Related
Given two arrays, a and b, how to find efficiently all combinations of elements in b that have equal value in a?
here is an example:
Given
a = [0, 0, 0, 1, 1, 2, 2, 2, 2]
b = [1, 2, 4, 5, 9, 3, 7, 22, 10]
how would you calculate
c = [[1, 2],
[1, 4],
[2, 4],
[5, 9],
[3, 7],
[3, 22],
[3, 10],
[7, 22],
[7, 10],
[22, 10]]
?
a can be assumed to be sorted.
I can do this with loops, a la:
import torch
a = torch.tensor([0, 0, 0, 1, 1, 2, 2, 2, 2])
b = torch.tensor([1, 2, 4, 5, 9, 3, 7, 22, 10])
jumps = torch.cat((torch.tensor([0]),
torch.where(a.diff() > 0)[0] + 1,
torch.tensor([len(a)])))
cs = []
for i in range(len(jumps) - 1):
cs.append(torch.combinations(b[jumps[i]:jumps[i + 1]]))
c = torch.cat(cs)
Is there any efficient way to avoid the loop? The solution should work for CPU and CUDA.
Also, the solution should have runtime O(m * m), where m is the largest number of equal elements in a and not O(n * n) where n is the length of of a.
I prefer solutions for pytorch, but I am curious for solution for numpy as well.
I think the overhead of using torch is only justified for bigger datasets, as there is basically no computational difficulty in the function, imho you can achieve same results with:
from collections import Counter
def find_combinations1(a, b):
count_a = Counter(a)
combinations = []
for x in set(b):
if count_a[x] == b.count(x):
combinations.append(x)
return combinations
or even a simpler:
def find_combinations2(a, b):
return list(set(a) & set(b))
With pytorch I assume the most simple approach is:
import torch
def find_combinations3(a, b):
a = torch.tensor(a)
b = torch.tensor(b)
eq = torch.eq(a, b.view(-1, 1))
indices = torch.nonzero(eq)
return indices[:, 1]
This option has of course a time complexity of O(n*m) where n is the size of a and m is the size of b, and O(n+m) is the memory for the tensors.
I am trying to access a pytorch tensor by a matrix of indices and I recently found this bit of code that I cannot find the reason why it is not working.
The code below is split into two parts. The first half proves to work, whilst the second trips an error. I fail to see the reason why. Could someone shed some light on this?
import torch
import numpy as np
a = torch.rand(32, 16)
m, n = a.shape
xx, yy = np.meshgrid(np.arange(m), np.arange(m))
result = a[xx] # WORKS for a torch.tensor of size M >= 32. It doesn't work otherwise.
a = torch.rand(16, 16)
m, n = a.shape
xx, yy = np.meshgrid(np.arange(m), np.arange(m))
result = a[xx] # IndexError: too many indices for tensor of dimension 2
and if I change a = np.random.rand(16, 16) it does work as well.
To whoever comes looking for an answer: it looks like its a bug in pyTorch.
Indexing using numpy arrays is not well defined, and it works only if tensors are indexed using tensors. So, in my example code, this works flawlessly:
a = torch.rand(M, N)
m, n = a.shape
xx, yy = torch.meshgrid(torch.arange(m), torch.arange(m), indexing='xy')
result = a[xx] # WORKS
I made a gist to check it, and it's available here
First, let me give you a quick insight into the idea of indexing a tensor with a numpy array and another tensor.
Example: this is our target tensor to be indexed
numpy_indices = torch.tensor([[0, 1, 2, 7],
[0, 1, 2, 3]]) # numpy array
tensor_indices = torch.tensor([[0, 1, 2, 7],
[0, 1, 2, 3]]) # 2D tensor
t = torch.tensor([[1, 2, 3, 4], # targeted tensor
[5, 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]])
numpy_result = t[numpy_indices]
tensor_result = t[tensor_indices]
Indexing using a 2D numpy array: the index is read like pairs (x,y) tensor[row,column] e.g. t[0,0], t[1,1], t[2,2], and t[7,3].
print(numpy_result) # tensor([ 1, 6, 11, 32])
Indexing using a 2D tensor: walks through the index tensor in a row-wise manner and each value is an index of a row in the targeted tensor.
e.g. [ [t[0],t[1],t[2],[7]] , [[0],[1],[2],[3]] ] see the example below, the new shape of tensor_result after indexing is (tensor_indices.shape[0],tensor_indices.shape[1],t.shape[1])=(2,4,4).
print(tensor_result) # tensor([[[ 1, 2, 3, 4],
# [ 5, 6, 7, 8],
# [ 9, 10, 11, 12],
# [29, 30, 31, 32]],
# [[ 1, 2, 3, 4],
# [ 5, 6, 7, 8],
# [ 9, 10, 11, 12],
# [ 13, 14, 15, 16]]])
If you try to add a third row in numpy_indices, you will get the same error you have because the index will be represented by 3D e.g., (0,0,0)...(7,3,3).
indices = np.array([[0, 1, 2, 7],
[0, 1, 2, 3],
[0, 1, 2, 3]])
print(numpy_result) # IndexError: too many indices for tensor of dimension 2
However, this is not the case with indexing by tensor and the shape will be bigger (3,4,4).
