Related
I'm using python 3.X and I want to create such iterator that will allow me to iterate a matrix from cell [N,0] to [0,N]
I don't want to use indices-magic so I tried np.nditer which is not enough for that.
a = np.matrix(np.random.randint(0,3,(3,3)))
>>>([[0, 0, 1],
[1, 1, 2],
[1, 2, 2]])
it = np.nditer(a, flags=['f_index'])
for i in range(a.size):
print(it[0])
it.iternext()
>>>0 0 1 1 1 2 1 2 2
I want to get the following :
1,2,2,1,1,2,0,0,1
Is it possible using iterators of some kind?
In [29]: arr = np.array([[0,0,1],[1,1,2],[1,2,2]])
In [30]: arr[::-1,:]
Out[30]:
array([[1, 2, 2],
[1, 1, 2],
[0, 0, 1]])
In [31]: arr[::-1,:].ravel()
Out[31]: array([1, 2, 2, 1, 1, 2, 0, 0, 1])
If I want to create a matrix, I simply call
m = np.matrix([[x00, x01],
[x10, x11]])
, where x00, x01, x10 and x11 are numbers. However, I would like to vectorize this process. For example, if the x's are one-dimensional arrays with length l, then I would like m to become an array of matrices, or a lx2x2-dimensional array. Unfortunately,
zeros = np.zeros(10)
ones = np.ones(10)
m = np.matrix([[zeros, ones],
[zeros, ones]])
raises an error ("matrix must be 2-dimensional") and
m = np.array([[zeros, ones],
[zeros, ones]])
gives an 2x2xl-dimensional array instead. In order to solve this, I could call np.moveaxis(m, 2, 0), but I am looking for a direct solution that doesn't need to change the order of axes of a (potentially huge) array. This also only sets the axis-order right if I'm passing one-dimensional arrays as values for my matrix, not if they're higher dimensional.
Is there a general and efficient way of vectorizing the creation of matrices?
Let's try a 2d (4d after joining) case:
In [374]: ones = np.ones((3,4),int)
In [375]: arr = np.array([[ones*0, ones],[ones*2, ones*3]])
In [376]: arr
Out[376]:
array([[[[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]],
[[1, 1, 1, 1],
[1, 1, 1, 1],
[1, 1, 1, 1]]],
[[[2, 2, 2, 2],
[2, 2, 2, 2],
[2, 2, 2, 2]],
[[3, 3, 3, 3],
[3, 3, 3, 3],
[3, 3, 3, 3]]]])
In [377]: arr.shape
Out[377]: (2, 2, 3, 4)
Notice that the original array elements are 'together'. arr has its own databuffer, with copies of the original arrays, but it was made with relatively efficient block copies.
We can easily transpose axes:
In [378]: arr.transpose(2,3,0,1)
Out[378]:
array([[[[0, 1],
[2, 3]],
[[0, 1],
[2, 3]],
...
[[0, 1],
[2, 3]]]])
Now it's 12 (2,2) arrays. It is a view, using arr's databuffer. It just has a different shape and strides. Doing this transpose is quite efficient, and isn't any slower when arr is very big. And a lot of math on the transposed array will be nearly as efficient as on the original arr (because of stridded iteration). If there are differences in speed it will be because of caching at a deep level.
But some actions will require a copy. For example the transposed array can't be raveled without a copy. The original 0s,1s etc are no longer together.
In [379]: arr.transpose(2,3,0,1).ravel()
Out[379]:
array([0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1,
2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3,
0, 1, 2, 3])
I could construct the same 1d array with
In [380]: tarr = np.empty((3,4,2,2), int)
In [381]: tarr[...,0,0] = ones*0
In [382]: tarr[...,0,1] = ones*1
In [383]: tarr[...,1,0] = ones*2
In [384]: tarr[...,1,1] = ones*3
In [385]: tarr.ravel()
Out[385]:
array([0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1,
2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3,
0, 1, 2, 3])
This tarr is effectively what you are trying to produce 'directly'.
