Assume I have an array like [2,3,4], I am looking for a way in NumPy (or Tensorflow) to convert it to [0,0,1,1,1,2,2,2,2] to apply tf.math.segment_sum() on a tensor that has a size of 2+3+4.
No elegant idea comes to my mind, only loops and list comprehension.
Would something like this work for you?
import numpy
arr = numpy.array([2, 3, 4])
numpy.repeat(numpy.arange(arr.size), arr)
# array([0, 0, 1, 1, 1, 2, 2, 2, 2])
You don't need to use numpy. You can use nothing but list comprehensions:
>>> foo = [2,3,4]
>>> sum([[i]*foo[i] for i in range(len(foo))], [])
[0, 0, 1, 1, 1, 2, 2, 2, 2]
It works like this:
You can create expanded arrays by multiplying a simple one with a constant, so [0] * 2 == [0,0]. So for each index in the array, we expand with [i]*foo[i]. In other words:
>>> [[i]*foo[i] for i in range(len(foo))]
[[0, 0], [1, 1, 1], [2, 2, 2, 2]]
Then we use sum to reduce the lists into a single list:
>>> sum([[i]*foo[i] for i in range(len(foo))], [])
[0, 0, 1, 1, 1, 2, 2, 2, 2]
Because we are "summing" lists, not integers, we pass [] to sum to make an empty list the starting value of the sum.
(Note that this likely will be slower than numpy, though I have not personally compared it to something like #Patol75's answer.)
I really like the answer from #Patol75 since it's neat. However, there is no pure tensorflow solution yet, so I provide one which maybe kinda complex. Just for reference and fun!
BTW, I didn't see tf.repeat this API in tf master. Please check this PR which adds tf.repeat support equivalent to numpy.repeat.
import tensorflow as tf
repeats = tf.constant([2,3,4])
values = tf.range(tf.size(repeats)) # [0,1,2]
max_repeats = tf.reduce_max(repeats) # max repeat is 4
tiled = tf.tile(tf.reshape(values, [-1,1]), [1,max_repeats]) # [[0,0,0,0],[1,1,1,1],[2,2,2,2]]
mask = tf.sequence_mask(repeats, max_repeats) # [[1,1,0,0],[1,1,1,0],[1,1,1,1]]
res = tf.boolean_mask(tiled, mask) # [0,0,1,1,1,2,2,2,2]
Patol75's answer uses Numpy but Gort the Robot's answer is actually faster (on your example list at least).
I'll keep this answer up as another solution, but it's slower than both.
Given that a = [2,3,4] this could be done using a loop like so:
b = []
for i in range(len(a)):
for j in range(a[i]):
b.append(range(len(a))[i])
Which, as a list comprehension one-liner, is this diabolical thing:
b = [range(len(a))[i] for i in range(len(a)) for j in range(a[i])]
Both end up with b = [0,0,1,1,1,2,2,2,2].
I am a little bit confused reading the documentation of argmin function in numpy.
It looks like it should do the job:
Reading this
Return the indices of the minimum values along an axis.
I might assume that
np.argmin([5, 3, 2, 1, 1, 1, 6, 1])
will return an array of all indices: which will be [3, 4, 5, 7]
But instead of this it returns only 3. Where is the catch, or what should I do to get my result?
That documentation makes more sense when you think about multidimensional arrays.
>>> x = numpy.array([[0, 1],
... [3, 2]])
>>> x.argmin(axis=0)
array([0, 0])
>>> x.argmin(axis=1)
array([0, 1])
With an axis specified, argmin takes one-dimensional subarrays along the given axis and returns the first index of each subarray's minimum value. It doesn't return all indices of a single minimum value.
To get all indices of the minimum value, you could do
numpy.where(x == x.min())
See the documentation for numpy.argmax (which is referred to by the docs for numpy.argmin):
In case of multiple occurrences of the maximum values, the indices corresponding to the first occurrence are returned.
The phrasing of the documentation ("indices" instead of "index") refers to the multidimensional case when axis is provided.
So, you can't do it with np.argmin. Instead, this will work:
np.where(arr == arr.min())
I would like to quickly add that as user grofte mentioned, np.where returns a tuple and it states that it is a shorthand for nonzero which has a corresponding method flatnonzero which returns an array directly.
