I have two vectors / one-dimensional numpy arrays and a function I want to apply:
arr1 = np.arange(1, 5)
arr2 = np.arange(2, 6)
func = lambda x, y: x * y
I now want to construct a n * m matrix (with n, m being the lengths of arr1, and arr2 respectively) containing the values of the function outputs. The naive approach using for loops would look like this:
np.array([[func(x, y) for x in arr1] for y in arr2])
I was wondering if there is a smarter vectorized approach using the arr1[:, None] syntax to apply my function - please note my actual function is significantly more complicated and can't be broken down to simple numpy operations (arr1[:, None] * arr2[None, :] won't work).
When you have numpy.array, One approach can be numpy.einsum. Because you want to compute this : arr1_i * arr2_j -> insert to arr_result_ji.
>>> np.einsum('i, j -> ji', arr1, arr2)
array([[ 2, 4, 6, 8],
[ 3, 6, 9, 12],
[ 4, 8, 12, 16],
[ 5, 10, 15, 20]])
Or you can use numpy.matmul or use #.
>>> np.matmul(arr2[:,None], arr1[None,:])
# OR
>>> arr2[:,None] # arr1[None,:]
# Or by thanks #hpaulj by elementwise multiplication with broadcasting
>>> arr2[:,None] * arr1[None,:]
array([[ 2, 4, 6, 8],
[ 3, 6, 9, 12],
[ 4, 8, 12, 16],
[ 5, 10, 15, 20]])
Here is some comparison between your loop approach and #I'mahdi 's approach:
import time
arr1 = np.arange(1, 10000)
arr2 = np.arange(2, 10001)
start = time.time()
np.array([[func(x, y) for x in arr1] for y in arr2])
print('loop: __time__', time.time()-start)
start = time.time()
(arr1[:, None]*arr2[None, :]).T
print('* __time__', time.time()-start)
start = time.time()
np.einsum('i, j -> ji', arr1, arr2)
print('einsum __time__', time.time()-start)
start = time.time()
np.matmul(arr2[:,None], arr1[None,:])
print('matmul __time__', time.time()-start)
Output:
loop: __time__ 70.3061535358429
* __time__ 0.43536829948425293
einsum __time__ 0.508014440536499
matmul __time__ 0.7149899005889893
Suppose you have 3 tensors of the same size:
a = torch.randn(3,3)
a = ([[ 0.1945, 0.8583, 2.6479],
[-0.1000, 1.2136, -0.3706],
[-0.0094, 0.4279, -0.6840]])
b = torch.randn(3, 3)
b = ([[-1.1155, 0.2106, -0.2183],
[ 1.6610, -0.6953, 0.0052],
[-0.8955, 0.0953, -0.7737]])
c = torch.randn(3, 3)
c = ([[-0.2303, -0.3427, -0.4990],
[-1.1254, 0.4432, 0.3999],
[ 0.2489, -0.9459, -0.5576]])
In Lua (torch7), they have this function:
[self] map2(tensor1, tensor2, function(x, xt1, xt2))
which applies the given function to all elements of self.
My questions are:
Is there any similar function in python (pytorch)?
Is there any pythonic method to iterate over the 3 tensors and get the respective elements of each tensor without using for loop and indices?
For example:
0.1945 -1.1155 -0.2303
0.8583 0.2106 -0.3427
2.6479 -0.2183 -0.4990
-0.1000 1.6610 -1.1254
...