Finally, as you see the outputs of the two types of indexing are completely different. To solve your problem, you can use
xx = torch.tensor(xx).long() # convert a numpy array to a tensor
What happens in the case of advanced indexing (rows of numpy_indices > 3 ) as your situation is still ambiguous and unsolved and you can check 1 , 2, 3.
I'm looking for a way to select multiple slices from a numpy array at once. Say we have a 1D data array and want to extract three portions of it like below:
data_extractions = []
for start_index in range(0, 3):
data_extractions.append(data[start_index: start_index + 5])
Afterwards data_extractions will be:
data_extractions = [
data[0:5],
data[1:6],
data[2:7]
]
Is there any way to perform above operation without the for loop? Some sort of indexing scheme in numpy that would let me select multiple slices from an array and return them as that many arrays, say in an n+1 dimensional array?
I thought maybe I can replicate my data and then select a span from each row, but code below throws an IndexError
replicated_data = np.vstack([data] * 3)
data_extractions = replicated_data[[range(3)], [slice(0, 5), slice(1, 6), slice(2, 7)]
You can use the indexes to select the rows you want into the appropriate shape.
For example:
data = np.random.normal(size=(100,2,2,2))
# Creating an array of row-indexes
indexes = np.array([np.arange(0,5), np.arange(1,6), np.arange(2,7)])
# data[indexes] will return an element of shape (3,5,2,2,2). Converting
# to list happens along axis 0
data_extractions = list(data[indexes])
np.all(data_extractions[1] == data[1:6])
True
The final comparison is against the original data.
stride_tricks can do that
a = np.arange(10)
b = np.lib.stride_tricks.as_strided(a, (3, 5), 2 * a.strides)
b
# array([[0, 1, 2, 3, 4],
# [1, 2, 3, 4, 5],
# [2, 3, 4, 5, 6]])
Please note that b references the same memory as a, in fact multiple times (for example b[0, 1] and b[1, 0] are the same memory address). It is therefore safest to make a copy before working with the new structure.
nd can be done in a similar fashion, for example 2d -> 4d
a = np.arange(16).reshape(4, 4)
b = np.lib.stride_tricks.as_strided(a, (3,3,2,2), 2*a.strides)
b.reshape(9,2,2) # this forces a copy
# array([[[ 0, 1],
# [ 4, 5]],
# [[ 1, 2],
# [ 5, 6]],
# [[ 2, 3],
# [ 6, 7]],
# [[ 4, 5],
# [ 8, 9]],
# [[ 5, 6],
# [ 9, 10]],
# [[ 6, 7],
# [10, 11]],
# [[ 8, 9],
# [12, 13]],
# [[ 9, 10],
# [13, 14]],
# [[10, 11],
# [14, 15]]])
In this post is an approach with strided-indexing scheme using np.lib.stride_tricks.as_strided that basically creates a view into the input array and as such is pretty efficient for creation and being a view occupies nomore memory space.
Also, this works for ndarrays with generic number of dimensions.
Here's the implementation -
def strided_axis0(a, L):
# Store the shape and strides info
shp = a.shape
s = a.strides
# Compute length of output array along the first axis
nd0 = shp[0]-L+1
# Setup shape and strides for use with np.lib.stride_tricks.as_strided
# and get (n+1) dim output array
shp_in = (nd0,L)+shp[1:]
strd_in = (s[0],) + s
return np.lib.stride_tricks.as_strided(a, shape=shp_in, strides=strd_in)
Sample run for a 4D array case -
In [44]: a = np.random.randint(11,99,(10,4,2,3)) # Array
In [45]: L = 5 # Window length along the first axis
In [46]: out = strided_axis0(a, L)
In [47]: np.allclose(a[0:L], out[0]) # Verify outputs
Out[47]: True
In [48]: np.allclose(a[1:L+1], out[1])
Out[48]: True
In [49]: np.allclose(a[2:L+2], out[2])
Out[49]: True
You can slice your array with a prepared slicing array
a = np.array(list('abcdefg'))
b = np.array([
[0, 1, 2, 3, 4],
[1, 2, 3, 4, 5],
[2, 3, 4, 5, 6]
])
a[b]
However, b doesn't have to generated by hand in this way. It can be more dynamic with
b = np.arange(5) + np.arange(3)[:, None]
In the general case you have to do some sort of iteration - and concatenation - either when constructing the indexes or when collecting the results. It's only when the slicing pattern is itself regular that you can use a generalized slicing via as_strided.
