frequencies using scale2frequencies from PyWavelet which uses scaled central frequency and I additionally scale it by sampling period.

The picture I am getting is close to what I expect but not quite. COI is over-predicted which I think is due to the Fourier wavelength calculation or frequency-scale relationship. Could you perhaps tell me how Fourier wavelength is derived?

Hi,
I am currently looking/ comparing different wavelet packages

have to say yours is a very nice package.

however there may be an issue in the correct calculation of the scaling.
In the standard Torrence/Compo wavelet package the default minimum scale (j0) is set to 2

Now according to your code (transform.py line 321) you should do the same (if I understood correctly)

However if I set dt =1 i get s0 = 1.93602661839, NOT 2.0!

Is that a mistake/ a rounding error coming from how you calculate s0 ?

I am using the python code hosted here for comparison (and the idl code also)
http://atoc.colorado.edu/research/wavelets/software.html

I mean, the result is overall the same, but the scales are slightly shifted, which is annoying.