Avoiding for loops to speed up the computation time with numpy - python

I am working with a 1800 x 900 matrix. For each cell in the matrix, I need to do a similar operation, which can be described with an example:
a1b1 a1b2 a1b3 .... a1b900
a2b1 a2b2 ..
.
a1800b1 ... ... ..a1800b900
For each cell in a new matrix, I want its value to be equivalent to, the sum of its top left, top and top right neighbor. So, I want a2b2 = a1b1 + a1b2 + a1b3. For the top row, it could be the same value as the current cell.
I can easily do this using 2 for loops, but is there I way I can vectorize this so that it speeds up the whole process?

you can vectorize it like below
arr = np.arange(16).reshape(4,4)
arr2 = np.pad(arr, [(1,0), (1,1)], mode="constant")[:-1, :]
result = arr+arr2[:,0:-2]+arr2[:,1:-1]+arr2[:,2:]
Description
np.pad(arr, [(1,0), (1,1)], mode="constant") pad zeros to left, right and top of matrix. index [:-1, :] will remove the last row effectively shifting the matrix by one row
arr2[:,0:-2], arr2[:,1:-1], arr2[:,2:] will give top left, top and top right element for each corresponding element of arr. If there is no element it gives 0 (the zeros padded in previous step)
Result
## arr
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]])
## result
array([[ 0, 1, 2, 3],
[ 5, 8, 12, 12],
[17, 24, 28, 24],
[29, 40, 44, 36]])

Related

Is there a way to conditionally index 3D-numpy array?

Having an array A with the shape (2,6, 60), is it possible to index it based on a binary array B of shape (6,)?
The 6 and 60 is quite arbitrary, they are simply the 2D data I wish to access.
The underlying thing I am trying to do is to calculate two variants of the 2D data (in this case, (6,60)) and then efficiently select the ones with the lowest total sum - that is where the binary (6,) array comes from.
Example: For B = [1,0,1,0,1,0] what I wish to receive is equal to stacking
A[1,0,:]
A[0,1,:]
A[1,2,:]
A[0,3,:]
A[1,4,:]
A[0,5,:]
but I would like to do it by direct indexing and not a for-loop.
I have tried A[B], A[:,B,:], A[B,:,:] A[:,:,B] with none of them providing the desired (6,60) matrix.
import numpy as np
A = np.array([[4, 4, 4, 4, 4, 4], [1, 1, 1, 1, 1, 1]])
A = np.atleast_3d(A)
A = np.tile(A, (1,1,60)
B = np.array([1, 0, 1, 0, 1, 0])
A[B]
Expected results are a (6,60) array containing the elements from A as described above, the received is either (2,6,60) or (6,6,60).
Thank you in advance,
Linus
You can generate a range of the indices you want to iterate over, in your case from 0 to 5:
count = A.shape[1]
indices = np.arange(count) # np.arange(6) for your particular case
>>> print(indices)
array([0, 1, 2, 3, 4, 5])
And then you can use that to do your advanced indexing:
result_array = A[B[indices], indices, :]
If you always use the full range from 0 to length - 1 (i.e. 0 to 5 in your case) of the second axis of A in increasing order, you can simplify that to:
result_array = A[B, indices, :]
# or the ugly result_array = A[B, np.arange(A.shape[1]), :]
Or even this if it's always 6:
result_array = A[B, np.arange(6), :]
An alternative solution using np.take_along_axis (from version 1.15 - docs)
import numpy as np
x = np.arange(2*6*6).reshape((2,6,6))
m = np.zeros(6, int)
m[0] = 1
#example: [1, 0, 0, 0, 0, 0]
np.take_along_axis(x, m[None, :, None], 0) #add dimensions to mask to match array dimensions
>>array([[[36, 37, 38, 39, 40, 41],
[ 6, 7, 8, 9, 10, 11],
[12, 13, 14, 15, 16, 17],
[18, 19, 20, 21, 22, 23],
[24, 25, 26, 27, 28, 29],
[30, 31, 32, 33, 34, 35]]])

From a 2D array, create 2nd 2D array of Unique(non-repeated) random selected values from 1st array (values not shared among rows) without using a loop

