I have a piece of code in Matlab that I want to convert into Python/numpy.
I have a matrix ind which has the dimensions (32768, 24). I have another matrix X which has the dimensions (98304, 6). When I perform the operation
result = X(ind)
the shape of the matrix is (32768, 24).
but in numpy when I perform the same shape
result = X[ind]
I get the shape of the result matrix as (32768, 24, 6).
I would greatly appreciate it if someone can help me with why I can these two different results and how can I fix them. I would want to get the shape (32768, 24) for the result matrix in numpy as well
In Octave, if I define:
>> X=diag([1,2,3,4])
X =
Diagonal Matrix
1 0 0 0
0 2 0 0
0 0 3 0
0 0 0 4
>> idx = [6 7;10 11]
idx =
6 7
10 11
then the indexing selects a block:
>> X(idx)
ans =
2 0
0 3
The numpy equivalent is
In [312]: X=np.diag([1,2,3,4])
In [313]: X
Out[313]:
array([[1, 0, 0, 0],
[0, 2, 0, 0],
[0, 0, 3, 0],
[0, 0, 0, 4]])
In [314]: idx = np.array([[5,6],[9,10]]) # shifted for 0 base indexing
In [315]: np.unravel_index(idx,(4,4)) # raveled to unraveled conversion
Out[315]:
(array([[1, 1],
[2, 2]]),
array([[1, 2],
[1, 2]]))
In [316]: X[_] # this indexes with a tuple of arrays
Out[316]:
array([[2, 0],
[0, 3]])
another way:
In [318]: X.flat[idx]
Out[318]:
array([[2, 0],
[0, 3]])
For example, I have got an array like this:
([ 1, 5, 7, 9, 4, 6, 3, 3, 7, 9, 4, 0, 3, 3, 7, 8, 1, 5 ])
I need to find all duplicated sequences , not values, but sequences of at least two values one by one.
The result should be like this:
of length 2: [1, 5] with indexes (0, 16);
of length 3: [3, 3, 7] with indexes (6, 12); [7, 9, 4] with indexes (2, 8)
The long sequences should be excluded, if they are not duplicated. ([5, 5, 5, 5]) should NOT be taken as [5, 5] on indexes (0, 1, 2)! It's not a duplicate sequence, it's one long sequence.
I can do it with pandas.apply function, but it calculates too slow, swifter did not help me.
And in real life I need to find all of them, with length from 10 up to 100 values one by one on database with 1500 columns with 700 000 values each. So i really do need a vectorized decision.
Is there a vectorized decision for finding all at once? Or at least for finding only 10-values sequences? Or only 4-values sequences? Anything, that will be fully vectorized?
One possible implementation (although not fully vectorized) that finds all sequences of size n that appear more than once is the following:
import numpy as np
def repeated_sequences(arr, n):
Na = arr.size
r_seq = np.arange(n)
n_seqs = arr[np.arange(Na - n + 1)[:, None] + r_seq]
unique_seqs = np.unique(n_seqs, axis=0)
comp = n_seqs == unique_seqs[:, None]
M = np.all(comp, axis=-1)
if M.any():
matches = np.array(
[np.convolve(M[i], np.ones((n), dtype=int)) for i in range(M.shape[0])]
)
repeated_inds = np.count_nonzero(matches, axis=-1) > n
repeated_matches = matches[repeated_inds]
idxs = np.argwhere(repeated_matches > 0)[::n]
grouped_idxs = np.split(
idxs[:, 1], np.unique(idxs[:, 0], return_index=True)[1][1:]
)
else:
return [], []
return unique_seqs[repeated_inds], grouped_idxs
In theory, you could replace
matches = np.array(
[np.convolve(M[i], np.ones((n), dtype=int)) for i in range(M.shape[0])]
)
with
matches = scipy.signal.convolve(
M, np.ones((1, n), dtype=int), mode="full"
).astype(int)
which would make the whole thing "fully vectorized", but my tests showed that this was 3 to 4 times slower than the for-loop. So I'd stick with that. Or simply,
matches = np.apply_along_axis(np.convolve, -1, M, np.ones((n), dtype=int))
which does not have any significant speed-up, since it's basically a hidden loop (see this).