So, the cleanest version seems to be
my_list = np.array([5, 3, 2, 1, 1, 1, 6, 1])
np.flatnonzero(my_list == my_list.min())
=> array([3, 4, 5, 7])
Assuming that you want the indices of a list, not a numpy array, try
import numpy as np
my_list = [5, 3, 2, 1, 1, 1, 6, 1]
np.where(np.array(my_list) == min(my_list))[0]
The index [0] is because numpy returns a tuple of your answer and nothing (answer as a numpy array). Don't ask me why.
Recommended way (by numpy documents) to get all indices of the minimum value is:
x = np.array([5, 3, 2, 1, 1, 1, 6, 1])
a, = np.nonzero(x == x.min()) # a=>array([3, 4, 5, 7])
The type of matrix I am dealing with was created from a vector as shown below:
Start with a 1-d vector V of length L.
To create a matrix A from V with N rows, make the i'th column of A the first N entries of V, starting from the i'th entry of V, so long as there are enough entries left in V to fill up the column. This means A has L - N + 1 columns.
Here is an example:
V = [0, 1, 2, 3, 4, 5]
N = 3
A =
[0 1 2 3
1 2 3 4
2 3 4 5]
Representing the matrix this way requires more memory than my machine has. Is there any reasonable way of storing this matrix sparsely? I am currently storing N * (L - N + 1) values, when I only need to store L values.
You can take a view of your original vector as follows:
>>> import numpy as np
>>> from numpy.lib.stride_tricks import as_strided
>>>
>>> v = np.array([0, 1, 2, 3, 4, 5])
>>> n = 3
>>>
>>> a = as_strided(v, shape=(n, len(v)-n+1), strides=v.strides*2)
>>> a
array([[0, 1, 2, 3],
[1, 2, 3, 4],
[2, 3, 4, 5]])
This is a view, not a copy of your original data, e.g.
>>> v[3] = 0
>>> v
array([0, 1, 2, 0, 4, 5])
>>> a
array([[0, 1, 2, 0],
[1, 2, 0, 4],
[2, 0, 4, 5]])
But you have to be careful no to do any operation on a that triggers a copy, since that would send your memory use through the ceiling.
If you're already using numpy, use its strided or sparse arrays, as Jaime explained.
If you're not already using numpy, you may to strongly consider using it.
If you need to stick with pure Python, there are three obvious ways to do this, depending on your use case.
For strided or sparse-but-clustered arrays, you could do effectively the same thing as numpy.
Or you could use a simple run-length-encoding scheme, plus maybe a higher-level list of runs for, or list of pointers to every Nth element, or even a whole stack of such lists (one for every 100 elements, one for every 10000, etc.).
But for mostly-uniformly-dense arrays, the easiest thing is to simply store a dict or defaultdict mapping indices to values. Random-access lookups or updates are still O(1)—albeit with a higher constant factor—and the storage you waste storing (in effect) a hash, key, and value instead of just a value for each non-default element is more than made up for by not storing values for the default elements, as long as you're less than 0.33 density.
I have a numerical list:
myList = [1, 2, 3, 100, 5]
Now if I sort this list to obtain [1, 2, 3, 5, 100].
What I want is the indices of the elements from the
original list in the sorted order i.e. [0, 1, 2, 4, 3]
--- ala MATLAB's sort function that returns both
values and indices.
If you are using numpy, you have the argsort() function available:
>>> import numpy
>>> numpy.argsort(myList)
array([0, 1, 2, 4, 3])
http://docs.scipy.org/doc/numpy/reference/generated/numpy.argsort.html
This returns the arguments that would sort the array or list.
