How to achieve matrix mapping without using for loop using numpy in python?

Question

I have three arrays, a src array of shape (width, height, 3) , a template array of shape (640, 480, 3) and an img_map array of shape (width, height, 2) .

img_map stores the x,y relationship between src and template, i.e. src[4, 5] = template[img_map[0,4,5], img_map[1,4,5]]. When index suggested by the mapping is out of bound, no value will be assigned.

How to achieve such value assignment without for loop using numpy in python?

Using for loop will be like this, but it is too slow.

for x in range(0, width): 
  for y in range(0, height):
    u = img_map[x,y,0]
    v = img_map[x,y,1]
    if (u >= 0) and (v >= 0) and (u < 640) and (v < 480):
      src[x,y,:] = template[u,v,:]

if img_map.shape is (width, height, 2) then img_map[1,4,5] will always throw an error. please make sure your code is runnable before asking. — Daniel F
– Daniel F, Commented Nov 9, 2020 at 11:09

bousof · Accepted Answer · 2020-11-09 11:50:55Z

Here is a one liner version that should do the job:

src = template[img_map[:,:,0].ravel(),img_map[:,:,1].ravel(),:].reshape((width,height,3))

Here is a small code that verify the accuracy of results and evaluate the performance improvement:

import numpy as np
import time

# Preparation of testdata
(width, height, depth) = (1000, 1000, 3)
template = np.random.randint(0, 256, (640,480,depth))
img_map = np.concatenate((
    np.random.randint(0, template.shape[0], (width*height,1)),
    np.random.randint(0, template.shape[1], (width*height,1))
  ), axis=1
).reshape(width,height,2)

# For loop verion
t_start = time.time()
src = np.zeros((img_map.shape[0],img_map.shape[1],template.shape[2]))
for x in range(0, img_map.shape[0]): 
  for y in range(0, img_map.shape[1]):
    u = img_map[x,y,0]
    v = img_map[x,y,1]
    if (u >= 0) and (v >= 0) and (u < template.shape[0]) and (v < template.shape[1]):
      src[x,y,:] = template[u,v,:]
print(f'Timer 1: {time.time()-t_start}s')

# One line version
t_start = time.time()
src2 = template[img_map[:,:,0].ravel(),img_map[:,:,1].ravel(),:].reshape((width,height,depth))
print(f'Timer 2: {time.time()-t_start}s')

# Verify that both gives the same result
print('Results are equal' if np.linalg.norm(src-src2)==0 else 'Results differ')

The results are equal and the execution time is significantly improved:

% python3 script.py
Timer 1: 2.7333829402923584s
Timer 2: 0.0499570369720459s
Results are equal

Don't hesitate if you have questions.

Collectives™ on Stack Overflow

How to achieve matrix mapping without using for loop using numpy in python?

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related