The accepted answer constructs an indexing array, one row per slice. So that is iterating over the slices, and arange itself is a (fast) iteration. And np.array concatenates them on a new axis (np.stack generalizes this).
In [264]: np.array([np.arange(0,5), np.arange(1,6), np.arange(2,7)])
Out[264]:
array([[0, 1, 2, 3, 4],
[1, 2, 3, 4, 5],
[2, 3, 4, 5, 6]])
indexing_tricks convenience methods to do the same thing:
In [265]: np.r_[0:5, 1:6, 2:7]
Out[265]: array([0, 1, 2, 3, 4, 1, 2, 3, 4, 5, 2, 3, 4, 5, 6])
This takes the slicing notation, expands it with arange and concatenates. It even lets me expand and concatenate into 2d
In [269]: np.r_['0,2',0:5, 1:6, 2:7]
Out[269]:
array([[0, 1, 2, 3, 4],
[1, 2, 3, 4, 5],
[2, 3, 4, 5, 6]])
In [270]: data=np.array(list('abcdefghijk'))
In [272]: data[np.r_['0,2',0:5, 1:6, 2:7]]
Out[272]:
array([['a', 'b', 'c', 'd', 'e'],
['b', 'c', 'd', 'e', 'f'],
['c', 'd', 'e', 'f', 'g']],
dtype='<U1')
In [273]: data[np.r_[0:5, 1:6, 2:7]]
Out[273]:
array(['a', 'b', 'c', 'd', 'e', 'b', 'c', 'd', 'e', 'f', 'c', 'd', 'e',
'f', 'g'],
dtype='<U1')
Concatenating results after indexing also works.
In [274]: np.stack([data[0:5],data[1:6],data[2:7]])
My memory from other SO questions is that relative timings are in the same order of magnitude. It may vary for example with the number of slices versus their length. Overall the number of values that have to be copied from source to target will be the same.
If the slices vary in length, you'd have to use the flat indexing.
No matter which approach you choose, if 2 slices contain same element, it doesn't support mathematical operations correctly unlesss you use ufunc.at which can be more inefficient than loop. For testing:
def as_strides(arr, window_size, stride, writeable=False):
'''Get a strided sub-matrices view of a 4D ndarray.
Args:
arr (ndarray): input array with shape (batch_size, m1, n1, c).
window_size (tuple): with shape (m2, n2).
stride (tuple): stride of windows in (y_stride, x_stride).
writeable (bool): it is recommended to keep it False unless needed
Returns:
subs (view): strided window view, with shape (batch_size, y_nwindows, x_nwindows, m2, n2, c)
See also numpy.lib.stride_tricks.sliding_window_view
'''
batch_size = arr.shape[0]
m1, n1, c = arr.shape[1:]
m2, n2 = window_size
y_stride, x_stride = stride
view_shape = (batch_size, 1 + (m1 - m2) // y_stride,
1 + (n1 - n2) // x_stride, m2, n2, c)
strides = (arr.strides[0], y_stride * arr.strides[1],
x_stride * arr.strides[2]) + arr.strides[1:]
subs = np.lib.stride_tricks.as_strided(arr,
view_shape,
strides=strides,
writeable=writeable)
return subs
import numpy as np
np.random.seed(1)
Xs = as_strides(np.random.randn(1, 5, 5, 2), (3, 3), (2, 2), writeable=True)[0]
print('input\n0,0\n', Xs[0, 0])
np.add.at(Xs, np.s_[:], 5)
print('unbuffered sum output\n0,0\n', Xs[0,0])
np.add.at(Xs, np.s_[:], -5)
Xs = Xs + 5
print('normal sum output\n0,0\n', Xs[0, 0])
We can use list comprehension for this
data=np.array([1,2,3,4,5,6,7,8,9,10])
data_extractions=[data[b:b+5] for b in [1,2,3,4,5]]
data_extractions
Results
[array([2, 3, 4, 5, 6]), array([3, 4, 5, 6, 7]), array([4, 5, 6, 7, 8]), array([5, 6, 7, 8, 9]), array([ 6, 7, 8, 9, 10])]
Lets say I have a Python Numpy array a.
a = numpy.array([1,2,3,4,5,6,7,8,9,10,11])
I want to create a matrix of sub sequences from this array of length 5 with stride 3. The results matrix hence will look as follows:
numpy.array([[1,2,3,4,5],[4,5,6,7,8],[7,8,9,10,11]])
One possible way of implementing this would be using a for-loop.
result_matrix = np.zeros((3, 5))
for i in range(0, len(a), 3):
result_matrix[i] = a[i:i+5]
Is there a cleaner way to implement this in Numpy?