Edit_1: I have also tried itertools.zip_longest and zip, but the results are not as I expected as mentioned above
You can use Python's map function similar to what you have mentioned. Like this:
>>> tensor_list = [torch.tensor([i, i, i]) for i in range(3)]
>>> list(map(lambda x: x**2, tensor_list))
[tensor([0, 0, 0]), tensor([1, 1, 1]), tensor([4, 4, 4])]
>>>
EDIT: For a PyTorch only approach you can use torch.Tensor.apply_ (Note this does the changes in place and doesn't return a new tensor)
>>> x = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
>>> x.apply_(lambda y: y ** 2)
tensor([[ 1, 4, 9],
[16, 25, 36],
[49, 64, 81]])
>>>
Is there a more efficient way in determining the averages of a certain area in a given numpy array? For simplicity, lets say I have a 5x5 array:
values = np.array([[0, 1, 2, 3, 4],
[1, 2, 3, 4, 5],
[2, 3, 4, 5, 6],
[3, 4, 5, 6, 7],
[4, 5, 6, 7, 8]])
I would like to get the averages of each coordinate, with a specified area size, assuming the array wraps around. Lets say the certain area is size 2, thus anything around a certain point within distance 2 will be considered. For example, to get the average of the area from coordinate (2,2), we need to consider
2,
2, 3, 4,
2, 3, 4, 5, 6
4, 5, 6,
6,
Thus, the average will be 4.
For coordinate (4, 4) we need to consider:
6,
6, 7, 3,
6, 7, 8, 4, 5
3, 4, 0,
5,
Thus the average will be 4.92.
Currently, I have the following code below. But since I have a for loop I feel like it could be improved. Is there a way to just use numpy built in functions?
Is there a way to use np.vectorize to gather the subarrays (area), place it all in an array, then use np.einsum or something.
def get_average(matrix, loc, dist):
sum = 0
num = 0
size, size = matrix.shape
for y in range(-dist, dist + 1):
for x in range(-dist + abs(y), dist - abs(y) + 1):
y_ = (y + loc.y) % size
x_ = (x + loc.x) % size
sum += matrix[y_, x_]
num += 1
return sum/num
class Coord():
def __init__(self, x, y):
self.x = x
self.y = y
values = np.array([[0, 1, 2, 3, 4],
[1, 2, 3, 4, 5],
[2, 3, 4, 5, 6],
[3, 4, 5, 6, 7],
[4, 5, 6, 7, 8]])
height, width = values.shape
averages = np.zeros((height, width), dtype=np.float16)
for r in range(height):
for c in range(width):
loc = Coord(c, r)
averages[r][c] = get_average(values, loc, 2)
print(averages)
Output:
[[ 3.07617188 2.92382812 3.5390625 4.15234375 4. ]
[ 2.92382812 2.76953125 3.38476562 4. 3.84570312]
[ 3.5390625 3.38476562 4. 4.6171875 4.4609375 ]
[ 4.15234375 4. 4.6171875 5.23046875 5.078125 ]
[ 4. 3.84570312 4.4609375 5.078125 4.921875 ]]
This solution is less efficient (slower) than yours but is just an example using numpy.ma module.
Required libraries:
import numpy as np
import numpy.ma as ma
Define methods to do the job:
# build the shape of the area as a rhomboid
def rhomboid2(dim):
size = 2*dim + 1
matrix = np.ones((size,size))
for y in range(-dim, dim + 1):
for x in range(-dim + abs(y), dim - abs(y) + 1):
matrix[(y + dim) % size, (x + dim) % size] = 0
return matrix
# build a mask using the area shaped
def mask(matrix_shape, rhom_dim):
mask = np.zeros(matrix_shape)
bound = 2*rhom_dim+1
rhom = rhomboid2(rhom_dim)
mask[0:bound, 0:bound] = rhom
# roll to set the position of the rhomboid to 0,0
mask = np.roll(mask,-rhom_dim, axis = 0)
mask = np.roll(mask,-rhom_dim, axis = 1)
return mask
Then, iterate to build the result:
mask_ = mask((5,5), 2) # call the mask sized as values array with a rhomboid area of size 2
averages = np.zeros_like(values, dtype=np.float16) # initialize the recipient
# iterate over the mask to calculate the average
for y in range(len(mask_)):
for x in range(len(mask_)):
masked = ma.array(values, mask = mask_)
averages[y,x] = np.mean(masked)
mask_ = np.roll(mask_, 1, axis = 1)
mask_ = np.roll(mask_, 1, axis = 0)
Which returns
# [[3.076 2.924 3.54 4.152 4. ]
# [2.924 2.77 3.385 4. 3.846]
# [3.54 3.385 4. 4.617 4.46 ]
# [4.152 4. 4.617 5.23 5.08 ]
# [4. 3.846 4.46 5.08 4.92 ]]
import numpy
square = numpy.reshape(range(0,16),(4,4))
square
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]])
In the above array, how do I access the primary diagonal and secondary diagonal of any given element? For example 9.
by primary diagonal, I mean - [4,9,14],
by secondary diagonal, I mean - [3,6,9,12]
I can't use numpy.diag() cause it takes the entire array to get the diagonal.
