1
Performance Measurement and Improvement Techniques {#tutorial_py_optimization}
2
==================================================
7
In image processing, since you are dealing with large number of operations per second, it is
8
mandatory that your code is not only providing the correct solution, but also in the fastest manner.
9
So in this chapter, you will learn
11
- To measure the performance of your code.
12
- Some tips to improve the performance of your code.
13
- You will see these functions : **cv2.getTickCount**, **cv2.getTickFrequency** etc.
15
Apart from OpenCV, Python also provides a module **time** which is helpful in measuring the time of
16
execution. Another module **profile** helps to get detailed report on the code, like how much time
17
each function in the code took, how many times the function was called etc. But, if you are using
18
IPython, all these features are integrated in an user-friendly manner. We will see some important
19
ones, and for more details, check links in **Additional Resouces** section.
21
Measuring Performance with OpenCV
22
---------------------------------
24
**cv2.getTickCount** function returns the number of clock-cycles after a reference event (like the
25
moment machine was switched ON) to the moment this function is called. So if you call it before and
26
after the function execution, you get number of clock-cycles used to execute a function.
28
**cv2.getTickFrequency** function returns the frequency of clock-cycles, or the number of
29
clock-cycles per second. So to find the time of execution in seconds, you can do following:
31
e1 = cv2.getTickCount()
33
e2 = cv2.getTickCount()
34
time = (e2 - e1)/ cv2.getTickFrequency()
36
We will demonstrate with following example. Following example apply median filtering with a kernel
37
of odd size ranging from 5 to 49. (Don't worry about what will the result look like, that is not our
40
img1 = cv2.imread('messi5.jpg')
42
e1 = cv2.getTickCount()
43
for i in xrange(5,49,2):
44
img1 = cv2.medianBlur(img1,i)
45
e2 = cv2.getTickCount()
46
t = (e2 - e1)/cv2.getTickFrequency()
49
# Result I got is 0.521107655 seconds
51
@note You can do the same with time module. Instead of cv2.getTickCount, use time.time() function.
52
Then take the difference of two times.
54
Default Optimization in OpenCV
55
------------------------------
57
Many of the OpenCV functions are optimized using SSE2, AVX etc. It contains unoptimized code also.
58
So if our system support these features, we should exploit them (almost all modern day processors
59
support them). It is enabled by default while compiling. So OpenCV runs the optimized code if it is
60
enabled, else it runs the unoptimized code. You can use **cv2.useOptimized()** to check if it is
61
enabled/disabled and **cv2.setUseOptimized()** to enable/disable it. Let's see a simple example.
63
# check if optimization is enabled
64
In [5]: cv2.useOptimized()
67
In [6]: %timeit res = cv2.medianBlur(img,49)
68
10 loops, best of 3: 34.9 ms per loop
71
In [7]: cv2.setUseOptimized(False)
73
In [8]: cv2.useOptimized()
76
In [9]: %timeit res = cv2.medianBlur(img,49)
77
10 loops, best of 3: 64.1 ms per loop
79
See, optimized median filtering is \~2x faster than unoptimized version. If you check its source,
80
you can see median filtering is SIMD optimized. So you can use this to enable optimization at the
81
top of your code (remember it is enabled by default).
83
Measuring Performance in IPython
84
--------------------------------
86
Sometimes you may need to compare the performance of two similar operations. IPython gives you a
87
magic command %timeit to perform this. It runs the code several times to get more accurate results.
88
Once again, they are suitable to measure single line codes.
90
For example, do you know which of the following addition operation is better, x = 5; y = x\*\*2,
91
x = 5; y = x\*x, x = np.uint8([5]); y = x\*x or y = np.square(x) ? We will find it with %timeit in
96
In [11]: %timeit y=x**2
97
10000000 loops, best of 3: 73 ns per loop
99
In [12]: %timeit y=x*x
100
10000000 loops, best of 3: 58.3 ns per loop
102
In [15]: z = np.uint8([5])
104
In [17]: %timeit y=z*z
105
1000000 loops, best of 3: 1.25 us per loop
107
In [19]: %timeit y=np.square(z)
108
1000000 loops, best of 3: 1.16 us per loop
110
You can see that, x = 5 ; y = x\*x is fastest and it is around 20x faster compared to Numpy. If you
111
consider the array creation also, it may reach upto 100x faster. Cool, right? *(Numpy devs are
112
working on this issue)*
114
@note Python scalar operations are faster than Numpy scalar operations. So for operations including
115
one or two elements, Python scalar is better than Numpy arrays. Numpy takes advantage when size of
116
array is a little bit bigger.
118
We will try one more example. This time, we will compare the performance of **cv2.countNonZero()**
119
and **np.count_nonzero()** for same image.
122
In [35]: %timeit z = cv2.countNonZero(img)
123
100000 loops, best of 3: 15.8 us per loop
125
In [36]: %timeit z = np.count_nonzero(img)
126
1000 loops, best of 3: 370 us per loop
128
See, OpenCV function is nearly 25x faster than Numpy function.
130
@note Normally, OpenCV functions are faster than Numpy functions. So for same operation, OpenCV
131
functions are preferred. But, there can be exceptions, especially when Numpy works with views
134
More IPython magic commands
135
---------------------------
137
There are several other magic commands to measure the performance, profiling, line profiling, memory
138
measurement etc. They all are well documented. So only links to those docs are provided here.
139
Interested readers are recommended to try them out.
141
Performance Optimization Techniques
142
-----------------------------------
144
There are several techniques and coding methods to exploit maximum performance of Python and Numpy.
145
Only relevant ones are noted here and links are given to important sources. The main thing to be
146
noted here is that, first try to implement the algorithm in a simple manner. Once it is working,
147
profile it, find the bottlenecks and optimize them.
149
-# Avoid using loops in Python as far as possible, especially double/triple loops etc. They are
151
2. Vectorize the algorithm/code to the maximum possible extent because Numpy and OpenCV are
152
optimized for vector operations.
153
3. Exploit the cache coherence.
154
4. Never make copies of array unless it is needed. Try to use views instead. Array copying is a
157
Even after doing all these operations, if your code is still slow, or use of large loops are
158
inevitable, use additional libraries like Cython to make it faster.
163
-# [Python Optimization Techniques](http://wiki.python.org/moin/PythonSpeed/PerformanceTips)
164
2. Scipy Lecture Notes - [Advanced
165
Numpy](http://scipy-lectures.github.io/advanced/advanced_numpy/index.html#advanced-numpy)
166
3. [Timing and Profiling in IPython](http://pynash.org/2013/03/06/timing-and-profiling.html)