if you are interested in becoming a collaborator let me know by creating an issue or gmail me nickgeoca

This implements a 1-D Continuous Wavelet Transform (CWT) in tensorflow. The benefit is that it runs parallel on GPUs.

The following wavelets are available:

• Ricker wavelet - cwtRicker
• Mortlet wavelet - cwtMortlet

Regarding CWT performance of Tensorflow vs Pywavelet, Pywavelet is about 13 times faster. However, this is a CPU only benchmark without using performance extensions, like AVX, on Tensorflow.

This can be aquired by running python benchmark.py

• DWT - sampleSize = 10000000
  • pywavelet dwt haar: 0.06824707984924316
  • pywavelet dwt db2: 0.08141493797302246
  • pywavelet dwt db8: 0.14669179916381836
• CWT - sampleSize = 10000; cwtWidth = 256
  • pywavelet cwt mortlet: 1.1284675598144531
  • tensorflow cwt mortlet: 14.783239364624023

• wavExample.py. The audio sample rate is scaled down to 8000 samples per second (instead of typical 44100).

• sinExample.py. It produces the plot below. The wavelet used is shown below (scale=32).

• The wavelet can be undersampled if the scale is too small. An example of this is seen below- the scale was set to 1.

• This cwt and scipy's cwt both limit the Ricker wavelet samples to 10x the scale size to improve accuracy.

• Add this line of code similar to scipy's cwt: samples = min(10*width, len(wav))
• consier scipy's ability to specify the wavelet scale

# Scipy's cwt can specify the wavelet scales in detail. This api can't do that.
cwt(wav, signal.ricker, [1,1.5,2,2.5,3])
# This api is equivilent to calling scipy's cwt as below.
cwt(wav, signal.ricker, range(1,n))

• Maybe add 2d verison