This is based off #Divakar's answer here that dealt with a very similar problem, in which the sequence to look for was provided. I simply made it so that it could follow this procedure for all possible sequences of size n, which are found inside the function with n_seqs = arr[np.arange(Na - n + 1)[:, None] + r_seq]; unique_seqs = np.unique(n_seqs, axis=0).
For example,
>>> a = np.array([1, 5, 7, 9, 4, 6, 3, 3, 7, 9, 4, 0, 3, 3, 7, 8, 1, 5])
>>> repeated_seqs, inds = repeated_sequences(a, n)
>>> for i, seq in enumerate(repeated_seqs[:10]):
...: print(f"{seq} with indexes {inds[i]}")
...:
[3 3 7] with indexes [ 6 12]
[7 9 4] with indexes [2 8]
Disclaimer
The long sequences should be excluded, if they are not duplicated. ([5, 5, 5, 5]) should NOT be taken as [5, 5] on indexes (0, 1, 2)! It's not a duplicate sequence, it's one long sequence.
This is not directly taken into account and the sequence [5, 5] would appear more than once according to this algorithm. You could do something like this, based off #Paul's answer here, but it involves a loop:
import numpy as np
repeated_matches = np.array([[0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]])
idxs = np.argwhere(repeated_matches > 0)
grouped_idxs = np.split(
idxs[:, 1], np.unique(idxs[:, 0], return_index=True)[1][1:]
)
>>> print(grouped_idxs)
[array([ 6, 7, 8, 12, 13, 14], dtype=int64),
array([ 7, 8, 9, 10], dtype=int64)]
# If there are consecutive numbers in grouped_idxs, that means that there is a long
# sequence that should be excluded. So, you'd have to check for consecutive numbers
filtered_idxs = []
for idx in grouped_idxs:
if not all((idx[1:] - idx[:-1]) == 1):
filtered_idxs.append(idx)
>>> print(filtered_idxs)
[array([ 6, 7, 8, 12, 13, 14], dtype=int64)]
Some tests:
>>> n = 3
>>> a = np.array([1, 5, 7, 9, 4, 6, 3, 3, 7, 9, 4, 0, 3, 3, 7, 8, 1, 5])
>>> %timeit repeated_sequences(a, n)
414 µs ± 5.88 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
>>> n = 4
>>> a = np.random.randint(0, 10, (10000,))
>>> %timeit repeated_sequences(a, n)
3.88 s ± 54 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
>>> result, _ = repeated_sequences(a, n)
>>> result.shape
(2637, 4)
This is not the most efficient implementation by far, but it works as a 2D approach. Plus, if there aren't any repeated sequences, it returns empty lists.
EDIT: Full implementation
I vectorized the routine I added in the Disclaimer section as a possible solution to the long sequence problem and ended up with the following:
import numpy as np
# Taken from:
# https://stackoverflow.com/questions/53051560/stacking-numpy-arrays-of-different-length-using-padding
def stack_padding(it):
def resize(row, size):
new = np.array(row)
new.resize(size)
return new
row_length = max(it, key=len).__len__()
mat = np.array([resize(row, row_length) for row in it])
return mat
def repeated_sequences(arr, n):
Na = arr.size
r_seq = np.arange(n)
n_seqs = arr[np.arange(Na - n + 1)[:, None] + r_seq]
unique_seqs = np.unique(n_seqs, axis=0)
comp = n_seqs == unique_seqs[:, None]
M = np.all(comp, axis=-1)
repeated_seqs = []
idxs_repeated_seqs = []
if M.any():
matches = np.apply_along_axis(np.convolve, -1, M, np.ones((n), dtype=int))
repeated_inds = np.count_nonzero(matches, axis=-1) > n
if repeated_inds.any():
repeated_matches = matches[repeated_inds]
idxs = np.argwhere(repeated_matches > 0)
grouped_idxs = np.split(
idxs[:, 1], np.unique(idxs[:, 0], return_index=True)[1][1:]
)
# Additional routine
# Pad this uneven array with zeros so that we can use it normally
grouped_idxs = np.array(grouped_idxs, dtype=object)
padded_idxs = stack_padding(grouped_idxs)
