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.
| Col1 | Col2 | Result | Notes |
|---|---|---|---|
| Tensorflow CWT (GPU) | Tensorflow CWT (CPU) | GPU ~8x faster | old i5 vs GTX 750 TI ~1,400 GFLOPS |
| Tensorflow CWT (CPU) | Pywavelet CWT (CPU) | Pywavelet CWT ~13x faster | Tensorflow w/o AVX extensions, etc |
| Tensorflow CWT (CPU) | Pywavelet DWT (CPU) | Pywavelet DWT ~200,000x faster | Haar wavelet; Tensorflow w/o AVX extensions, etc |
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