This is a follow up on this question.
From a 2d array, create another 2d array composed of randomly selected values from original array (values not shared among rows) without using a loop
I am looking for a way to create a 2D array whose rows are randomly selected unique values (non-repeating) from another row, without using a loop.
Here is a way to do it With using a loop.
pool = np.random.randint(0, 30, size=[4,5])
seln = np.empty([4,3], int)
for i in range(0, pool.shape[0]):
seln[i] =np.random.choice(pool[i], 3, replace=False)
print('pool = ', pool)
print('seln = ', seln)
>pool = [[ 1 11 29 4 13]
[29 1 2 3 24]
[ 0 25 17 2 14]
[20 22 18 9 29]]
seln = [[ 8 12 0]
[ 4 19 13]
[ 8 15 24]
[12 12 19]]
Here is a method that does not uses a loop, however, it can select the same value multiple times in each row.
pool = np.random.randint(0, 30, size=[4,5])
print(pool)
array([[ 4, 18, 0, 15, 9],
[ 0, 9, 21, 26, 9],
[16, 28, 11, 19, 24],
[20, 6, 13, 2, 27]])
# New array shape
new_shape = (pool.shape[0],3)
# Indices where to randomly choose from
ix = np.random.choice(pool.shape[1], new_shape)
array([[0, 3, 3],
[1, 1, 4],
[2, 4, 4],
[1, 2, 1]])
ixs = (ix.T + range(0,np.prod(pool.shape),pool.shape[1])).T
array([[ 0, 3, 3],
[ 6, 6, 9],
[12, 14, 14],
[16, 17, 16]])
pool.flatten()[ixs].reshape(new_shape)
array([[ 4, 15, 15],
[ 9, 9, 9],
[11, 24, 24],
[ 6, 13, 6]])
I am looking for a method that does not use a loop, and if a particular value from a row is selected, that value can Not be selected again.
Here is a way without explicit looping. However, it requires generating an array of random numbers of the size of the original array. That said, the generation is done using compiled code so it should be pretty fast. It can fail if you happen to generate two identical numbers, but the chance of that happening is essentially zero.
m,n = 4,5
pool = np.random.randint(0, 30, size=[m,n])
new_width = 3
mask = np.argsort(np.random.rand(m,n))<new_width
pool[mask].reshape(m,3)
How it works:
We generate a random array of floats, and argsort it. By default, when artsort is applied to a 2d array it is applied along axis 1 so the value of the i,j entry of the argsorted list is what place the j-th entry of the i-th row would appear if you sorted the i-th row.
We then find all the values in this array where the entries whose values are less than new_width. Each row contains the numbers 0,...,n-1 in a random order, so exactly new_width of them will be less than new_width. This means each row of mask will have exactly new_width number of entries which are True, and the rest will be False (when you use a boolean operator between a ndarray and a scalar it applies it component-wise).
Finally, the boolean mask is applied to the original data to grab new_width many entries from each row.
You could also use np.vectorize for your loop solution, although that is just shorthand for a loop.

Skipping rows in numpy ndarrays slicing

Suppose I have a numpy array img, with img.shape == (468,832,3). What does img[::2, ::2] do? It reduces the shape to (234,416,3) Can you please explain the logic?
Let's read documentation together (Source).
(Just read the bold part first)
The basic slice syntax is i:j:k where i is the starting index, j is the stopping index, and k is the step (k \neq 0). This selects the m elements (in the corresponding dimension) with index values i, i + k, ..., i + (m - 1) k where m = q + (r\neq0) and q and r are the quotient and remainder obtained by dividing j - i by k: j - i = q k + r, so that i + (m - 1) k < j.
...
Assume n is the number of elements in the dimension being sliced.
Then, if i is not given it defaults to 0 for k > 0 and n - 1 for k < 0
. If j is not given it defaults to n for k > 0 and -n-1 for k < 0 . If
k is not given it defaults to 1. Note that :: is the same as : and
means select all indices along this axis.
Now looking at your part.
[::2, ::2] will be translated to [0:468:2, 0:832:2] because you do not specify the first two or i and j in the documentation. (You only specify k here. Recall the i:j:k notation above.) You select elements on these axes at the step size 2 which means you select every other elements along the axes specified.
Because you did not specify for the 3rd dimension, all will be selected.
It slices every alternate row, and then every alternate column, from an array, returning an array of size (n // 2, n // 2, ...).
Here's an example of slicing with a 2D array -
>>> a = np.arange(16).reshape(4, 4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]])
>>> a[::2, ::2]
array([[ 0, 2],
[ 8, 10]])
And, here's another example with a 3D array -
>>> a = np.arange(27).reshape(3, 3, 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, 25, 26]]])
>>> a[::2, ::2] # same as a[::2, ::2, :]
array([[[ 0, 1, 2],
[ 6, 7, 8]],
[[18, 19, 20],
[24, 25, 26]]])
Well, we have the RGB image as a 3D array of shape:
img.shape=(468,832,3)
Now, what does img[::2, ::2] do?
we're just downsampling the image (i.e. we're shrinking the image size by half by taking only every other pixel from the original image and we do this by using a step size of 2, which means to skip one pixel). This should be clear from the example below.
Let's take a simple grayscale image for easier understanding.
In [13]: arr
Out[13]:
array([[10, 11, 12, 13, 14, 15],
[20, 21, 22, 23, 24, 25],
[30, 31, 32, 33, 34, 35],
[40, 41, 42, 43, 44, 45],
[50, 51, 52, 53, 54, 55],
[60, 61, 62, 63, 64, 65]])
In [14]: arr.shape
Out[14]: (6, 6)
In [15]: arr[::2, ::2]
Out[15]:
array([[10, 12, 14],
[30, 32, 34],
[50, 52, 54]])
In [16]: arr[::2, ::2].shape
Out[16]: (3, 3)
Notice which pixels are in the sliced version. Also, observe how the array shape changes after slicing (i.e. it is reduced by half).
Now, this downsampling happens for all three channels in the image since there's no slicing happening in the third axis. Thus, you will get the shape reduced only for the first two axis in your example.
(468, 832, 3)
. . |
. . |
(234, 416, 3)

