Related
I have an array for an example:
import numpy as np
data=np.array([[4,4,4,0,1,1,1,0,0,0,0,1,0,0,1],
[3,0,0,1,1,1,1,1,1,1,1,0,0,1,0],
[6,0,0,1,1,1,1,1,0,0,0,0,1,0,0]])
Requirement :
In the data array, if element 1's are consecutive as the square size
of ((3,3)) and more than square size no changes. Otherwise, replace
element value 1 with zero except the square size.
Expected output :
[[4 4 4 0 1 1 1 0 0 0 0 0 0 0 0]
[3 0 0 0 1 1 1 0 0 0 0 0 0 0 0]
[6 0 0 0 1 1 1 0 0 0 0 0 0 0 0]]
I will provide here as solutions two different approaches. One which doesn't and one which is using Python loops. Let's start with the common header:
import numpy as np
from skimage.util import view_as_windows as winview
data=np.array([[4,4,4,0,1,1,1,0,0,0,0,1,0,0,1],
[3,0,0,1,1,1,1,1,1,1,1,0,0,1,0],
[6,0,0,1,1,1,1,1,0,0,0,0,1,0,0]])
Below an approach without using Python loops resulting in shortest code, but requiring import of an additional module skimage:
clmn = np.where(np.all(winview(data,(3,3))[0],axis=(1,2)))[0][0]
data[data == 1] = 0 # set all ONEs to zero
data[0:3,clmn+3:] = 0 # set after match to zero
data[0:3,clmn:clmn+3] = 1 # restore ONEs
Another one is using Python loops and only two lines longer:
for clmn in range(0,data.shape[1]):
if np.all(data[0:3,clmn:clmn+3]):
data[data==1] = 0
data[0:3,clmn+3:] = 0
data[0:3,clmn:clmn+3] = 1
break
Instead of explaining how the above code using loops works I have put the 'explanations' into the names of the used variables so the code becomes hopefully self-explaining. With this explanations and some redundant code you can use the code below for another shaped haystack to search for in another array of same kind. For an array with more rows as the shape of the sub-array there will be necessary to loop also over the rows and optimize the code skipping some unnecessary checks.
import numpy as np
data=np.array([[4,4,4,0,1,1,1,0,0,0,0,1,0,0,1],
[3,0,0,1,1,1,1,1,1,1,1,0,0,1,0],
[6,0,0,1,1,1,1,1,0,0,0,0,1,0,0]])
indx_of_clmns_in_shape = 1
indx_of_rows_in_shape = 0
subarr_shape = (3, 3)
first_row = 0
first_clmn = 0
for clmn in range(first_clmn,data.shape[indx_of_clmns_in_shape],1):
sub_data = data[
first_row:first_row+subarr_shape[indx_of_rows_in_shape],
clmn:clmn+subarr_shape[indx_of_clmns_in_shape]]
if np.all(sub_data):
data[data == 1] = 0
data[first_row : subarr_shape[indx_of_rows_in_shape],
clmn+subarr_shape[indx_of_clmns_in_shape] : ] = 0
data[first_row : subarr_shape[indx_of_rows_in_shape],
clmn : clmn+subarr_shape[indx_of_clmns_in_shape]] = 1
break
# print(sub_data)
print(data)
all three versions of the code give the same result:
[[4 4 4 0 1 1 1 0 0 0 0 0 0 0 0]
[3 0 0 0 1 1 1 0 0 0 0 0 0 0 0]
[6 0 0 0 1 1 1 0 0 0 0 0 0 0 0]]
Should be easy to do with a double for loop and a second array
rows = len(source_array)
columns = len(source_array[0])
# Create a result array of same size
result_array = [[0 for _ in range(rows)] for _ in range(columns)]
for i in range(rows):
for j in range(columns):
# Copy non 1s
if source_array[i][j] != 1:
result_array[i][j] = source_array[i][j]
# if enough rows left to check then check
if i < rows - 3:
if j < columns - 3:
