GitHub

Generating (time correlated) matrices

When generating testdata currently several options are available. Firstly the easiest solution is the function:

using AJD
data = AJD.get_test_data(:exact_diag, n_dims = 10, n_matrices = 10)
filter_ = diagonalize(data, algorithm = FFDiag())

which will create a set of 10 matrices, which are normal and commuting with dimension $10 \times 10$ and calculate the filter containing $V$. Instead of FFDiag() algorithm JDiagEdourdPineau() or all other algorithms can be used (see getting started).

If you want to generate time correlation matrices from a testset the extension AJDExtTimedTestdata is currently available using:

using AJD
using WAV
data = AJD.get_test_data(:approx_diag,"")
filter_ = diagonalize(data, algorithm = FFDiag())

which will create a time delayed correlation matrix with 3 measurement signals from the file channels3_room69_mix.wav, which can be found in the test_data folder and diagonalizes the set of matrices from the correlation matrix. The matrix in data has dimension $5 \\times 10$.

For larger testdata the :approx_diag_large symbol can be used.

Implementation of the time delayed correlation matrices used are explained here (p.16).

Note

The testdata functionality as well is implemented as an extension to AJD.jl. This may require you to add the WAV package to your Julia parent environment. This is done to limit the dependency footprint of AJD.jl.

Another way to generate a testset of correlation matrices from random vectors is provided with:

using AJD
using WAV
#you can adjust this!
no_of_signals,no_of_samples = (10,1000)
#get extension to generate signal
ext = Base.get_extension(AJD,:AJDExtTimedTestdata)
meas = ext.generate_random_measurements(no_of_signals,no_of_samples)
testset_data = ext.generate_testdata(meas, delay = 10,no_of_segments=11,show_warning = false)
filter_ = diagonalize(testset_data)

which generates random time correlated matrices. See documentation of generate_testdata for further information about the keywords. Currently the delay has to be smaller or equal to the elements in a segment e.g. if no_of_segments is 10 and data has thousand elements maximum delay is 100.

For more advanced testdata generation the following might be interesting.

Further information on generate_testdata

For realising discrete test data the files channels2_room69_mix.wav, channels3_room69_mix.wav and channels3_room69_mix_shortened.wav are included in the repository. You can load them into the get_test_data function by specifying the absolute path to the file and :exact_diag_large as the symbol. Might however not work when working from the temporary env.

Generating testset from https://github.com/fakufaku/bssspeechdataset

The data was generated using the github code found here.

If you want to generate your own test data or additional datasets refer to the github repository linked above and clone it. You'll need python!

For the code of the repo to work the following adaptations have to be made:

  • the generate_samples.py has to be changed on line 120. Change the deprecated np.floatto float

The generation will take a while and (probably) uses around 10GB of storage space. You can always cancel the generation of testdata during the process and get some testsets generated until then. After generating your testset you might want to edit the soundfiles since some of the testsets have

Known Issues

The calculation of time correlation matrices won't work if one of the segments is a vector of only zeros, since the correlation of this vector will become NaN due to the variance being zero! The function will throw an error in that case!

Make sure to decrease number of segments to counteract.