Another way to look at this construction, is to assign the values to the array's .flat with strides - insert 0s at every 4th slot, 1s at the adjacent ones, etc.:
In [386]: tarr.flat[0::4] = ones*0
In [387]: tarr.flat[1::4] = ones*1
In [388]: tarr.flat[2::4] = ones*2
In [389]: tarr.flat[3::4] = ones*3
Here's another 'direct' way - use np.stack (a version of concatenate) to create a (3,4,4) array, which can then be reshaped:
np.stack((ones*0,ones*1,ones*2,ones*3),2).reshape(3,4,2,2)
That stack is, in essence:
In [397]: ones1 = ones[...,None]
In [398]: np.concatenate((ones1*0, ones1*1, ones1*2, ones1*3),axis=2)
Notice that this target (3,4,2,2) could be reshaped to (12,4) (and v.v) at no cost. So the original problem becomes: is it easier to construct a (4,12) and transpose, or construct the (12,4) first? It's really a 2d problem, not a (m+n)d one.
np.matrix must be a 2D array. From numpy documentation of np.matrix
Returns a matrix from an array-like object, or from a string of data.
A matrix is a specialized 2-D array that retains its 2-D nature
through operations. It has certain special operators, such as *
(matrix multiplication) and ** (matrix power).
Note
It is no longer recommended to use this class, even for linear
algebra. Instead use regular arrays. The class may be removed in the
future.
Is there any reason you want np.matrix? Most numpy operations should be doable in the array object as the matrix class is quasi-deprecated.
From your example I tried using the transpose (.T) method:
zeros = np.zeros(10)
ones = np.ones(10)
twos = np.ones(10) * 2
threes = np.ones(10) * 3
m = np.array([[zeros, ones], [twos, threes]]).T
>> array([[0,2],[1,3]],...)
or
m = np.transpose(np.array([[zeros, ones], [twos, threes]]), (2,0,1))
>> array([[0,1],[2,3]],...)
This yields a (10, 2, 2) array
I have a list of unique rows and another larger array of data (called test_rows in example). I was wondering if there was a faster way to get the location of each unique row in the data. The fastest way that I could come up with is...
import numpy
uniq_rows = numpy.array([[0, 1, 0],
[1, 1, 0],
[1, 1, 1],
[0, 1, 1]])
test_rows = numpy.array([[0, 1, 1],
[0, 1, 0],
[0, 0, 0],
[1, 1, 0],
[0, 1, 0],
[0, 1, 1],
[0, 1, 1],
[1, 1, 1],
[1, 1, 0],
[1, 1, 1],
[0, 1, 0],
[0, 0, 0],
[1, 1, 0]])
# this gives me the indexes of each group of unique rows
for row in uniq_rows.tolist():
print row, numpy.where((test_rows == row).all(axis=1))[0]
This prints...
[0, 1, 0] [ 1 4 10]
[1, 1, 0] [ 3 8 12]
[1, 1, 1] [7 9]
[0, 1, 1] [0 5 6]
Is there a better or more numpythonic (not sure if that word exists) way to do this? I was searching for a numpy group function but could not find it. Basically for any incoming dataset I need the fastest way to get the locations of each unique row in that data set. The incoming dataset will not always have every unique row or the same number.
EDIT:
This is just a simple example. In my application the numbers would not be just zeros and ones, they could be anywhere from 0 to 32000. The size of uniq rows could be between 4 to 128 rows and the size of test_rows could be in the hundreds of thousands.
Numpy
From version 1.13 of numpy you can use numpy.unique like np.unique(test_rows, return_counts=True, return_index=True, axis=1)
Pandas
df = pd.DataFrame(test_rows)
uniq = pd.DataFrame(uniq_rows)
uniq
0 1 2
0 0 1 0
1 1 1 0
2 1 1 1
3 0 1 1
Or you could generate the unique rows automatically from the incoming DataFrame
uniq_generated = df.drop_duplicates().reset_index(drop=True)
yields
0 1 2
0 0 1 1
1 0 1 0
2 0 0 0
3 1 1 0
4 1 1 1
and then look for it
d = dict()
for idx, row in uniq.iterrows():
d[idx] = df.index[(df == row).all(axis=1)].values
This is about the same as your where method
d
{0: array([ 1, 4, 10], dtype=int64),
1: array([ 3, 8, 12], dtype=int64),
2: array([7, 9], dtype=int64),
3: array([0, 5, 6], dtype=int64)}
There are a lot of solutions here, but I'm adding one with vanilla numpy. In most cases numpy will be faster than list comprehensions and dictionaries, although the array broadcasting may cause memory to be an issue if large arrays are used.
np.where((uniq_rows[:, None, :] == test_rows).all(2))
Wonderfully simple, eh? This returns a tuple of unique row indices and the corresponding test row.
