Kazane: simple sinc interpolation for 1D signal in PyTorch
Kazane utilize FFT based convolution to provide fast sinc interpolation for 1D signal when your sample rate only needs to change by an integer amounts; If you need to change by a fraction amounts, checkout julius.
Installation
pip install kazane
or
pip install git+https://github.com/yoyololicon/kazane
for latest version.
Usage
import kazane
import torch
signal = torch.randn(8, 2, 44100)
# downsample by an amount of 3
decimater = kazane.Decimate(3)
resampled_signal = decimater(signal)
# upsample by an amount of 2
upsampler = kazane.Upsample(2)
resampled_signal = upsampler(signal)
# you can also control number of zeros, roll-off frequency of the sinc interpolation kernel
decimater = kazane.Decimate(3, num_zeros=24, roll_off=0.9)
# use other types of window function for the sinc kernel
upsampler = kazane.Upsample(2, window_func=torch.blackman_window)
Benchmarks on CUDA
Using the benchmark scripts at bench, you can see that FFT can gives some speed improvements when the sample rate changes with some common integer numbers.
[---------- Down sample ----------]
| julius | kazane
2 threads: ------------------------
rate: 2 | 52.2 | 52.4
rate: 3 | 66.5 | 36.1
rate: 5 | 94.8 | 30.0
rate: 7 | 121.7 | 42.3
Times are in milliseconds (ms).
[----------- Up sample -----------]
| julius | kazane
2 threads: ------------------------
rate: 2 | 48.8 | 39.0
rate: 3 | 68.1 | 51.6
rate: 5 | 112.5 | 78.9
rate: 7 | 159.4 | 108.0
Times are in milliseconds (ms).
86 Dec 28, 2022
3k Jan 4, 2023
1 Jan 4, 2022
1 Jan 10, 2022
28 Dec 9, 2022
63 Dec 16, 2022
8 Oct 29, 2022
51 Nov 18, 2022
57 Nov 21, 2022
59 Dec 22, 2022
280 Dec 23, 2022
49 Nov 22, 2022
95 Dec 4, 2022
66 Dec 9, 2022
195 Dec 29, 2022
544 Dec 19, 2022
100 Nov 30, 2022
49 Jan 7, 2023