Something like next:
>>> myList = [1, 2, 3, 100, 5]
>>> [i[0] for i in sorted(enumerate(myList), key=lambda x:x[1])]
[0, 1, 2, 4, 3]
enumerate(myList) gives you a list containing tuples of (index, value):
[(0, 1), (1, 2), (2, 3), (3, 100), (4, 5)]
You sort the list by passing it to sorted and specifying a function to extract the sort key (the second element of each tuple; that's what the lambda is for. Finally, the original index of each sorted element is extracted using the [i[0] for i in ...] list comprehension.
myList = [1, 2, 3, 100, 5]
sorted(range(len(myList)),key=myList.__getitem__)
[0, 1, 2, 4, 3]
I did a quick performance check on these with perfplot (a project of mine) and found that it's hard to recommend anything else but
np.argsort(x)
(note the log scale):
Code to reproduce the plot:
import perfplot
import numpy as np
def sorted_enumerate(seq):
return [i for (v, i) in sorted((v, i) for (i, v) in enumerate(seq))]
def sorted_enumerate_key(seq):
return [x for x, y in sorted(enumerate(seq), key=lambda x: x[1])]
def sorted_range(seq):
return sorted(range(len(seq)), key=seq.__getitem__)
b = perfplot.bench(
setup=np.random.rand,
kernels=[sorted_enumerate, sorted_enumerate_key, sorted_range, np.argsort],
n_range=[2 ** k for k in range(15)],
xlabel="len(x)",
)
b.save("out.png")
The answers with enumerate are nice, but I personally don't like the lambda used to sort by the value. The following just reverses the index and the value, and sorts that. So it'll first sort by value, then by index.
sorted((e,i) for i,e in enumerate(myList))
Updated answer with enumerate and itemgetter:
sorted(enumerate(a), key=lambda x: x[1])
# [(0, 1), (1, 2), (2, 3), (4, 5), (3, 100)]
Zip the lists together: The first element in the tuple will the index, the second is the value (then sort it using the second value of the tuple x[1], x is the tuple)
Or using itemgetter from the operatormodule`:
from operator import itemgetter
sorted(enumerate(a), key=itemgetter(1))
Essentially you need to do an argsort, what implementation you need depends if you want to use external libraries (e.g. NumPy) or if you want to stay pure-Python without dependencies.
The question you need to ask yourself is: Do you want the
indices that would sort the array/list
indices that the elements would have in the sorted array/list
Unfortunately the example in the question doesn't make it clear what is desired because both will give the same result:
>>> arr = np.array([1, 2, 3, 100, 5])
>>> np.argsort(np.argsort(arr))
array([0, 1, 2, 4, 3], dtype=int64)
>>> np.argsort(arr)
array([0, 1, 2, 4, 3], dtype=int64)
Choosing the argsort implementation
If you have NumPy at your disposal you can simply use the function numpy.argsort or method numpy.ndarray.argsort.
An implementation without NumPy was mentioned in some other answers already, so I'll just recap the fastest solution according to the benchmark answer here
def argsort(l):
return sorted(range(len(l)), key=l.__getitem__)
Getting the indices that would sort the array/list
To get the indices that would sort the array/list you can simply call argsort on the array or list. I'm using the NumPy versions here but the Python implementation should give the same results
>>> arr = np.array([3, 1, 2, 4])
>>> np.argsort(arr)
array([1, 2, 0, 3], dtype=int64)
The result contains the indices that are needed to get the sorted array.
Since the sorted array would be [1, 2, 3, 4] the argsorted array contains the indices of these elements in the original.
The smallest value is 1 and it is at index 1 in the original so the first element of the result is 1.
The 2 is at index 2 in the original so the second element of the result is 2.
The 3 is at index 0 in the original so the third element of the result is 0.
The largest value 4 and it is at index 3 in the original so the last element of the result is 3.
Getting the indices that the elements would have in the sorted array/list
In this case you would need to apply argsort twice:
>>> arr = np.array([3, 1, 2, 4])
>>> np.argsort(np.argsort(arr))
array([2, 0, 1, 3], dtype=int64)
In this case :
the first element of the original is 3, which is the third largest value so it would have index 2 in the sorted array/list so the first element is 2.
the second element of the original is 1, which is the smallest value so it would have index 0 in the sorted array/list so the second element is 0.
the third element of the original is 2, which is the second-smallest value so it would have index 1 in the sorted array/list so the third element is 1.
the fourth element of the original is 4 which is the largest value so it would have index 3 in the sorted array/list so the last element is 3.
If you do not want to use numpy,
sorted(range(len(seq)), key=seq.__getitem__)
is fastest, as demonstrated here.
The other answers are WRONG.
Running argsort once is not the solution.