(array([0, 0, 0, 1, 1, 1, 2, 2, 3, 3, 3]),
array([ 1, 4, 10, 3, 8, 12, 7, 9, 0, 5, 6]))
How it works:
(uniq_rows[:, None, :] == test_rows)
Uses array broadcasting to compare each element of test_rows with each row in uniq_rows. This results in a 4x13x3 array. all is used to determine which rows are equal (all comparisons returned true). Finally, where returns the indices of these rows.
With the np.unique from v1.13 (downloaded from the source link on the latest documentation, https://github.com/numpy/numpy/blob/master/numpy/lib/arraysetops.py#L112-L247)
In [157]: aset.unique(test_rows, axis=0,return_inverse=True,return_index=True)
Out[157]:
(array([[0, 0, 0],
[0, 1, 0],
[0, 1, 1],
[1, 1, 0],
[1, 1, 1]]),
array([2, 1, 0, 3, 7], dtype=int32),
array([2, 1, 0, 3, 1, 2, 2, 4, 3, 4, 1, 0, 3], dtype=int32))
In [158]: a,b,c=_
In [159]: c
Out[159]: array([2, 1, 0, 3, 1, 2, 2, 4, 3, 4, 1, 0, 3], dtype=int32)
In [164]: from collections import defaultdict
In [165]: dd = defaultdict(list)
In [166]: for i,v in enumerate(c):
...: dd[v].append(i)
...:
In [167]: dd
Out[167]:
defaultdict(list,
{0: [2, 11],
1: [1, 4, 10],
2: [0, 5, 6],
3: [3, 8, 12],
4: [7, 9]})
or indexing the dictionary with the unique rows (as hashable tuple):
In [170]: dd = defaultdict(list)
In [171]: for i,v in enumerate(c):
...: dd[tuple(a[v])].append(i)
...:
In [172]: dd
Out[172]:
defaultdict(list,
{(0, 0, 0): [2, 11],
(0, 1, 0): [1, 4, 10],
(0, 1, 1): [0, 5, 6],
(1, 1, 0): [3, 8, 12],
(1, 1, 1): [7, 9]})
This will do the job:
import numpy as np
uniq_rows = np.array([[0, 1, 0],
[1, 1, 0],
[1, 1, 1],
[0, 1, 1]])
test_rows = np.array([[0, 1, 1],
[0, 1, 0],
[0, 0, 0],
[1, 1, 0],
[0, 1, 0],
[0, 1, 1],
[0, 1, 1],
[1, 1, 1],
[1, 1, 0],
[1, 1, 1],
[0, 1, 0],
[0, 0, 0],
[1, 1, 0]])
indices=np.where(np.sum(np.abs(np.repeat(uniq_rows,len(test_rows),axis=0)-np.tile(test_rows,(len(uniq_rows),1))),axis=1)==0)[0]
loc=indices//len(test_rows)
indices=indices-loc*len(test_rows)
res=[[] for i in range(len(uniq_rows))]
for i in range(len(indices)):
res[loc[i]].append(indices[i])
print(res)
[[1, 4, 10], [3, 8, 12], [7, 9], [0, 5, 6]]
This will work for all the cases including the cases in which not all the rows in uniq_rows are present in test_rows. However, if somehow you know ahead that all of them are present, you could replace the part
res=[[] for i in range(len(uniq_rows))]
for i in range(len(indices)):
res[loc[i]].append(indices[i])
with just the row:
res=np.split(indices,np.where(np.diff(loc)>0)[0]+1)
Thus avoiding loops entirely.
Not very 'numpythonic', but for a bit of an upfront cost, we can make a dict with the keys as a tuple of your row, and a list of indices:
test_rowsdict = {}
for i,j in enumerate(test_rows):
test_rowsdict.setdefault(tuple(j),[]).append(i)
test_rowsdict
{(0, 0, 0): [2, 11],
(0, 1, 0): [1, 4, 10],
(0, 1, 1): [0, 5, 6],
(1, 1, 0): [3, 8, 12],
(1, 1, 1): [7, 9]}
Then you can filter based on your uniq_rows, with a fast dict lookup: test_rowsdict[tuple(row)]:
out = []
for i in uniq_rows:
out.append((i, test_rowsdict.get(tuple(i),[])))
For your data, I get 16us for just the lookup, and 66us for building and looking up, versus 95us for your np.where solution.