# Create set on the selected partition
elements = set(source_array[i][j:j+3] + source_array[i+1][j:j+3] + source_array[i+2][j:j+3])
# Copy 1s to new array
if len(elements) == 1 and 1 in elements:
for sq_i in range(i,i+3):
for sq_j in range(j,j+3):
result_array[sq_i][sq_j] = 1
Given a matrix M n*n (containing only 0 and 1), I want to build the matrix that contains a 1 in position (i, j) if and only if there is at least a 1 in the bottom-right submatrix M[i:n, j:n]
Please note that I know there are optimal algorithm to compute this, but for performance reasons, I'm looking for a solution using numpy (so the algorithm is fully compiled)
Example:
Given this matrix:
0 0 0 0 1
0 0 1 0 0
0 0 0 0 1
1 0 1 0 0
I'm looking for a way to compute this matrix:
0 0 0 0 1
0 0 1 1 1
0 0 1 1 1
1 1 1 1 1
Thanks
Using numpy, you can accumulate the maximum value over each axis:
import numpy as np
M = np.array([[0,0,0,0,1],
[0,0,1,0,0],
[0,0,0,0,1],
[1,0,1,0,0]])
M = np.maximum.accumulate(M)
M = np.maximum.accumulate(M,axis=1)
print(M)
[[0 0 0 0 1]
[0 0 1 1 1]
[0 0 1 1 1]
[1 1 1 1 1]]
Note: This matches your example result (presence of 1 in top-left quadrant). Your explanations of the logic would produce a different result however
If we go with M[i:n,j:n] (bottom-right):
M = np.array([[0,0,0,0,1],
[0,0,1,0,0],
[0,0,0,0,1],
[1,0,1,0,0]])
M = np.maximum.accumulate(M[::-1,:])[::-1,:]
M = np.maximum.accumulate(M[:,::-1],axis=1)[:,::-1]
print(M)
[[1 1 1 1 1]
[1 1 1 1 1]
[1 1 1 1 1]
[1 1 1 0 0]]
It is essentially the same approach except with reversed accumulation on the axes
I want to create a 64 components array showing all the squares in which the two rooks of an empty chessboard could move from their current position. So far I am doing it with for and while loops.
I first create a function just to better visualize the board:
import numpy as np
def from_array_to_matrix(v):
m=np.zeros((8,8)).astype('int')
for row in range(8):
for column in range(8):
m[row,column]=v[row*8+column]
return m
and here I show how I actually build the array:
# positions of the two rooks
a=np.zeros(64).astype('int')
a[15] = 1
a[25] = 1
print from_array_to_matrix(a)
# attack_a will be all the squares where they could move in the empty board
attack_a=np.zeros(64).astype('int')
for piece in np.where(a)[0]:
j=0
square=piece+j*8
while square<64:
attack_a[square]=1
j+=1
square=piece+j*8
j=0
square=piece-j*8
while square>=0:
attack_a[square]=1
j+=1
square=piece-j*8
j=0
square=piece+j
while square<8*(1+piece//8):
attack_a[square]=1
j+=1
square=piece+j
j=0
square=piece-j
while square>=8*(piece//8):
attack_a[square]=1
j+=1
square=piece-j
print attack_a
print from_array_to_matrix(attack_a)
I have been advised to avoid for and while loops whenever it is possible to use other ways, because they tend to be time consuming. Is there any way to achieve the same result without iterating the process with for and while loops ?
Perhaps using the fact that the indices to which I want to assign the value 1 can be determined by a function.
There are a couple of different ways to do this. The simplest thing is of course to work with matrices.