Approach #1 : Using broadcasting -
def broadcasting_app(a, L, S ): # Window len = L, Stride len/stepsize = S
nrows = ((a.size-L)//S)+1
return a[S*np.arange(nrows)[:,None] + np.arange(L)]
Approach #2 : Using more efficient NumPy strides -
def strided_app(a, L, S ): # Window len = L, Stride len/stepsize = S
nrows = ((a.size-L)//S)+1
n = a.strides[0]
return np.lib.stride_tricks.as_strided(a, shape=(nrows,L), strides=(S*n,n))
Sample run -
In [143]: a
Out[143]: array([ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11])
In [144]: broadcasting_app(a, L = 5, S = 3)
Out[144]:
array([[ 1, 2, 3, 4, 5],
[ 4, 5, 6, 7, 8],
[ 7, 8, 9, 10, 11]])
In [145]: strided_app(a, L = 5, S = 3)
Out[145]:
array([[ 1, 2, 3, 4, 5],
[ 4, 5, 6, 7, 8],
[ 7, 8, 9, 10, 11]])
Starting in Numpy 1.20, we can make use of the new sliding_window_view to slide/roll over windows of elements.
And coupled with a stepping [::3], it simply becomes:
from numpy.lib.stride_tricks import sliding_window_view
# values = np.array([1,2,3,4,5,6,7,8,9,10,11])
sliding_window_view(values, window_shape = 5)[::3]
# array([[ 1, 2, 3, 4, 5],
# [ 4, 5, 6, 7, 8],
# [ 7, 8, 9, 10, 11]])
where the intermediate result of the sliding is:
sliding_window_view(values, window_shape = 5)
# array([[ 1, 2, 3, 4, 5],
# [ 2, 3, 4, 5, 6],
# [ 3, 4, 5, 6, 7],
# [ 4, 5, 6, 7, 8],
# [ 5, 6, 7, 8, 9],
# [ 6, 7, 8, 9, 10],
# [ 7, 8, 9, 10, 11]])
Modified version of #Divakar's code with checking to ensure that memory is contiguous and that the returned array cannot be modified. (Variable names changed for my DSP application).
def frame(a, framelen, frameadv):
"""frame - Frame a 1D array
a - 1D array
framelen - Samples per frame
frameadv - Samples between starts of consecutive frames
Set to framelen for non-overlaping consecutive frames
Modified from Divakar's 10/17/16 11:20 solution:
https://stackoverflow.com/questions/40084931/taking-subarrays-from-numpy-array-with-given-stride-stepsize
CAVEATS:
Assumes array is contiguous
Output is not writable as there are multiple views on the same memory
"""
if not isinstance(a, np.ndarray) or \
not (a.flags['C_CONTIGUOUS'] or a.flags['F_CONTIGUOUS']):
raise ValueError("Input array a must be a contiguous numpy array")
# Output
nrows = ((a.size-framelen)//frameadv)+1
oshape = (nrows, framelen)
# Size of each element in a
n = a.strides[0]
# Indexing in the new object will advance by frameadv * element size
ostrides = (frameadv*n, n)
return np.lib.stride_tricks.as_strided(a, shape=oshape,
strides=ostrides, writeable=False)
Suppose I have the array [1,2,3,4,5].
I want to "add" the array [2,4,6,8] to it so I get
[[3,5,7,9],
[4,6,8,10],
[5,7,9,11],
[6,8,10,12],
[7,9,11,13]]
(or its transpose).
There is probably a function for this but I can't seem to find it because I'm not sure what to search for.
As suggested by #Divakar, the best way is to use add.outer:
a1 = np.array([1,2,3,4,5])
a2 = np.array([2,4,6,8])
np.add.outer(a1,a2)
But you can also explicitely broadcast your a1 array to the proper shape, then add to a2:
a1[:,None]+a2
# essentially equivalent to:
# a1.reshape(-1,1)+a2
Both produce:
array([[ 3, 5, 7, 9],
[ 4, 6, 8, 10],
[ 5, 7, 9, 11],
[ 6, 8, 10, 12],
[ 7, 9, 11, 13]])