Approach #1
Here's one approach, not sure about the level of "NumPythonic-ness" though to such a tricky problem -
def get1Ds(a, b): # Get 1D views of each row from the two inputs
# check that casting to void will create equal size elements
assert a.shape[1:] == b.shape[1:]
assert a.dtype == b.dtype
# compute dtypes
void_dt = np.dtype((np.void, a.dtype.itemsize * a.shape[1]))
# convert to 1d void arrays
a = np.ascontiguousarray(a)
b = np.ascontiguousarray(b)
a_void = a.reshape(a.shape[0], -1).view(void_dt).ravel()
b_void = b.reshape(b.shape[0], -1).view(void_dt).ravel()
return a_void, b_void
def matching_row_indices(uniq_rows, test_rows):
A, B = get1Ds(uniq_rows, test_rows)
validA_mask = np.in1d(A,B)
sidx_A = A.argsort()
validA_mask = validA_mask[sidx_A]
sidx = B.argsort()
sortedB = B[sidx]
split_idx = np.flatnonzero(sortedB[1:] != sortedB[:-1])+1
all_split_indx = np.split(sidx, split_idx)
match_mask = np.in1d(B,A)[sidx]
valid_mask = np.logical_or.reduceat(match_mask, np.r_[0, split_idx])
locations = [e for i,e in enumerate(all_split_indx) if valid_mask[i]]
return uniq_rows[sidx_A[validA_mask]], locations
Scope(s) of improvement (on performance) :
np.split could be replaced by a for-loop for splitting using slicing.
np.r_ could be replaced by np.concatenate.
Sample run -
In [331]: unq_rows, idx = matching_row_indices(uniq_rows, test_rows)
In [332]: unq_rows
Out[332]:
array([[0, 1, 0],
[0, 1, 1],
[1, 1, 0],
[1, 1, 1]])
In [333]: idx
Out[333]: [array([ 1, 4, 10]),array([0, 5, 6]),array([ 3, 8, 12]),array([7, 9])]
Approach #2
Another approach to beat the setup overhead from the previous one and making use of get1Ds from it, would be -
A, B = get1Ds(uniq_rows, test_rows)
idx_group = []
for row in A:
idx_group.append(np.flatnonzero(B == row))
The numpy_indexed package (disclaimer: I am its author) was created to solve problems of this kind in an elegant and efficient manner:
import numpy_indexed as npi
indices = np.arange(len(test_rows))
unique_test_rows, index_groups = npi.group_by(test_rows, indices)
If you dont care about the indices of all rows, but only those present in test_rows, npi has a bunch of simple ways of tackling that problem too; f.i:
subset_indices = npi.indices(unique_test_rows, unique_rows)
As a sidenote; it might be useful to take a look at the examples in the npi library; in my experience, most of the time people ask a question of this kind, these grouped indices are just a means to an end, and not the endgoal of the computation. Chances are that using the functionality in npi you can reach that end goal more efficiently, without ever explicitly computing those indices. Do you care to give some more background to your problem?
EDIT: if you arrays are indeed this big, and always consist of a small number of columns with binary values, wrapping them with the following encoding might boost efficiency a lot further still:
def encode(rows):
return (rows * [[2**i for i in range(rows.shape[1])]]).sum(axis=1, dtype=np.uint8)
I am seeing behaviour with numpy bincount that I cannot make sense of. I want to bin the values in a 2D array in a row-wise manner and see the behaviour below. Why would it work with dbArray but fail with simarray?