Base on your description, with np.where, np.diagonal and np.fliplr
import numpy as np
x,y=np.where(square==9)
np.diagonal(square, offset=-(x-y))
Out[382]: array([ 4, 9, 14])
x,y=np.where(np.fliplr(square)==9)
np.diagonal(np.fliplr(square), offset=-(x-y))
# base on the op's comment it should be np.diagonal(np.fliplr(square), offset=-(x-y))
Out[396]: array([ 3, 6, 9, 12])
For the first diagonal, use the fact that both x_coordiante and y_coordinate increase with 1 each step:
def first_diagonal(x, y, length_array):
if x < y:
return zip(range(x, length_array), range(length_array - x))
else:
return zip(range(length_array - y), range(y, length_array))
For the secondary diagonal, use the fact that the x_coordinate + y_coordinate = constant.
def second_diagonal(x, y, length_array):
tot = x + y
return zip(range(tot+1), range(tot, -1, -1))
This gives you two lists you can use to access your matrix.
Of course, if you have a non square matrix these functions will have to be reshaped a bit.
To illustrate how to get the desired output:
a = np.reshape(range(0,16),(4,4))
first = first_diagonal(1, 2, len(a))
second = second_diagonal(1,2, len(a))
primary_diagonal = [a[i[0]][i[1]] for i in first]
secondary_diagonal = [a[i[0]][i[1]] for i in second]
print(primary_diagonal)
print(secondary_diagonal)
this outputs:
[4, 9, 14]
[3, 6, 9, 12]
Is there a way to slice a 2d array in numpy into smaller 2d arrays?
Example
[[1,2,3,4], -> [[1,2] [3,4]
[5,6,7,8]] [5,6] [7,8]]
So I basically want to cut down a 2x4 array into 2 2x2 arrays. Looking for a generic solution to be used on images.
There was another question a couple of months ago which clued me in to the idea of using reshape and swapaxes. The h//nrows makes sense since this keeps the first block's rows together. It also makes sense that you'll need nrows and ncols to be part of the shape. -1 tells reshape to fill in whatever number is necessary to make the reshape valid. Armed with the form of the solution, I just tried things until I found the formula that works.
You should be able to break your array into "blocks" using some combination of reshape and swapaxes:
def blockshaped(arr, nrows, ncols):
"""
Return an array of shape (n, nrows, ncols) where
n * nrows * ncols = arr.size
If arr is a 2D array, the returned array should look like n subblocks with
each subblock preserving the "physical" layout of arr.
"""
h, w = arr.shape
assert h % nrows == 0, f"{h} rows is not evenly divisible by {nrows}"
assert w % ncols == 0, f"{w} cols is not evenly divisible by {ncols}"
return (arr.reshape(h//nrows, nrows, -1, ncols)
.swapaxes(1,2)
.reshape(-1, nrows, ncols))
turns c
np.random.seed(365)
c = np.arange(24).reshape((4, 6))
print(c)
[out]:
[[ 0 1 2 3 4 5]
[ 6 7 8 9 10 11]
[12 13 14 15 16 17]
[18 19 20 21 22 23]]
into
print(blockshaped(c, 2, 3))
[out]:
[[[ 0 1 2]
[ 6 7 8]]
[[ 3 4 5]
[ 9 10 11]]
[[12 13 14]
[18 19 20]]
[[15 16 17]
[21 22 23]]]
I've posted an inverse function, unblockshaped, here, and an N-dimensional generalization here. The generalization gives a little more insight into the reasoning behind this algorithm.
