Vectorizing feature hashing in python

Question

I'm wondering if anyone knows how to vectorize feature hashing in Python. For example, this is my code:

    import numpy as np
    hashlen = 5
    x = np.array([4, 7, 4, 2, 6, 8, 0, 6, 3, 1])
    h = np.array([0, 3, 1, 2, 4, 2, 1, 0, 3, 1])

In feature hashing, h represents the indices of the new vector I am hashing x to, i.e the index 0 of the hashed vector should have 4 and 6 summed up, index 1 should have 4, 0 and 1 summed up, etc. The resulting hashed vector should be:

    w = np.array([ 10, 5, 10, 10, 6])

One way of doing this is of course by looping through the hash indices, i.e:

    for itr in range(hashlen):
        w[itr] = np.sum(x[np.where(h==itr)])

For large vectors, the complexity is a function of hashlen (the length of the hashed vector). It could take too long, especially with a np.where() in it.

I want to do something like:

    w = np.zeros(hashlen)
    w[h]+= x

However, the result of this is the same as doing

    w = np.zeros(hashlen)
    w[h] = x

Can anyone let me know if I'm missing something here? Or if there's an 'easy' way of doing the feature hashing that doesn't involve too many computations?

Answer 1

You can use bincount with weights to do what you are asking:

>>> np.bincount(h,weights=x)
array([ 10.,   5.,  10.,  10.,   6.])

For matrices:

>>> import numpy as np
>>> a=np.random.randint(0,5,(50,50))
>>> rand=np.random.rand(5)
>>> rand
array([ 0.10899745,  0.35296303,  0.21127571,  0.56433924,  0.27895281])
>>> b=np.take(rand,a)

#Unfortunately you cannot do it like this:
>>> np.bincount(a,weights=b)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: object too deep for desired array

#There we go:
>>> np.bincount(a.flat,weights=b.flat)
array([  55.04371257,  172.59892108,   96.34172236,  297.40677707,
        145.89232039])

This used fancy indexing to see what was happening:

>>> np.bincount(a.flat)
array([505, 489, 456, 527, 523])
>>> np.bincount(a.flat)*rand
array([  55.04371257,  172.59892108,   96.34172236,  297.40677707,
        145.89232039])