>>> dbArray
array([[1, 0, 1, 0, 1],
[1, 1, 1, 1, 1],
[1, 1, 0, 1, 1],
[1, 0, 0, 0, 0],
[0, 0, 0, 1, 1],
[0, 1, 0, 1, 0]])
>>> N.apply_along_axis(N.bincount,1,dbArray)
array([[2, 3],
[0, 5],
[1, 4],
[4, 1],
[3, 2],
[3, 2]], dtype=int64)
>>> simarray
array([[2, 0, 2, 0, 2],
[2, 1, 2, 1, 2],
[2, 1, 1, 1, 2],
[2, 0, 1, 0, 1],
[1, 0, 1, 1, 2],
[1, 1, 1, 1, 1]])
>>> N.apply_along_axis(N.bincount,1,simarray)
Traceback (most recent call last):
File "<pyshell#31>", line 1, in <module>
N.apply_along_axis(N.bincount,1,simarray)
File "C:\Python27\lib\site-packages\numpy\lib\shape_base.py", line 118, in apply_along_axis
outarr[tuple(i.tolist())] = res
ValueError: could not broadcast input array from shape (2) into shape (3)
The problem is that bincount isn't always returning the same shaped objects, in particular when values are missing. For example:
>>> m = np.array([[0,0,1],[1,1,0],[1,1,1]])
>>> np.apply_along_axis(np.bincount, 1, m)
array([[2, 1],
[1, 2],
[0, 3]])
>>> [np.bincount(m[i]) for i in range(m.shape[1])]
[array([2, 1]), array([1, 2]), array([0, 3])]
works, but:
>>> m = np.array([[0,0,0],[1,1,0],[1,1,0]])
>>> m
array([[0, 0, 0],
[1, 1, 0],
[1, 1, 0]])
>>> [np.bincount(m[i]) for i in range(m.shape[1])]
[array([3]), array([1, 2]), array([1, 2])]
>>> np.apply_along_axis(np.bincount, 1, m)
Traceback (most recent call last):
File "<ipython-input-49-72e06e26a718>", line 1, in <module>
np.apply_along_axis(np.bincount, 1, m)
File "/usr/local/lib/python2.7/dist-packages/numpy/lib/shape_base.py", line 117, in apply_along_axis
outarr[tuple(i.tolist())] = res
ValueError: could not broadcast input array from shape (2) into shape (1)
won't.
You could use the minlength parameter and pass it using a lambda or partial or something:
>>> np.apply_along_axis(lambda x: np.bincount(x, minlength=2), axis=1, arr=m)
array([[3, 0],
[1, 2],
[1, 2]])
As #DSM has already mentioned, bincount of a 2d array cannot be done without knowing the maximum value of the array, because it would mean an inconsistency of array sizes.
But thanks to the power of numpy's indexing, it was fairly easy to make a faster implementation of 2d bincount, as it doesn't use concatenation or anything.
def bincount2d(arr, bins=None):
if bins is None:
bins = np.max(arr) + 1
count = np.zeros(shape=[len(arr), bins], dtype=np.int64)
indexing = np.arange(len(arr))
for col in arr.T:
count[indexing, col] += 1
return count
t = np.array([[1,2,3],[4,5,6],[3,2,2]], dtype=np.int64)
print(bincount2d(t))
P.S.
This:
t = np.empty(shape=[10000, 100], dtype=np.int64)
s = time.time()
bincount2d(t)
e = time.time()
print(e - s)
gives ~2 times faster result, than this:
t = np.empty(shape=[100, 10000], dtype=np.int64)
s = time.time()
bincount2d(t)
e = time.time()
print(e - s)
because of the for loop iterating over columns. So, it's better to transpose your 2d array, if shape[0] < shape[1].
UPD
Better than this can't be done (using python alone, I mean):
def bincount2d(arr, bins=None):
if bins is None:
bins = np.max(arr) + 1
count = np.zeros(shape=[len(arr), bins], dtype=np.int64)
indexing = (np.ones_like(arr).T * np.arange(len(arr))).T
np.add.at(count, (indexing, arr), 1)
return count
This is a function that does exactly what you want, but without any loops.
def sub_sum_partition(a, partition):
"""
Generalization of np.bincount(partition, a).
Sums rows of a matrix for each value of array of non-negative ints.
:param a: array_like
:param partition: array_like, 1 dimension, nonnegative ints
:return: matrix of shape ('one larger than the largest value in partition', a.shape[1:]). The i's element is
the sum of rows j in 'a' s.t. partition[j] == i
"""
assert partition.shape == (len(a),)
n = np.prod(a.shape[1:], dtype=int)
bins = ((np.tile(partition, (n, 1)) * n).T + np.arange(n, dtype=int)).reshape(-1)
sums = np.bincount(bins, a.reshape(-1))
if n > 1:
sums = sums.reshape(-1, *a.shape[1:])
return sums
I went through these threads:
Find unique rows in numpy.array
Removing duplicates in each row of a numpy array
Pandas: unique dataframe
and they all discuss several methods for computing the matrix with unique rows and columns.
