What is np.view
When slicing or transpose takes place on an array, we actually get a np.view object that shares the underlying memory with the array.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
import numpy as np
input_image = np.array([[1, 2, 3, 4],
[5, 6, 7, 8],
[9, 10, 11, 12]])
# This is slicing
reverse_arr = input_image[:,::-1]
# True
print(f'Are reverse_arr and input_image sharing memory? {np.may_share_memory(input_image, reverse_arr)}')
# False
con_reverse_arr = np.ascontiguousarray(reverse_arr)
print(f'Are con_reverse_arr and input_image sharing memory? {np.may_share_memory(input_image, con_reverse_arr)}')
# This is transpose
transposed_arr = input_image.T
# True
print(f'Are con_reverse_arr and input_image sharing memory? {np.may_share_memory(input_image, transposed_arr)}')
np.may_share_memory()checks if two arrays share the same underlying memory
How does it work?
An np.array has 4 important compoenents:
databuffershapestridesdtype
In the above array, we have
1
2
3
4
5
6
[[ 1 2 3 4]
[ 5 6 7 8]
[ 9 10 11 12]]
Strides: (32, 8)
Shape: (3, 4)
Dtype: int64
The databuffer is 1D array.
stridesmeans that to go to the next row, on the 1D array, one needs to jump32 bytes (32/8=4 int).- To go to the next column, one needs to jump
8/8 = 1 int.
On the transpose array:
1
2
3
4
5
6
7
[[ 1 5 9]
[ 2 6 10]
[ 3 7 11]
[ 4 8 12]]
Strides: (8, 32)
Shape: (4, 3)
Dtype: int64
- To go to the next row: one needs to jump 1 int. To go to the next column, one needs to jump 4 int.
When Is np.ascontiguous Necessary
When interfacing with other libraries, like torch, OpenCV, or a low level C library, arrays are often expected to be contiguous. When we have an array that’s actually a view object, we can use np.ascontiguous() to copy the array into a new array.
1
2
3
4
5
6
7
8
9
10
11
12
13
import numpy as np
input_image = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
# This is slicing
reverse_arr = input_image[:,::-1]
print(reverse_arr.flags['C_CONTIGUOUS']) # Output: False
# True
print(f'Are reverse_arr and input_image sharing memory? {np.may_share_memory(input_image, reverse_arr)}')
con_reverse_arr = np.ascontiguousarray(reverse_arr)
print(con_reverse_arr.flags['C_CONTIGUOUS']) # Output: True
# False
print(f'Are con_reverse_arr and input_image sharing memory? {np.may_share_memory(input_image, con_reverse_arr)}')
