cv2.imwrite/cv2.imread are faster than np.save/np.load on numpy matrices with integers

In one of the ongoing projects, I generate a lot of small matrices (~1000 entries). These are often written to the disk and read from it. I was trying to optimize this step since this step run in a REST api.

The standard way is numpy.save and numpy.load which works fine. My experience with opencv made me wonder if cv2.imwrite and cv2.imread can be used. The later saves the matrices to png file which works without any loss of data if you have 8 or 16 bit integer values in your matrices. I do.

Here are some benchmarks using %timeit.

Only saving matrices

I generate a random matrix of 400 entries and save it to a png file using OpenCV and also using numpy. OpenCV is roughly 3.6 times faster.

In [8]: %timeit cv2.imwrite('/tmp/a.png', np.random.randint(0, 255, (20,20)))                                                                                     
142 µs ± 1.05 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

In [10]: %timeit np.save('/tmp/a.npy', np.random.randint(0, 255, (20,20)))                                                                                        
517 µs ± 96.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

The trend continues (OpenCV is ever so slightly faster here) for matrices with 900 entries. I did not test with larger matrices.

In [14]: %timeit cv2.imwrite('/tmp/a.png', np.random.randint(0, 255, (30,30))); a = cv2.imread('/tmp/a.png')                                                      
215 µs ± 2.93 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [15]: %timeit np.save('/tmp/a.npy', np.random.randint(0, 255, (30,30))); np.load('/tmp/a.npy')                                                                 
791 µs ± 29.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Saving and reading matrices

Reading back into matrices is also faster with OpenCV. Following micro-benchmarks write and read a randomly generated matrix using OpenCV and numpy.

In [12]: %timeit cv2.imwrite('/tmp/a.png', np.random.randint(0, 255, (20,20))); a = cv2.imread('/tmp/a.png')                                                      
192 µs ± 2.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [13]: %timeit np.save('/tmp/a.npy', np.random.randint(0, 255, (20,20))); np.load('/tmp/a.npy')                                                                 
770 µs ± 23.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

I guess I’ll be using OpenCV to save data whenever I have integer matrices.

I used Python3.6 on OpenSUSE-15.1 server.