But you can vectorize operations on the raveled array as well. For example, say you had a rook at position 0 <= n < 64 in the linear array. To set the row to one, use integer division:
array[8 * (n // 8):8 * (n // 8 + 1)] = True
To set the column, use modulo:
array[n % 8::8] = True
You can convert to a matrix using reshape:
matrix = array.reshape(8, 8)
And back using ravel:
array = martix.ravel()
Or reshape:
array = matrix.reshape(-1)
Setting ones in a matrix is even simpler, given a specific row 0 <= m < 8 and column 0 <= n < 8:
matrix[m, :] = matrix[:, n] = True
Now the only question is how to vectorize multiple indices simultaneously. As it happens, you can use a fancy index in one axis. I.e, the expression above can be used with an m and n containing multiple elements:
m, n = np.nonzero(matrix)
matrix[m, :] = matrix[:, n] = True
You could even play games and do this with the array, also using fancy indexing:
n = np.nonzero(array)[0]
r = np.linspace(8 * (n // 8), 8 * (n // 8 + 1), 8, False).T.ravel()
c = np.linspace(n % 8, n % 8 + 64, 8, False)
array[r] = array[c] = True
Using linspace allows you to generate multiple sequences of the same size simultaneously. Each sequence is a column, so we transpose before raveling, although this is not required.
Use reshaping to convert 1-D array to 8x8 2-D matrix and then numpy advance indexing to select rows and columns to set to 1:
import numpy as np
def from_array_to_matrix(v):
return v.reshape(8,8)
# positions of the two rooks
a=np.zeros(64).astype('int')
a[15] = 1
a[25] = 1
a = from_array_to_matrix(a)
# attack_a will be all the squares where they could move in the empty board
attack_a=np.zeros(64).astype('int')
attack_a = from_array_to_matrix(attack_a)
#these two lines replace your for and while loops
attack_a[np.where(a)[0],:] = 1
attack_a[:,np.where(a)[1]] = 1
output:
a:
[[0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 1]
[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 0]
[0 0 0 0 0 0 0 0]
[0 0 0 0 0 0 0 0]]
attack_a:
[[0 1 0 0 0 0 0 1]
[1 1 1 1 1 1 1 1]
[0 1 0 0 0 0 0 1]
[1 1 1 1 1 1 1 1]
[0 1 0 0 0 0 0 1]
[0 1 0 0 0 0 0 1]
[0 1 0 0 0 0 0 1]
[0 1 0 0 0 0 0 1]]
I have a 3D image with size: Deep x Weight x Height (for example: 10x20x30, means 10 images, and each image has size 20x30.
Given a patch size is pd x pw x ph (such as pd <Deep, pw<Weight, ph<Height), for example patch size: 4x4x4. The center point location of the path will be: pd/2 x pw/2 x ph/2. Let's call the distance between time t and time t+1 of the center point be stride, for example stride=2.
I want to extract the original 3D image into patches with size and stride given above. How can I do it in python? Thank you
.
Use np.lib.stride_tricks.as_strided. This solution does not require the strides to divide the corresponding dimensions of the input stack. It even allows for overlapping patches (Just do not write to the result in this case, or make a copy.). It therefore is more flexible than other approaches:
import numpy as np
from numpy.lib import stride_tricks
def cutup(data, blck, strd):
sh = np.array(data.shape)
blck = np.asanyarray(blck)
strd = np.asanyarray(strd)
nbl = (sh - blck) // strd + 1
strides = np.r_[data.strides * strd, data.strides]
dims = np.r_[nbl, blck]
data6 = stride_tricks.as_strided(data, strides=strides, shape=dims)
return data6#.reshape(-1, *blck)
#demo
x = np.zeros((5, 6, 12), int)
y = cutup(x, (2, 2, 3), (3, 3, 5))
y[...] = 1
print(x[..., 0], '\n')
print(x[:, 0, :], '\n')
print(x[0, ...], '\n')
Output:
[[1 1 0 1 1 0]
[1 1 0 1 1 0]
[0 0 0 0 0 0]
[1 1 0 1 1 0]
[1 1 0 1 1 0]]
[[1 1 1 0 0 1 1 1 0 0 0 0]
[1 1 1 0 0 1 1 1 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0]
[1 1 1 0 0 1 1 1 0 0 0 0]
[1 1 1 0 0 1 1 1 0 0 0 0]]
[[1 1 1 0 0 1 1 1 0 0 0 0]
[1 1 1 0 0 1 1 1 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0]
[1 1 1 0 0 1 1 1 0 0 0 0]
[1 1 1 0 0 1 1 1 0 0 0 0]
[0 0 0 0 0 0 0 0 0 0 0 0]]
Explanation. Numpy arrays are organised in terms of strides, one for each dimension, data point [x,y,z] is located in memory at address base + stridex * x + stridey * y + stridez * z.