Note that there is also superbatfish's
blockwise_view. It arranges the
blocks in a different format (using more axes) but it has the advantage of (1)
always returning a view and (2) being capable of handling arrays of any
dimension.
It seems to me that this is a task for numpy.split or some variant.
e.g.
a = np.arange(30).reshape([5,6]) #a.shape = (5,6)
a1 = np.split(a,3,axis=1)
#'a1' is a list of 3 arrays of shape (5,2)
a2 = np.split(a, [2,4])
#'a2' is a list of three arrays of shape (2,5), (2,5), (1,5)
If you have a NxN image you can create, e.g., a list of 2 NxN/2 subimages, and then divide them along the other axis.
numpy.hsplit and numpy.vsplit are also available.
There are some other answers that seem well-suited for your specific case already, but your question piqued my interest in the possibility of a memory-efficient solution usable up to the maximum number of dimensions that numpy supports, and I ended up spending most of the afternoon coming up with possible method. (The method itself is relatively simple, it's just that I still haven't used most of the really fancy features that numpy supports so most of the time was spent researching to see what numpy had available and how much it could do so that I didn't have to do it.)
def blockgen(array, bpa):
"""Creates a generator that yields multidimensional blocks from the given
array(_like); bpa is an array_like consisting of the number of blocks per axis
(minimum of 1, must be a divisor of the corresponding axis size of array). As
the blocks are selected using normal numpy slicing, they will be views rather
than copies; this is good for very large multidimensional arrays that are being
blocked, and for very large blocks, but it also means that the result must be
copied if it is to be modified (unless modifying the original data as well is
intended)."""
bpa = np.asarray(bpa) # in case bpa wasn't already an ndarray
# parameter checking
if array.ndim != bpa.size: # bpa doesn't match array dimensionality
raise ValueError("Size of bpa must be equal to the array dimensionality.")
if (bpa.dtype != np.int # bpa must be all integers
or (bpa < 1).any() # all values in bpa must be >= 1
or (array.shape % bpa).any()): # % != 0 means not evenly divisible
raise ValueError("bpa ({0}) must consist of nonzero positive integers "
"that evenly divide the corresponding array axis "
"size".format(bpa))
# generate block edge indices
rgen = (np.r_[:array.shape[i]+1:array.shape[i]//blk_n]
for i, blk_n in enumerate(bpa))
# build slice sequences for each axis (unfortunately broadcasting
# can't be used to make the items easy to operate over
c = [[np.s_[i:j] for i, j in zip(r[:-1], r[1:])] for r in rgen]
# Now to get the blocks; this is slightly less efficient than it could be
# because numpy doesn't like jagged arrays and I didn't feel like writing
# a ufunc for it.
for idxs in np.ndindex(*bpa):
blockbounds = tuple(c[j][idxs[j]] for j in range(bpa.size))
yield array[blockbounds]
You question practically the same as this one. You can use the one-liner with np.ndindex() and reshape():
def cutter(a, r, c):
lenr = a.shape[0]/r
lenc = a.shape[1]/c
np.array([a[i*r:(i+1)*r,j*c:(j+1)*c] for (i,j) in np.ndindex(lenr,lenc)]).reshape(lenr,lenc,r,c)
To create the result you want:
a = np.arange(1,9).reshape(2,1)
#array([[1, 2, 3, 4],
# [5, 6, 7, 8]])
cutter( a, 1, 2 )
#array([[[[1, 2]],
# [[3, 4]]],
# [[[5, 6]],
# [[7, 8]]]])
Some minor enhancement to TheMeaningfulEngineer's answer that handles the case when the big 2d array cannot be perfectly sliced into equally sized subarrays
def blockfy(a, p, q):
'''
Divides array a into subarrays of size p-by-q
p: block row size
q: block column size
'''
m = a.shape[0] #image row size
n = a.shape[1] #image column size
# pad array with NaNs so it can be divided by p row-wise and by q column-wise
bpr = ((m-1)//p + 1) #blocks per row
bpc = ((n-1)//q + 1) #blocks per column
M = p * bpr
N = q * bpc
A = np.nan* np.ones([M,N])
A[:a.shape[0],:a.shape[1]] = a
block_list = []
previous_row = 0
for row_block in range(bpc):
previous_row = row_block * p
previous_column = 0
for column_block in range(bpr):
previous_column = column_block * q
block = A[previous_row:previous_row+p, previous_column:previous_column+q]
# remove nan columns and nan rows
nan_cols = np.all(np.isnan(block), axis=0)
block = block[:, ~nan_cols]
nan_rows = np.all(np.isnan(block), axis=1)
block = block[~nan_rows, :]
## append
if block.size:
block_list.append(block)
return block_list
Examples:
a = np.arange(25)
a = a.reshape((5,5))
out = blockfy(a, 2, 3)
a->
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19],
[20, 21, 22, 23, 24]])
out[0] ->
array([[0., 1., 2.],
[5., 6., 7.]])