# Find the indices where there are padded zeros
pad_positions = padded_idxs == 0
# Perform the "consecutive-numbers check" (this will take one
# item off the original array, so we have to correct for its shape).
idxs_to_remove= np.pad(
(padded_idxs[:, 1:] - padded_idxs[:, :-1]) == 1,
[(0, 0), (0, 1)],
constant_values=True,
)
pad_positions = np.argwhere(pad_positions)
i = pad_positions[:, 0]
j = pad_positions[:, 1] - 1 # Shift by one (shape correction)
idxs_to_remove[i, j] = True # Masking, since we don't want pad indices
# Obtain a final mask (boolean opposite of indices to remove)
final_mask = ~idxs_to_remove.all(axis=-1)
grouped_idxs = grouped_idxs[final_mask] # Filter the long sequences
repeated_seqs = unique_seqs[repeated_inds][final_mask]
# In order to get the correct indices, we must first limit the
# search to a shape (on axis=1) of the closest multiple of n.
# This will avoid taking more indices than we should to show where
# each repeated sequence begins
to = padded_idxs.shape[1] & (-n)
# Build the final list of indices (that goes from 0 - to with
# a step of n
idxs_repeated_seqs = [
grouped_idxs[i][:to:n] for i in range(grouped_idxs.shape[0])
]
return repeated_seqs, idxs_repeated_seqs
For example,
n = 2
examples = [
# First example is your original example array.
np.array([1, 5, 7, 9, 4, 6, 3, 3, 7, 9, 4, 0, 3, 3, 7, 8, 1, 5]),
# Second example has a long sequence of 5's, and since there aren't
# any [5, 5] anywhere else, it's not taken into account and therefore
# should not come out.
np.array([1, 5, 5, 5, 5, 6, 3, 3, 7, 9, 4, 0, 3, 3, 7, 8, 1, 5]),
# Third example has the same long sequence but since there is a [5, 5]
# later, then it should take it into account and this sequence should
# be found.
np.array([1, 5, 5, 5, 5, 6, 5, 5, 7, 9, 4, 0, 3, 3, 7, 8, 1, 5]),
# Fourth example has a [5, 5] first and later it has a long sequence of
# 5's which are uneven and the previous implementation got confused with
# the indices to show as the starting indices. In this case, it should be
# 1, 13 and 15 for [5, 5].
np.array([1, 5, 5, 9, 4, 6, 3, 3, 7, 9, 4, 0, 3, 5, 5, 5, 5, 5]),
]
for a in examples:
print(f"\nExample: {a}")
repeated_seqs, inds = repeated_sequences(a, n)
for i, seq in enumerate(repeated_seqs):
print(f"\t{seq} with indexes {inds[i]}")
Output (as expected):
Example: [1 5 7 9 4 6 3 3 7 9 4 0 3 3 7 8 1 5]
[1 5] with indexes [0 16]
[3 3] with indexes [6 12]
[3 7] with indexes [7 13]
[7 9] with indexes [2 8]
[9 4] with indexes [3 9]
Example: [1 5 5 5 5 6 3 3 7 9 4 0 3 3 7 8 1 5]
[1 5] with indexes [0 16]
[3 3] with indexes [6 12]
[3 7] with indexes [7 13]
Example: [1 5 5 5 5 6 5 5 7 9 4 0 3 3 7 8 1 5]
[1 5] with indexes [ 0 16]
[5 5] with indexes [1 3 6]
Example: [1 5 5 9 4 6 3 3 7 9 4 0 3 5 5 5 5 5]
[5 5] with indexes [ 1 13 15]
[9 4] with indexes [3 9]
You can test it out yourself with more examples and more cases. Keep in mind this is what I understood from your disclaimer. If you want to count the long sequences as one, even if multiple sequences are in there (for example, [5, 5] appears twice in [5, 5, 5, 5]), this won't work for you and you'd have to come up with something else.
I'd like to select elements from an array along a specific axis given an index array. For example, given the arrays
a = np.arange(30).reshape(5,2,3)
idx = np.array([0,1,1,0,0])
I'd like to select from the second dimension of a according to idx, such that the resulting array is of shape (5,3). Can anyone help me with that?
