Look into contourpy and check whether it is faster than `find_contours` from scikit-image

[paulmueller]

contourpy is what matplotlib uses.

https://contourpy.readthedocs.io

[RaghavaAlajangi]

Test

import time

import contourpy as cp
import dclab
import matplotlib.pylab as plt
import numpy as np
from skimage.measure import find_contours

p = r"S:\Data\RTDC_analysis\reproducability_measurements\20240610_Felix_blood_reproducibility_refocus\sample_11\M001_data.rtdc"

ds = dclab.new_dataset(p)

image = ds["mask"]

n = 6000

t1 = time.time()
test = [
    find_contours(
        im, level=0.9999, positive_orientation="low", fully_connected="high"
    )
    for im in image[:n]
]
fc = time.time() - t1
print("skimage contours time: ", fc)

t2 = time.time()
test1 = [cp.contour_generator(z=im).lines(level=0.7) for im in image[:n]]
cg = time.time() - t2
print("Contourpy contours time: ", cg)

print("improvement:", (fc / cg))

Result:

skimage contours time:  2.5009407997131348
Contourpy contours time:  1.8994224071502686
improvement: 1.3166848986820858 times

[paulmueller]

Thanks for looking into this and posting the results here ❤

We might consider contourpy for the planned get_contourlines method. I am just a bit hesitant, because it means an additional dependencies. I have to think about it a bit. Let's keep talking.

[RaghavaAlajangi]

Yeah, but it is much smaller than skimage. We can eliminate the hardcoded code as well. I can see a clear advantage.

[paulmueller]

find_contours from skimage is a part of dclab: https://github.com/DC-analysis/dclab/tree/master/dclab/external/skimage (skimage is not a dependency of dclab).

It would be grand, if we could do the same thing with contourpy.

[RaghavaAlajangi]

Since it is written in C++, I doubt whether it supports the same thing as skimage.