This package implements a staged-programming approach for a highly specialized version of a Savitzky-Golay filter. It is a generalized, performant algorithm of rolling/moving functions, such as a rolling/moving average.
If you would like to read a full blog post on the implementation on it, you can find some here. For a video explanation by Stefan Karpinski, check out the following video
Many people contributed to the eventual development of this code:
- Jiahao Chen, Matt Bauman and others for the original implementation
- Stefan Karpinski for some branch elisions
- Steven G Johnson, for help when benchmarking a PyCall code and being careful not to measure the conversion back and forth.
If you know the measurement to be missing / not be as fair as it could be, please reach out and let me know - we're trying to compare best in class implementations side by side, not cheat out. In particular, I don't know if the savgol method I'm calling is in place or not, and would appreciate if somone could point out how to opt into that.
Run the following script:
python -m timeit -s "import scipy" -s "import numpy as np" -s "data = np.array([i for i in range(10_000_000)], dtype = np.float32)" -s "f = scipy.signal.savgol_filter" "f(data, 5, 2, mode = 'wrap')"
python -m timeit -s "import scipy" -s "import numpy as np" -s "data = np.array([i for i in range(10_000_000)], dtype = np.float64)" -s "f = scipy.signal.savgol_filter" "f(data, 5, 2, mode = 'wrap')"or by running python -m IPython and copy pasting the test/bench.py script in.