You could use fancy indexing
a[np.arange(5),idx]
Output:
array([[ 0, 1, 2],
[ 9, 10, 11],
[15, 16, 17],
[18, 19, 20],
[24, 25, 26]])
To make this more verbose this is the same as:
x,y,z = np.arange(a.shape[0]), idx, slice(None)
a[x,y,z]
x and y are being broadcasted to the shape (5,5). z could be used to select any columns in the output.
I think this gives the results you are after - it uses np.take_along_axis, but first you need to reshape your idx array so that it is also a 3d array:
a = np.arange(30).reshape(5, 2, 3)
idx = np.array([0, 1, 1, 0, 0]).reshape(5, 1, 1)
results = np.take_along_axis(a, idx, 1).reshape(5, 3)
Giving:
[[ 0 1 2]
[ 9 10 11]
[15 16 17]
[18 19 20]
[24 25 26]]
I have the following array
arr = np.array([[[1, 2, 3], [4, 5, 6]],
[[7, 8, 9], [10, 11, 12]]])
I want to go through each element and sum on axis 0, so I do:
lst = []
for x in arr:
for y in np.sum(x,axis=0):
lst.append(y)
where now the lst is
[5, 7, 9, 17, 19, 21]
However I want the output to be in the following form:
[[5, 7, 9], [17, 19, 21]]
to then take the mean of its axis 0 namely (5+17)/2 and so on. The final output should look like
[11., 13., 15.]
I wonder how can I do this? Is it possible to write this whole operation in a compact form as list comprehension?
Update: To get the final output I can do:
np.mean(np.reshape(lst, (len(arr),-1)),axis=0)
Yet I am sure there is a Pythonic way of doing this
In [5]: arr = np.array([[[1, 2, 3], [4, 5, 6]],
...: [[7, 8, 9], [10, 11, 12]]])
In [7]: arr
Out[7]:
array([[[ 1, 2, 3],
[ 4, 5, 6]],
[[ 7, 8, 9],
[10, 11, 12]]])
The for iterates on the 1st dimension, as though it was a list of arrays:
In [8]: for x in arr:print(x)
[[1 2 3]
[4 5 6]]
[[ 7 8 9]
[10 11 12]]
list(arr) also makes a list (but it is slower than `arr.tolist()).
One common way of iterating on other dimensions is to use an index:
In [10]: for i in range(2):print(arr[:,i])
[[1 2 3]
[7 8 9]]
[[ 4 5 6]
[10 11 12]]
You could also transpose the array placing the desired axis first.
But you don't need to iterate
In [13]: arr.sum(axis=1)
Out[13]:
array([[ 5, 7, 9],
[17, 19, 21]])
In [14]: arr.sum(axis=1).mean(axis=0)
Out[14]: array([11., 13., 15.])
I want to remap a numpy array according to a dictionary.
Let us assume I have a numpy array with N rows and 3 columns. Now I want to remap the values according to its indices which are written in tuples in a dictionary.
This works fine:
import numpy as np
a = np.arange(6).reshape(2,3)
b = np.zeros(6).reshape(2,3)
print a
print A
dictt = { (0,0):(0,2), (0,1):(0,1), (0,2):(0,0), (1,0):(1,2), (1,1):(1,1), (1,2):(1,0) }
for key in dictt:
b[key] = a[dictt[key]]
print b
a = [[0 1 2]
[3 4 5]]
b = [[ 2. 1. 0.]
[ 5. 4. 3.]]
Let us assume I have N rows, where N is an even number. Now I want to apply the same mapping (which are valid for those 2 rows in the upper example) to all the other rows.
Hence I want to have an array from:
[[ 0 1 2]
[ 3 4 5]
[ 6 7 8]
[ 9 10 11]]
to:
b = [[ 2. 1. 0.]
[ 5. 4. 3.]]
[[ 8. 7. 6.]
[ 11. 10. 9.]]
Any ideas? I would like to do it fast since these are 192000 entries in each array which should be remapped.
For simplicity I would just use [::-1].
a = [[0, 1, 2], [3, 4, 5], [6, 7, 8]]
b = [item[::-1] for item in a]
>>> b
[[2, 1, 0], [5, 4, 3], [8, 7, 6]]