Torch sum a tensor along an axis

How do I sum over the columns of a tensor?
torch.Size([10, 100]) ---> torch.Size([10])
The simplest and best solution is to use torch.sum().
To sum all elements of a tensor:
torch.sum(x) # gives back a scalar
To sum over all rows (i.e. for each column):
torch.sum(x, dim=0) # size = [ncol]
To sum over all columns (i.e. for each row):
torch.sum(x, dim=1) # size = [nrow]
It should be noted that the dimension summed over is eliminated from the resulting tensor.
Alternatively, you can use tensor.sum(axis) where axis indicates 0 and 1 for summing over rows and columns respectively, for a 2D tensor.
In [210]: X
Out[210]:
tensor([[ 1, -3, 0, 10],
[ 9, 3, 2, 10],
[ 0, 3, -12, 32]])
In [211]: X.sum(1)
Out[211]: tensor([ 8, 24, 23])
In [212]: X.sum(0)
Out[212]: tensor([ 10, 3, -10, 52])
As, we can see from the above outputs, in both cases, the output is a 1D tensor. If you, on the other hand, wish to retain the dimension of the original tensor in the output as well, then you've set the boolean kwarg keepdim to True as in:
In [217]: X.sum(0, keepdim=True)
Out[217]: tensor([[ 10, 3, -10, 52]])
In [218]: X.sum(1, keepdim=True)
Out[218]:
tensor([[ 8],
[24],
[23]])
If you have tensor my_tensor, and you wish to sum across the second array dimension (that is, the one with index 1, which is the column-dimension, if the tensor is 2-dimensional, as yours is), use torch.sum(my_tensor,1) or equivalently my_tensor.sum(1) see documentation here.
One thing that is not mentioned explicitly in the documentation is: you can sum across the last array-dimension by using -1 (or the second-to last dimension, with -2, etc.)
So, in your example, you could use: outputs.sum(1) or torch.sum(outputs,1), or, equivalently, outputs.sum(-1) or torch.sum(outputs,-1). All of these would give the same result, an output tensor of size torch.Size([10]), with each entry being the sum over the all rows in a given column of the tensor outputs.
To illustrate with a 3-dimensional tensor:
In [1]: my_tensor = torch.arange(24).view(2, 3, 4)
Out[1]:
tensor([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
In [2]: my_tensor.sum(2)
Out[2]:
tensor([[ 6, 22, 38],
[54, 70, 86]])
In [3]: my_tensor.sum(-1)
Out[3]:
tensor([[ 6, 22, 38],
[54, 70, 86]])
Based on doc https://pytorch.org/docs/stable/generated/torch.sum.html
it should be
dim (int or tuple of python:ints) – the dimension or dimensions to reduce.
dim=0 means reduce row dimensions: condense all rows = sum by col
dim=1 means reduce col dimensions: condense cols= sum by row
Torch sum along multiple axis or dimensions
Just for the sake of completeness (I could not find it easily) I include how to sum along multiple dimensions with torch.sum which is heavily used in computer vision tasks where you have to reduce along H and W dimensions.
If you have an image x with shape C x H x W and want to compute the average pixel intensity value per channel you could do:
avg = torch.sum(x, dim=(1,2)) / (H*W) # Sum along (H,W) and norm

Python with numpy: How to delete an element from each row of a 2-D array according to a specific index

Say I have a 2-D numpy array A of size 20 x 10.
I also have an array of length 20, del_ind.
I want to delete an element from each row of A according to del_ind, to get a resultant array of size 20 x 9.
How can I do this?
I looked into np.delete with a specified axis = 1, but this only deletes element from the same position for each row.
Thanks for the help
You will probably have to build a new array.
Fortunately you can avoid python loops for this task, using fancy indexing:
h, w = 20, 10
A = np.arange(h*w).reshape(h, w)
del_ind = np.random.randint(0, w, size=h)
mask = np.ones((h,w), dtype=bool)
mask[range(h), del_ind] = False
A_ = A[mask].reshape(h, w-1)
Demo with a smaller dataset:
>>> h, w = 5, 4
>>> %paste
A = np.arange(h*w).reshape(h, w)
del_ind = np.random.randint(0, w, size=h)
mask = np.ones((h,w), dtype=bool)
mask[range(h), del_ind] = False
A_ = A[mask].reshape(h, w-1)
## -- End pasted text --
>>> A
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15],
[16, 17, 18, 19]])
>>> del_ind
array([2, 2, 1, 1, 0])
>>> A_
array([[ 0, 1, 3],
[ 4, 5, 7],
[ 8, 10, 11],
[12, 14, 15],
[17, 18, 19]])
Numpy isn't known for inplace edits; it's mainly intended for statically sized matrices. For that reason, I'd recommend doing this by copying the intended elements to a new array.
Assuming that it's sufficient to delete one column from every row:
def remove_indices(arr, indices):
result = np.empty((arr.shape[0], arr.shape[1] - 1))
for i, (delete_index, row) in enumerate(zip(indices, arr)):
result[i] = np.delete(row, delete_index)
return result

Categories