However, the solutions look a bit convoluted, at least to the untrained eye. Here is for example top solution from the first thread, which (correct me if I am wrong) I believe it is the safest and fastest:
np.unique(a.view(np.dtype((np.void, a.dtype.itemsize*a.shape[1])))).view(a.dtype).reshape(-1,
a.shape[1])
Either way, the above solution only returns the matrix of unique rows. What I am looking for is something along the original functionality of np.unique
u, indices = np.unique(a, return_inverse=True)
which returns, not only the list of unique entries, but also the membership of each item to each unique entry found, but how can I do this for columns?
Here is an example of what I am looking for:
array([[0, 2, 0, 2, 2, 0, 2, 1, 1, 2],
[0, 1, 0, 1, 1, 1, 2, 2, 2, 2]])
We would have:
u = array([0,1,2,3,4])
indices = array([0,1,0,1,1,3,4,4,3])
Where the different values in u represent the set of unique columns in the original array:
0 -> [0,0]
1 -> [2,1]
2 -> [0,1]
3 -> [2,2]
4 -> [1,2]
First lets get the unique indices, to do so we need to start by transposing your array:
>>> a=a.T
Using a modified version of the above to get unique indices.
>>> ua, uind = np.unique(np.ascontiguousarray(a).view(np.dtype((np.void,a.dtype.itemsize * a.shape[1]))),return_inverse=True)
>>> uind
array([0, 3, 0, 3, 3, 1, 4, 2, 2, 4])
#Thanks to #Jamie
>>> ua = ua.view(a.dtype).reshape(ua.shape + (-1,))
>>> ua
array([[0, 0],
[0, 1],
[1, 2],
[2, 1],
[2, 2]])
For sanity:
>>> np.all(a==ua[uind])
True
To reproduce your chart:
>>> for x in range(ua.shape[0]):
... print x,'->',ua[x]
...
0 -> [0 0]
1 -> [0 1]
2 -> [1 2]
3 -> [2 1]
4 -> [2 2]
To do exactly what you ask, but will be a bit slower if it has to convert the array:
>>> b=np.asfortranarray(a).view(np.dtype((np.void,a.dtype.itemsize * a.shape[0])))
>>> ua,uind=np.unique(b,return_inverse=True)
>>> uind
array([0, 3, 0, 3, 3, 1, 4, 2, 2, 4])
>>> ua.view(a.dtype).reshape(ua.shape+(-1,),order='F')
array([[0, 0, 1, 2, 2],
[0, 1, 2, 1, 2]])
#To return this in the previous order.
>>> ua.view(a.dtype).reshape(ua.shape + (-1,))
Essentially, you want np.unique to return the indexes of the unique columns, and the indices of where they're used? This is easy enough to do by transposing the matrix and then using the code from the other question, with the addition of return_inverse=True.
at = a.T
b = np.ascontiguousarray(at).view(np.dtype((np.void, at.dtype.itemsize * at.shape[1])))
_, u, indices = np.unique(b, return_index=True, return_inverse=True)
With your a, this gives:
In [35]: u
Out[35]: array([0, 5, 7, 1, 6])
In [36]: indices
Out[36]: array([0, 3, 0, 3, 3, 1, 4, 2, 2, 4])
It's not entirely clear to me what you want u to be, however. If you want it to be the unique columns, then you could use the following instead:
at = a.T
b = np.ascontiguousarray(at).view(np.dtype((np.void, at.dtype.itemsize * at.shape[1])))
_, idx, indices = np.unique(b, return_index=True, return_inverse=True)
u = a[:,idx]
This would give
In [41]: u
Out[41]:
array([[0, 0, 1, 2, 2],
[0, 1, 2, 1, 2]])
In [42]: indices
Out[42]: array([0, 3, 0, 3, 3, 1, 4, 2, 2, 4])
Not entirely sure what you are after, but have a look at the numpy_indexed package (disclaimer: I am its author); it is sure to make problems of this kind easier:
import numpy_indexed as npi
unique_columns = npi.unique(A, axis=1)
# or perhaps this is what you want?
unique_columns, indices = npi.group_by(A.T, np.arange(A.shape[1])))