Question

I have been trying to implement some modification to speed up this pseudo code:

>>> A=np.array([1,1,1,2,2,2,3,3,3])
>>> B=np.array([np.power(A,n) for n in [3,4,5]])
>>> B
array([[  1,   1,   1,   8,   8,   8,  27,  27,  27],
       [  1,   1,   1,  16,  16,  16,  81,  81,  81],
       [  1,   1,   1,  32,  32,  32, 243, 243, 243]])

Where elements of A are often repeated 10-20 times and the shape of B needs to be retained because it is multiplied by another array of the same shape later.

My first idea was to use the following code:

uA=np.unique(A)
uB=np.array([np.power(uA,n) for n in [3,4,5]])
B=[]
for num in range(uB.shape[0]):
    Temp=np.copy(A)
    for k,v in zip(uA,uB[num]): Temp[A==k] = v
    B.append(Temp)
B=np.array(B)
### Also any better way to create the numpy array B?

This seems fairly terrible and there is likely a better way. Any idea on how to speed this up would be much appreciated.

Thank you for your time.

Here is an update. I realized that my function was poorly coded. A thank you to everyone for the suggestions. I will try to rephrase my questions better in the future so that they show everything required.

Normal='''
import numpy as np
import scipy
def func(value,n):
    if n==0: return 1
    else: return np.power(value,n)/scipy.factorial(n,exact=0)+func(value,n-1)
A=np.random.randint(10,size=250)
A=np.unique(A)
B=np.array([func(A,n) for n in [6,8,10]])
'''

Me='''
import numpy as np
import scipy
def func(value,n):
    if n==0: return 1
    else: return np.power(value,n)/scipy.factorial(n,exact=0)+func(value,n-1)
A=np.random.randint(10,size=250)
uA=np.unique(A)
uB=np.array([func(A,n) for n in [6,8,10]])
B=[]
for num in range(uB.shape[0]):
    Temp=np.copy(A)
    for k,v in zip(uA,uB[num]): Temp[A==k] = v
    B.append(Temp)
B=np.array(B)
'''


Alex='''
import numpy as np
import scipy
A=np.random.randint(10,size=250)
power=np.arange(11)
fact=scipy.factorial(np.arange(11),exact=0).reshape(-1,1)
power=np.power(A,np.arange(11).reshape(-1,1))
value=power/fact
six=np.sum(value[:6],axis=0)
eight=six+np.sum(value[6:8],axis=0)
ten=eight+np.sum(value[8:],axis=0)
B=np.vstack((six,eight,ten))
'''
Alex='''
import numpy as np
import scipy
A=np.random.randint(10,size=250)
power=np.arange(11)
fact=scipy.factorial(np.arange(11),exact=0).reshape(-1,1)
power=np.power(A,np.arange(11).reshape(-1,1))
value=power/fact
six=np.sum(value[:6],axis=0)
eight=six+np.sum(value[6:8],axis=0)
ten=eight+np.sum(value[8:],axis=0)
B=np.vstack((six,eight,ten))
'''

Alex2='''
import numpy as np
import scipy
def find_count(the_list):
    count = list(the_list).count
    result = [count(item) for item in set(the_list)]
    return result
A=np.random.randint(10,size=250)
A_unique=np.unique(A)
A_counts = np.array(find_count(A_unique))
fact=scipy.factorial(np.arange(11),exact=0).reshape(-1,1)
power=np.power(A_unique,np.arange(11).reshape(-1,1))
value=power/fact
six=np.sum(value[:6],axis=0)
eight=six+np.sum(value[6:8],axis=0)
ten=eight+np.sum(value[8:],axis=0)
B_nodup=np.vstack((six,eight,ten))
B_list = [ np.transpose( np.tile( B_nodup[:,i], (A_counts[i], 1) ) ) for i in range(A_unique.shape[0]) ]
B = np.hstack( B_list )
'''


print timeit.timeit(Normal, number=10000)
print timeit.timeit(Me, number=10000)
print timeit.timeit(Alex, number=10000)
print timeit.timeit(Alex2, number=10000)

Normal: 10.7544178963
Me:     23.2039361
Alex:    4.85648703575
Alex2:   4.18024992943

Answer 1

Use a combination of numpy.tile() and numpy.hstack(), as follows:

A = np.array([1,2,3])
A_counts = np.array([3,3,3])
A_powers = np.array([[3],[4],[5]])
B_nodup = np.power(A, A_powers)
B_list = [ np.transpose( np.tile( B_nodup[:,i], (A_counts[i], 1) ) ) for i in range(A.shape[0]) ]
B = np.hstack( B_list )

The transpose and stack may be reversed, this may be faster:

B_list = [ np.tile( B_nodup[:,i], (A_counts[i], 1) ) for i in range(A.shape[0]) ]
B = np.transpose( np.vstack( B_list ) )

This is likely only worth doing if the function you are calculating is quite expensive, or it is duplicated many, many times (more than 10); doing a tile and stack to prevent calculating the power function an extra 10 times is likely not worth it. Please benchmark and let us know.

EDIT: Or, you could just use broadcasting to get rid of the list comprehension:

>>> A=np.array([1,1,1,2,2,2,3,3,3])
>>> B = np.power(A,[[3],[4],[5]])
>>> B
array([[  1,   1,   1,   8,   8,   8,  27,  27,  27],
       [  1,   1,   1,  16,  16,  16,  81,  81,  81],
       [  1,   1,   1,  32,  32,  32, 243, 243, 243]])

This is probably pretty fast, but doesn’t actually do what you asked.