out[1]->
array([[3., 4.],
[8., 9.]])
out[-1]->
array([[23., 24.]])
For now it just works when the big 2d array can be perfectly sliced into equally sized subarrays.
The code bellow slices
a ->array([[ 0, 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10, 11],
[12, 13, 14, 15, 16, 17],
[18, 19, 20, 21, 22, 23]])
into this
block_array->
array([[[ 0, 1, 2],
[ 6, 7, 8]],
[[ 3, 4, 5],
[ 9, 10, 11]],
[[12, 13, 14],
[18, 19, 20]],
[[15, 16, 17],
[21, 22, 23]]])
p ang q determine the block size
Code
a = arange(24)
a = a.reshape((4,6))
m = a.shape[0] #image row size
n = a.shape[1] #image column size
p = 2 #block row size
q = 3 #block column size
block_array = []
previous_row = 0
for row_block in range(blocks_per_row):
previous_row = row_block * p
previous_column = 0
for column_block in range(blocks_per_column):
previous_column = column_block * q
block = a[previous_row:previous_row+p,previous_column:previous_column+q]
block_array.append(block)
block_array = array(block_array)
If you want a solution that also handles the cases when the matrix is
not equally divided, you can use this:
from operator import add
half_split = np.array_split(input, 2)
res = map(lambda x: np.array_split(x, 2, axis=1), half_split)
res = reduce(add, res)
Here is a solution based on unutbu's answer that handle case where matrix cannot be equally divided. In this case, it will resize the matrix before using some interpolation. You need OpenCV for this. Note that I had to swap ncols and nrows to make it works, didn't figured why.
import numpy as np
import cv2
import math
def blockshaped(arr, r_nbrs, c_nbrs, interp=cv2.INTER_LINEAR):
"""
arr a 2D array, typically an image
r_nbrs numbers of rows
r_cols numbers of cols
"""
arr_h, arr_w = arr.shape
size_w = int( math.floor(arr_w // c_nbrs) * c_nbrs )
size_h = int( math.floor(arr_h // r_nbrs) * r_nbrs )
if size_w != arr_w or size_h != arr_h:
arr = cv2.resize(arr, (size_w, size_h), interpolation=interp)
nrows = int(size_w // r_nbrs)
ncols = int(size_h // c_nbrs)
return (arr.reshape(r_nbrs, ncols, -1, nrows)
.swapaxes(1,2)
.reshape(-1, ncols, nrows))
a = np.random.randint(1, 9, size=(9,9))
out = [np.hsplit(x, 3) for x in np.vsplit(a,3)]
print(a)
print(out)
yields
[[7 6 2 4 4 2 5 2 3]
[2 3 7 6 8 8 2 6 2]
[4 1 3 1 3 8 1 3 7]
[6 1 1 5 7 2 1 5 8]
[8 8 7 6 6 1 8 8 4]
[6 1 8 2 1 4 5 1 8]
[7 3 4 2 5 6 1 2 7]
[4 6 7 5 8 2 8 2 8]
[6 6 5 5 6 1 2 6 4]]
[[array([[7, 6, 2],
[2, 3, 7],
[4, 1, 3]]), array([[4, 4, 2],
[6, 8, 8],
[1, 3, 8]]), array([[5, 2, 3],
[2, 6, 2],
[1, 3, 7]])], [array([[6, 1, 1],
[8, 8, 7],
[6, 1, 8]]), array([[5, 7, 2],
[6, 6, 1],
[2, 1, 4]]), array([[1, 5, 8],
[8, 8, 4],
[5, 1, 8]])], [array([[7, 3, 4],
[4, 6, 7],
[6, 6, 5]]), array([[2, 5, 6],
[5, 8, 2],
[5, 6, 1]]), array([[1, 2, 7],
[8, 2, 8],
[2, 6, 4]])]]
I publish my solution. Notice that this code doesn't' actually create copies of original array, so it works well with big data. Moreover, it doesn't crash if array cannot be divided evenly (but you can easly add condition for that by deleting ceil and checking if v_slices and h_slices are divided without rest).