The stride_tricks.as_strided factory allows to directly manipulate the strides and shape of a new array sharing its memory with a given array. Try this only if you know what you're doing because no checks are performed, meaning you are allowed to shoot your foot by addressing out-of-bounds memory.
The code uses this function to split up each of the three existing dimensions into two new ones, one for the corresponding within block coordinate (this will have the same stride as the original dimension, because adjacent points in a block corrspond to adjacent points in the whole stack) and one dimension for the block index along this axis; this will have stride = original stride x block stride.
All the code does is computing the correct strides and dimensions (= block dimensions and block counts along the three axes).
Since the data are shared with the original array, when we set all points of the 6d array to 1, they are also set in the original array exposing the block structure in the demo. Note that the commented out reshape in the last line of the function breaks this link, because it forces a copy.
the skimage module offer you an integrated solution with view_as_blocks.
The source is on line.
Take care to choose Deep,Weight,Height multiple of pd, pw, ph, because as_strided do not check bounds.
I have a 2-dimensional array of integers, we'll call it "A".
I want to create a 3-dimensional array "B" of all 1s and 0s such that:
for any fixed (i,j) sum(B[i,j,:])==A[i.j], that is, B[i,j,:] contains A[i,j] 1s in it
the 1s are randomly placed in the 3rd dimension.
I know how I would do this using standard python indexing but this turns out to be very slow.
I am looking for a way to do this that takes advantage of the features that can make Numpy fast.
Here is how I would do it using standard indexing:
B=np.zeros((X,Y,Z))
indexoptions=range(Z)
for i in xrange(Y):
for j in xrange(X):
replacedindices=np.random.choice(indexoptions,size=A[i,j],replace=False)
B[i,j,[replacedindices]]=1
Can someone please explain how I can do this in a faster way?
Edit: Here is an example "A":
A=np.array([[0,1,2,3,4],[0,1,2,3,4],[0,1,2,3,4],[0,1,2,3,4],[0,1,2,3,4]])
in this case X=Y=5 and Z>=5
Essentially the same idea as #JohnZwinck and #DSM, but with a shuffle function for shuffling a given axis:
import numpy as np
def shuffle(a, axis=-1):
"""
Shuffle `a` in-place along the given axis.
Apply numpy.random.shuffle to the given axis of `a`.
Each one-dimensional slice is shuffled independently.
"""
b = a.swapaxes(axis,-1)
# Shuffle `b` in-place along the last axis. `b` is a view of `a`,
# so `a` is shuffled in place, too.
shp = b.shape[:-1]
for ndx in np.ndindex(shp):
np.random.shuffle(b[ndx])
return
def random_bits(a, n):
b = (a[..., np.newaxis] > np.arange(n)).astype(int)
shuffle(b)
return b
if __name__ == "__main__":
np.random.seed(12345)
A = np.random.randint(0, 5, size=(3,4))
Z = 6
B = random_bits(A, Z)
print "A:"
print A
print "B:"
print B
Output:
A:
[[2 1 4 1]
[2 1 1 3]
[1 3 0 2]]
B:
[[[1 0 0 0 0 1]
[0 1 0 0 0 0]
[0 1 1 1 1 0]
[0 0 0 1 0 0]]
[[0 1 0 1 0 0]
[0 0 0 1 0 0]
[0 0 1 0 0 0]
[1 0 1 0 1 0]]
[[0 0 0 0 0 1]
[0 0 1 1 1 0]
[0 0 0 0 0 0]
[0 0 1 0 1 0]]]