CODEWITHSUNDEEP
pythonpython-3.xnumpyvectorization
matte
matte

Reputation: 97

Hybridizing Numpy vectorization and plain Python

I have a 2-dimensional ndarray, the rows of which shall be scanned to check whether any one is equal to any other.

My first try actually works, but I feel it is not the optimal way. It takes time once the number of rows in the matrix approaches 1000.

My code is the following. X is the aforementioned array, Y also is a 2-dimensional ndarray.

for i in range(X.shape[0]-1):
    for j in range(i+1,X.shape[0]):
        if (np.all( (X[i,:] == X[j,:] ), axis = 0 )):
            Y[j,:] = Y[i,:]
        #endif
    #enddo
#enddo

I know that the nested loop is time consuming and should be avoided, but I could not find an alternative. List comprehension seems to me not suitable in that there is no need to save items.

The fact that the core of the procedure is the assignment operation Y[j,:] = Y[i,:], which is index-dependent, would lead me to exclude a list comprehension-like solution.

Question then is: is there a more efficient way to code such a search exploiting numpy vectorization?

Upvotes: 2

Views: 105

Answers (3)

Subhaneil Lahiri
Subhaneil Lahiri

Reputation: 458

I'm on my phone so I can't test this, but I think it will work

mask = np.all(X[:, None] == X[None], axis=-1)
ind1, ind2 = np.nonzero(mask)
ind1, ind2 = ind1[ind1 < ind2], ind2[ind1 < ind2]
Y[ind2] = Y[ind1]

Upvotes: 0

Divakar
Divakar

Reputation: 221534

Approach #1

We could leverage row-views to get the pairwise matches. Then, run the loop and assign those in Y. The idea is to minimize the work once we start running loop. Considering that there could be more than one index matching with other indices, a purely vectorized method would be hard to propose. The implementation would look something like this -

# https://stackoverflow.com/a/44999009/ @Divakar
def view1D(a): # a is array
    a = np.ascontiguousarray(a)
    void_dt = np.dtype((np.void, a.dtype.itemsize * a.shape[1]))
    return a.view(void_dt).ravel()

# Get 1D view
a1D = view1D(a)

# Perform broadcasting to get outer equality match
mask = a1D[:,None]==a1D

# Get indices of pairwise matches
n = len(mask)
mask[np.tri(n, dtype=bool)] = 0
idx = np.argwhere(mask)

# Run loop to assign equal rows in Y
for (i,j) in zip(idx[:,0],idx[:,1]):
    Y[j] = Y[i]

Alternative #1 : Use mask to directly assign

So, with mask, directly assign the rows in Y, like so -

for i,m in enumerate(mask):
    if m.any():
        Y[m] = Y[i]

This would be helpful if there are many matches.

Alternative #2 : Use reduced mask

If there are more than one row common between two rows, we might want to reduce those to have all of those linked to the first occurring ones. Hence, we can generate a reduced-mask and use that instead of the earlier mask -

mask0 = np.zeros_like(mask)
mask0[mask.argmax(0), np.arange(len(mask))] = 1
np.fill_diagonal(mask0,0)

Then, use mask0 instead of mask and assign.

Approach #2

Another method would be starting off with the 1D view method and using sorting-based method to setup the pairwise matching indices, like so -

sidx = a1D.argsort() # a1D from earlier approach
b = a1D[sidx]
m0 = b[:-1] == b[1:]
m1 = np.r_[False,m0,False]
idx = np.flatnonzero(m1[:-1]!=m1[1:]).reshape(-1,2)
for (i,j) in idx:
    row0,row1 = sidx[i],sidx[i+1:j+1]
    Y[row1] = Y[row0]

Upvotes: 3

Carlos Pedro Soberats
Carlos Pedro Soberats

Reputation: 71

See the following example: As an illustration, consider a 1-dimensional vector of True and False for which you want to count the number of “False to True” transitions in the sequence:

np.random.seed(444)
x = np.random.choice([False, True], size=100000)

With a Python for-loop, one way to do this would be to evaluate, in pairs, the truth value of each element in the sequence along with the element that comes right after it:

def count_transitions(x) -> int:
  count = 0
  for i, j in zip(x[:-1], x[1:]):
     if j and not i:
        count += 1
  return count

count_transitions(x)

In vectorized form, there’s no explicit for-loop or direct reference to the individual elements:

np.count_nonzero(x[:-1] < x[1:])

How do these two equivalent functions compare in terms of performance? In this particular case, the vectorized NumPy call wins out by a factor of about 70 times

https://realpython.com/numpy-array-programming/

Upvotes: 1

Related Questions