import numpy as np
from math import ceil
a = np.arange(9).reshape(3, 3)
p, q = 2, 2
width, height = a.shape
v_slices = ceil(width / p)
h_slices = ceil(height / q)
for h in range(h_slices):
for v in range(v_slices):
block = a[h * p : h * p + p, v * q : v * q + q]
# do something with a block
This code changes (or, more precisely, gives you direct access to part of an array) this:
[[0 1 2]
[3 4 5]
[6 7 8]]
Into this:
[[0 1]
[3 4]]
[[2]
[5]]
[[6 7]]
[[8]]
If you need actual copies, Aenaon code is what you are looking for.
If you are sure that big array can be divided evenly, you can use numpy splitting tools.
to add to #Aenaon answer and his blockfy function, if you are working with COLOR IMAGES/ 3D ARRAY here is my pipeline to create crops of 224 x 224 for 3 channel input
def blockfy(a, p, q):
'''
Divides array a into subarrays of size p-by-q
p: block row size
q: block column size
'''
m = a.shape[0] #image row size
n = a.shape[1] #image column size
# pad array with NaNs so it can be divided by p row-wise and by q column-wise
bpr = ((m-1)//p + 1) #blocks per row
bpc = ((n-1)//q + 1) #blocks per column
M = p * bpr
N = q * bpc
A = np.nan* np.ones([M,N])
A[:a.shape[0],:a.shape[1]] = a
block_list = []
previous_row = 0
for row_block in range(bpc):
previous_row = row_block * p
previous_column = 0
for column_block in range(bpr):
previous_column = column_block * q
block = A[previous_row:previous_row+p, previous_column:previous_column+q]
# remove nan columns and nan rows
nan_cols = np.all(np.isnan(block), axis=0)
block = block[:, ~nan_cols]
nan_rows = np.all(np.isnan(block), axis=1)
block = block[~nan_rows, :]
## append
if block.size:
block_list.append(block)
return block_list
then extended above to
for file in os.listdir(path_to_crop): ### list files in your folder
img = io.imread(path_to_crop + file, as_gray=False) ### open image
r = blockfy(img[:,:,0],224,224) ### crop blocks of 224 x 224 for red channel
g = blockfy(img[:,:,1],224,224) ### crop blocks of 224 x 224 for green channel
b = blockfy(img[:,:,2],224,224) ### crop blocks of 224 x 224 for blue channel
for x in range(0,len(r)):
img = np.array((r[x],g[x],b[x])) ### combine each channel into one patch by patch
img = img.astype(np.uint8) ### cast back to proper integers
img_swap = img.swapaxes(0, 2) ### need to swap axes due to the way things were proceesed
img_swap_2 = img_swap.swapaxes(0, 1) ### do it again
Image.fromarray(img_swap_2).save(path_save_crop+str(x)+"bounding" + file,
format = 'jpeg',
subsampling=0,
quality=100) ### save patch with new name etc