setting the values of sliding windows of an array in numpy - python

Suppose I have a 2D array with shape (3, 3), call it a, and an array of zeros with shape (7, 7, 5, 5), call it b. I want to modify b in the following way:
for p in range(5):
for q in range(5):
b[p:p + 3, q:q + 3, p, q] = a
Given:
a = np.array([[4, 2, 2],
[9, 0, 5],
[9, 9, 4]])
b = np.zeros((7, 7, 5, 5), dtype=int)
b would end up something like:
>>> b[:, :, 0, 0]
array([[4, 2, 2, 0, 0, 0, 0],
[9, 0, 5, 0, 0, 0, 0],
[9, 9, 4, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> b[:, :, 0, 1]
array([[0, 4, 2, 2, 0, 0, 0],
[0, 9, 0, 5, 0, 0, 0],
[0, 9, 9, 4, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])

One way to think about this to make a sliding window view of b (6D), slice out the parts you want (3D or 4D), and assign a to them.
However, there is a simpler way to do this altogether. The way a sliding window view works is by creating a dimension that steps along less than the full size of the dimension you are viewing. For example:
>>> x = np.array([1, 2, 3, 4])
array([1, 2, 3, 4])
>>> window = np.lib.stride_tricks.as_strided(
x, shape=(x.shape[0] - 2, 3),
strides=x.strides * 2)
[[1 2 3]
[2 3 4]]
I'm deliberately using np.lib.stride_tricks.as_strided rather than np.lib.stride_tricks.sliding_window_view here because it has a certain flexibility that you need.
You can have a stride that is larger than the axis you are viewing, as long as you are careful. Contiguous arrays are more forgiving in this case, but by no means a requirement. An example of this is np.diag. You can implement it something like this:
>>> x = np.arange(12).reshape(3, 4)
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> diag = np.lib.stride_tricks.as_strided(
x, shape=(min(x.shape),),
strides=(sum(x.strides),))
array([ 0, 5, 10])
The trick is to make a view of only the parts of b you care about in a way that makes the assignment easy. Because of broadcasting rules, you will want the last two dimensions of the view to be a.shape, and the strides to be b.strides[:2], since that's where you want to place a.
The first two dimensions of the view will be responsible for making the copies of a. You want 25 copies, so the shape will be (5, 5). The strides are the trickier part. Let's take a look at a 2D case, just because that's easier to visualize, and then attempt to generalize:
>>> a0 = np.array([1, 2])
>>> b0 = np.zeros((4, 3), dtype=int)
>>> b0[0:2, 0] = b0[1:3, 1] = b0[2:4, 2] = a0
The goal is to make a view that strides along the diagonal of b0 in the first axis. So:
>>> np.lib.stride_tricks.as_strided(
b0, shape=(b0.shape[0] - a0.shape[0] + 1, a0.shape[0]),
strides=(sum(b0.strides), b0.strides[0]))[:] = a0
>>> b0
array([[1, 0, 0],
[2, 1, 0],
[0, 2, 1],
[0, 0, 2]])
So that's what you do for b, but adding up every second dimension:
a = np.array([[4, 2, 2],
[9, 0, 5],
[9, 9, 4]])
b = np.zeros((7, 7, 5, 5), dtype=int)
vshape = (*np.subtract(b.shape[:a.ndim], a.shape) + 1,
*a.shape)
vstrides = (*np.add(b.strides[:a.ndim], b.strides[a.ndim:]),
*b.strides[:a.ndim])
np.lib.stride_tricks.as_strided(b, shape=vshape, strides=vstrides)[:] = a
TL;DR
def emplace_window(a, b):
vshape = (*np.subtract(b.shape[:a.ndim], a.shape) + 1, *a.shape)
vstrides = (*np.add(b.strides[:a.ndim], b.strides[a.ndim:]), *b.strides[:a.ndim])
np.lib.stride_tricks.as_strided(b, shape=vshape, strides=vstrides)[:] = a
I've phrased it this way, because now you can apply it to any number of dimensions. The only expectations is that 2 * a.ndim == b.ndim and that b.shape[a.ndim:] == b.shape[:a.ndim] - a.shape + 1.