For example, the following code:
import numpy as np
x = [3,1,2]
np.argsort(x)
yields array([1, 2, 0], dtype=int64) which is not what we want.
The answer should be to run argsort twice:
import numpy as np
x = [3,1,2]
np.argsort(np.argsort(x))
gives array([2, 0, 1], dtype=int64) as expected.
Most easiest way you can use Numpy Packages for that purpose:
import numpy
s = numpy.array([2, 3, 1, 4, 5])
sort_index = numpy.argsort(s)
print(sort_index)
But If you want that you code should use baisc python code:
s = [2, 3, 1, 4, 5]
li=[]
for i in range(len(s)):
li.append([s[i],i])
li.sort()
sort_index = []
for x in li:
sort_index.append(x[1])
print(sort_index)
We will create another array of indexes from 0 to n-1
Then zip this to the original array and then sort it on the basis of the original values
ar = [1,2,3,4,5]
new_ar = list(zip(ar,[i for i in range(len(ar))]))
new_ar.sort()
`
s = [2, 3, 1, 4, 5]
print([sorted(s, reverse=False).index(val) for val in s])
For a list with duplicate elements, it will return the rank without ties, e.g.
s = [2, 2, 1, 4, 5]
print([sorted(s, reverse=False).index(val) for val in s])
returns
[1, 1, 0, 3, 4]
Import numpy as np
FOR INDEX
S=[11,2,44,55,66,0,10,3,33]
r=np.argsort(S)
[output]=array([5, 1, 7, 6, 0, 8, 2, 3, 4])
argsort Returns the indices of S in sorted order
FOR VALUE
np.sort(S)
[output]=array([ 0, 2, 3, 10, 11, 33, 44, 55, 66])
Code:
s = [2, 3, 1, 4, 5]
li = []
for i in range(len(s)):
li.append([s[i], i])
li.sort()
sort_index = []
for x in li:
sort_index.append(x[1])
print(sort_index)
Try this, It worked for me cheers!
firstly convert your list to this:
myList = [1, 2, 3, 100, 5]
add a index to your list's item
myList = [[0, 1], [1, 2], [2, 3], [3, 100], [4, 5]]
next :
sorted(myList, key=lambda k:k[1])
result:
[[0, 1], [1, 2], [2, 3], [4, 5], [3, 100]]
A variant on RustyRob's answer (which is already the most performant pure Python solution) that may be superior when the collection you're sorting either:
Isn't a sequence (e.g. it's a set, and there's a legitimate reason to want the indices corresponding to how far an iterator must be advanced to reach the item), or
Is a sequence without O(1) indexing (among Python's included batteries, collections.deque is a notable example of this)
Case #1 is unlikely to be useful, but case #2 is more likely to be meaningful. In either case, you have two choices:
Convert to a list/tuple and use the converted version, or
Use a trick to assign keys based on iteration order
This answer provides the solution to #2. Note that it's not guaranteed to work by the language standard; the language says each key will be computed once, but not the order they will be computed in. On every version of CPython, the reference interpreter, to date, it's precomputed in order from beginning to end, so this works, but be aware it's not guaranteed. In any event, the code is:
sizediterable = ...
sorted_indices = sorted(range(len(sizediterable)), key=lambda _, it=iter(sizediterable): next(it))
All that does is provide a key function that ignores the value it's given (an index) and instead provides the next item from an iterator preconstructed from the original container (cached as a defaulted argument to allow it to function as a one-liner). As a result, for something like a large collections.deque, where using its .__getitem__ involves O(n) work (and therefore computing all the keys would involve O(n²) work), sequential iteration remains O(1), so generating the keys remains just O(n).
If you need something guaranteed to work by the language standard, using built-in types, Roman's solution will have the same algorithmic efficiency as this solution (as neither of them rely on the algorithmic efficiency of indexing the original container).
To be clear, for the suggested use case with collections.deque, the deque would have to be quite large for this to matter; deques have a fairly large constant divisor for indexing, so only truly huge ones would have an issue. Of course, by the same token, the cost of sorting is pretty minimal if the inputs are small/cheap to compare, so if your inputs are large enough that efficient sorting matters, they're large enough for efficient indexing to matter too.