-
Notifications
You must be signed in to change notification settings - Fork 53
/
benchmarks_zoom.py
141 lines (118 loc) · 5.61 KB
/
benchmarks_zoom.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
"""
Benchmark tests comparing fourier zoom to other techniques
imsize map_coordinates fourier_zoom
50 0.0523551 0.0214219
101 0.214316 0.092334
153 0.499386 0.245633
204 1.13154 0.506588
255 2.03562 1.33743
306 4.2135 3.11828
358 6.59853 5.83727
409 8.52711 9.29976
460 11.1365 12.3836
511 17.8496 17.9716
563 19.4098 24.3088
614 10.559 29.7606
665 11.6219 15.7843
716 15.1977 20.9938
768 16.6483 22.0162
819 19.7834 28.2969
870 21.6399 34.0512
921 24.7352 41.2564
973 27.3512 48.8034
1024 32.1365 51.9153
Somewhat disappointing, but the fourier approach can be parallelized pretty
easily, I don't know about map_coordinates.
Also, I'm more than a little suspicious of these results; the execution times
took a precipitous drop between 615 and 665 seconds, which I think is an
indication that I was running too much other junk in parallel.
It's also pretty weird that the fourier approach seems to go up more steeply
than the map_coordinates approach; it could indicate memory limitations on my
machine.
Another test:
imsize map_coordinates fourier_zoom skimage_zoom
50 0.086158 0.0248492 0.0535769
100 0.308121 0.114798 0.199375
150 0.762103 0.300909 0.53565
200 1.32903 0.753605 0.900321
250 2.13581 1.33399 1.39549
300 2.98052 2.05711 2.18081
350 4.3403 3.49586 2.69526
400 5.45967 4.82973 3.68139
450 7.97855 6.9673 4.62726
500 8.367 8.79732 3.99151
skimage is the clear winner, except for small images. Hrmph.
"""
import itertools
import timeit
import time
import numpy as np
zoomtimings = {'map_coordinates':[],
'fourier_zoom':[],
'skimage_zoom':[],
#'griddata_nearest':[],
#'griddata_linear':[],
#'griddata_cubic':[],
}
imsizes = np.round(np.linspace(50,1024,20))
imsizes = np.round(np.linspace(50,500,10))
for imsize in imsizes:
t0 = time.time()
setup = """
import numpy as np
#im = np.random.randn({imsize},{imsize})
yy,xx = np.indices([{imsize},{imsize}])
im = np.exp(-((xx-{imsize}/2.)**2+(yy-{imsize}/2.)**2)/(2**2*2.))
# upsample by factor of 2
yr = np.linspace(0,(({imsize}*2)-1)/2.,{imsize}*2)-0.25 # middle conventions...
xr = np.linspace(0,(({imsize}*2)-1)/2.,{imsize}*2)-0.25 # middle conventions...
xxnew,yynew = np.meshgrid(xr,yr)
import image_registration.fft_tools.zoom as fzm
import scipy.interpolate as si
import scipy.ndimage as snd
points = zip(xx.flat,yy.flat)
imflat = im.ravel()
import skimage.transform as skit
""".replace(" ","").format(imsize=imsize)
fzoom_timer = timeit.Timer("ftest=fzm.zoomnd(im,usfac=2,outshape=xxnew.shape)",
setup=setup)
# too slow!
# interp2d_timer = timeit.Timer("itest=si.interp2d(xr,yr,im)(xr-0.5,yr-0.5)",
# setup=setup)
mapzoom_timer = timeit.Timer("mtest=snd.map_coordinates(im,[yynew,xxnew])",
setup=setup)
skimagezoom_timer = timeit.Timer("stest=skit.resize(im,xxnew.shape)",setup=setup)
# all slopw
#grid_timer_nearest = timeit.Timer("gtest=si.griddata(points,imflat,(xx-0.5,yy-0.5), method='nearest')",
# setup=setup)
#grid_timer_linear = timeit.Timer("gtest=si.griddata(points,imflat,(xx-0.5,yy-0.5), method='linear')",
# setup=setup)
#grid_timer_cubic = timeit.Timer("gtest=si.griddata(points,imflat,(xx-0.5,yy-0.5), method='cubic')",
# setup=setup)
print "imsize %i fourier zoom " % imsize,
zoomtimings['fourier_zoom'].append( np.min(fzoom_timer.repeat(3,10)) )
print "imsize %i map_coordinates zoom " % imsize,
zoomtimings['map_coordinates'].append( np.min(mapzoom_timer.repeat(3,10)) )
print "imsize %i skimage zoom " % imsize,
zoomtimings['skimage_zoom'].append( np.min(skimagezoom_timer.repeat(3,10)) )
#zoomtimings['griddata_nearest'].append( np.min(grid_timer_nearest.repeat(3,10)) )
#zoomtimings['griddata_linear'].append( np.min(grid_timer_linear.repeat(3,10)) )
#zoomtimings['griddata_cubic'].append( np.min(grid_timer_cubic.repeat(3,10)) )
print "imsize %i done, %f seconds" % (imsize,time.time()-t0)
print "%10s " % "imsize"," ".join(["%16s" % t for t in zoomtimings.keys()])
for ii,sz in enumerate(imsizes):
print "%10i " % sz," ".join(["%16.6g" % t[ii] for t in zoomtimings.values()])
import scipy.optimize as scopt
def f(x,a,b):
return a*x**b
pm,err = scopt.curve_fit(f,imsizes[1:],zoomtimings['map_coordinates'][1:])
pf,err = scopt.curve_fit(f,imsizes[1:],zoomtimings['fourier_zoom'][1:])
ps,err = scopt.curve_fit(f,imsizes[1:],zoomtimings['skimage_zoom'][1:])
import matplotlib.pyplot as pl
pl.clf()
pl.loglog(imsizes,zoomtimings['map_coordinates'],'+')
pl.loglog(imsizes,zoomtimings['fourier_zoom'],'x')
pl.loglog(imsizes,zoomtimings['skimage_zoom'],'x')
pl.loglog(imsizes,f(imsizes,*pm))
pl.loglog(imsizes,f(imsizes,*pf))
pl.loglog(imsizes,f(imsizes,*ps))