Related

Zero-Padding an image without numpy

I am a beginner with Python and I am learning how to treat images.
Given a square image (NxN), I would like to make it into a (N+2)x(N+2) image with a new layer of zeros around it. I would prefer not to use numpy and only stick with the basic python programming. Any idea on how to do so ?
Right now, I used .extend to add zeros on the right side and on the bottom but can't do it up and left.
Thank you for your help!
We can create a padding function that adds layers of zeros around an image (padding it).
def pad(img,layers):
#img should be rectangular
return [[0]*(len(img[0])+2*layers)]*layers + \
[[0]*layers+r+[0]*layers for r in img] + \
[[0]*(len(img[0])+2*layers)]*layers
We can test with a sample image, such as:
i = [[1, 2, 3],
[4, 5, 6],
[7, 8, 9]]
So,
pad(i,2)
gives:
[[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 2, 3, 0, 0],
[0, 0, 4, 5, 6, 0, 0],
[0, 0, 7, 8, 9, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]]
Im assuming that by image we're talking about a matrix, in that case you could do this:
img = [[5, 5, 5], [5, 5, 5], [5, 5, 5]]
row_len = len(img)
col_len = len(img[0])
new_image = list()
for n in range(col_len+2): # Adding two more rows
if n == 0 or n == col_len + 1:
new_image.append([0] * (row_len + 2)) # First and last row is just zeroes
else:
new_image.append([0] + img[n - 1] + [0]) # Append a zero to the front and back of each row
print(new_image) # [[0, 0, 0, 0, 0], [0, 5, 5, 5, 0], [0, 5, 5, 5, 0], [0, 5, 5, 5, 0], [0, 0, 0, 0, 0]]

How to add an array twice by two TRUE conditions?

this is my code:
import numpy as np
from scipy.ndimage.interpolation import shift
B = np.array([[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]])
F = np.array([[0, 0, 0, 0, 0],
[0, 0, 0, 0, 1],
[0, 0, 0, 0, 1],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]])
M = np.array([[1, 2, 1, 2, 1],
[1, 2, 1, 2, 1],
[1, 2, 1, 2, 0],
[1, 2, 1, 2, 1],
[1, 2, 1, 2, 1]])
if F[2, 4] == 1:
B = np.add(M, B)
if F[1, 4] == 1:
M_shift = shift(M, (-1, 0), cval=0)
B = np.add(M_shift, B)
print(B)
I want to add M to B if the condition for F is true. In this example both if-conditions are true and i thought that my code will add two times M to B. But apparently it's not working? What is wrong?
Thanks in advance.
EDIT for arra:
My desired output is in this case, my code snippet should add the matrix with B and also the shifted version of M. Because both conditions are True. But as u said the first assignment is overwritten. How can I prevent the overwriting for this case?
The output, after the assignment should look like this:
B = np.array([[2, 4, 2, 4, 2],
[2, 4, 2, 4, 1],
[2, 4, 2, 4, 1],
[2, 4, 2, 4, 2],
[1, 2, 1, 2, 1]])
I guess a good way is to use a for-loop because i want to check all array entries in Matrix F, if there's a 1. But i reckon that if i use the np.add function I'll overwrite it again
You overwrite B in the second if, so any previous value will be erased. Not sure what you are trying to achieve in your code, but don't you want something like:
B = np.add(M_shift, B)
or
intermediateVal = np.add(M_shift, F)
B = np.add(B, intermediateVal)

Count how often integer y occurs right after integer x in a numpy array

I have a very large numpy.array of integers, where each integer is in the range [0, 31].
I would like to count, for every pair of integers (a, b) in the range [0, 31] (e.g. [0, 1], [7, 9], [18, 0]) how often b occurs right after a.
This would give me a (32, 32) matrix of counts.
I'm looking for an efficient way to do this with numpy. Raw python loops would be too slow.
Here's one way...
To make the example easier to read, I'll use a maximum value of 9 instead of 31:
In [178]: maxval = 9
Make a random input for the example:
In [179]: np.random.seed(123)
In [180]: x = np.random.randint(0, maxval+1, size=100)
Create the result, initially all 0:
In [181]: counts = np.zeros((maxval+1, maxval+1), dtype=int)
Now add 1 to each coordinate pair, using numpy.add.at to ensure that duplicates are counted properly:
In [182]: np.add.at(counts, (x[:-1], x[1:]), 1)
In [183]: counts
Out[183]:
array([[2, 1, 1, 0, 1, 0, 1, 1, 1, 1],
[2, 1, 1, 3, 0, 2, 1, 1, 1, 1],
[0, 2, 1, 1, 4, 0, 2, 0, 0, 0],
[1, 1, 1, 3, 3, 3, 0, 0, 1, 2],
[1, 1, 0, 1, 1, 0, 2, 2, 2, 0],
[1, 0, 0, 0, 0, 0, 1, 1, 0, 2],
[0, 4, 2, 3, 1, 0, 2, 1, 0, 1],
[0, 1, 1, 1, 0, 0, 2, 0, 0, 3],
[1, 2, 0, 1, 0, 0, 1, 0, 0, 0],
[2, 0, 2, 2, 0, 0, 2, 2, 0, 0]])
For example, the number of times 6 is followed by 1 is
In [188]: counts[6, 1]
Out[188]: 4
We can verify that with the following expression:
In [189]: ((x[:-1] == 6) & (x[1:] == 1)).sum()
Out[189]: 4
You can use numpy's built-in diff routine together with boolean arrays.
import numpy as np
test_array = np.array([1, 2, 3, 1, 2, 4, 5, 1, 2, 6, 7])
a, b = (1, 2)
sum(np.bitwise_and(test_array[:-1] == a, np.diff(test_array) == b - a))
# 3
If your array is multi-dimensional, you will need to flatten it first or make some small modifications to the code above.

How to speed up the creation of a label tensor from label map in Numpy?

Given a label map of dimensions W X H where each element can take values from {0,..,K-1} I want to output a label tensor of dimensions K X W x H where each element in the K'th map is 1 only if the corresponding value in the labelmap was K. Currently my implementation uses two for loops and is very slow.
p_label = Labelmap with one channel
label = np.zeros((K,p_label.shape[0], p_label.shape[1]))
for i in xrange(p_label.shape[0]):
for j in xrange(p_label.shape[1]):
label[p_label[i,j],i,j] = 1
Is there a better way to do this operation in Numpy using broadcasting?
You can use the == operator with broadcasting.
For example,
In [19]: W = 5
In [20]: H = 8
In [21]: K = 10
Create a p_label for the example:
In [22]: p_label = np.random.randint(0, K, size=(W, H))
kvals is simply an array containing [0, 1, ..., K-1]:
In [23]: kvals = np.arange(K)
kvals.reshape(-1, 1, 1) converts kvals to an array with shape (K, 1, 1). This is compared using == to p_label. Broadcasting applies, so the result of the comparison has shape (K, W, H). It is a boolean array of the values that you want. .astype(int) converts the result to an integer array. (You can remove that if a boolean array would work for you.)
In [24]: label = (p_label == kvals.reshape(-1, 1, 1)).astype(int)
Here's the original p_label. Note, for example, the locations of the value 0:
In [25]: p_label
Out[25]:
array([[3, 3, 2, 6, 2, 2, 9, 3],
[1, 8, 1, 1, 4, 3, 7, 8],
[5, 9, 1, 0, 7, 2, 8, 0],
[1, 3, 5, 4, 6, 0, 9, 5],
[5, 7, 2, 0, 6, 4, 5, 3]])
label[0] is 1 in the positions where p_label is 0.
In [26]: label[0]
Out[26]:
array([[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0, 1],
[0, 0, 0, 0, 0, 1, 0, 0],
[0, 0, 0, 1, 0, 0, 0, 0]])
Label[p_label, np.arange(p_label.shape[0])[:,None], np.arange(p_label.shape[1])] = 1
The 3 index arrays broadcast against each other.
==============================
lmap = np.arange(12).reshape(3,4)
lbl = np.zeros((12,3,4),int)
lbl[lmap,np.arange(3)[:,None],np.arange(4)] = 1
In [5]: lbl
Out[5]:
array([[[1, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]],
[[0, 1, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]],
...
[[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 1]]])

Numpy resize and fill with specific value

How can i resize a numpy array and fill it with a specific value (if some dimension is extended) ?
I find a way to extend my array with np.pad but I can't shorten it:
>>> import numpy as np
>>> a = np.ndarray((5, 5), dtype=np.uint16)
>>> a
array([[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint16)
>>> np.pad(a, ((0, 1), (0,3)), mode='constant', constant_values=9)
array([[0, 0, 0, 0, 0, 9, 9, 9],
[0, 0, 0, 0, 0, 9, 9, 9],
[0, 0, 0, 0, 0, 9, 9, 9],
[0, 0, 0, 0, 0, 9, 9, 9],
[0, 0, 0, 0, 0, 9, 9, 9],
[9, 9, 9, 9, 9, 9, 9, 9]], dtype=uint16)
And if i use resize i can't specify the value that I want to use.
>>> a.fill(5)
>>> a.resize((2, 7))
>>> a
array([[5, 5, 5, 5, 5, 5, 5],
[5, 5, 5, 5, 5, 5, 5]], dtype=uint16)
But i would like
>>> a
array([[5, 5, 5, 5, 5, 9, 9],
[5, 5, 5, 5, 5, 9, 9]], dtype=uint16)
After some test I create this function but it's only work when you change x_value or with a lower y_value, if you need to increase y dimension it doesn't work, why ?
VALUE_TO_FILL = 9
def resize(self, x_value, y_value):
x_diff = self.np_array.shape[0] - x_value
y_diff = self.np_array.shape[1] - y_value
self.np_array.resize((x_value, y_value), refcheck=False)
if x_diff < 0:
self.np_array[x_diff:, :] = VALUE_TO_FILL
if y_diff < 0:
self.np_array[:, y_diff:] = VALUE_TO_FILL
Your array has a fixed size data buffer. You can reshape the array without changing that buffer. You can take a slice (view) without changing the buffer. But you can't add values to the array without changing the buffer.
In general resize returns an new array with a new data buffer.
pad is a complex function to handle general cases. But the simplest approach is to create the empty target array, fill it, and then copy the input into the right place.
Alternatively pad could create the fill arrays and concatenate them with the original. But concatenate also makes the empty return and copies.
A do it yourself pad with clipping could be structured as:
n,m = X.shape
R = np.empty((k,l))
R.fill(value)
<calc slices from n,m,k,l>
R[slice1] = X[slice2]
Calculating the slices may require if-else tests or equivalent min/max. You can probably work out those details.
This may be all that is needed
R[:X.shape[0],:X.shape[1]]=X[:R.shape[0],:R.shape[1]]
That's because there's no problem if a slice is larger than the dimension.
In [37]: np.arange(5)[:10]
Out[37]: array([0, 1, 2, 3, 4])
Thus, for example:
In [38]: X=np.ones((3,4),int)
In [39]: R=np.empty((2,5),int)
In [40]: R.fill(9)
In [41]: R[:X.shape[0],:X.shape[1]]=X[:R.shape[0],:R.shape[1]]
In [42]: R
Out[42]:
array([[1, 1, 1, 1, 9],
[1, 1, 1, 1, 9]])
To shorten it, you can use negative values in slice :
>>> import numpy as np
>>> a = np.ndarray((5, 5), dtype=np.uint16)
>>> a
array([[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint16)
>>> b = a[0:-1,0:-3]
>>> b
array([[0, 0],
[0, 0],
[0, 0],
[0, 0]], dtype=uint16)

Categories

Resources