CODEWITHSUNDEEP
pythonpytableseuclidean-distancenumexpr
mrgloom
mrgloom

Reputation: 21662

euclidean norm using numexpr

I need to rewrite this code using numexpr, it's calculating euclidean norm matrix of matrix data [rows x cols] and vector [1 x cols].

d = ((data-vec)**2).sum(axis=1)

How can it be done? Also maybe there is another even faster method?

The problem that I use hdf5, and data matrix readed from it. For example this code gives error: objects are not aligned.

#naive numpy solution, can be parallel?
def test_bruteforce_knn():
    h5f = tables.open_file(fileName)

    t0= time.time()
    d = np.empty((rows*batches,))
    for i in range(batches):
        d[i*rows:(i+1)*rows] = ((h5f.root.carray[i*rows:(i+1)*rows]-vec)**2).sum(axis=1)
    print (time.time()-t0)
    ndx = d.argsort()
    print ndx[:k]

    h5f.close()

#using some tricks (don't work error: objects are not aligned )
def test_bruteforce_knn():
    h5f = tables.open_file(fileName)

    t0= time.time()
    d = np.empty((rows*batches,))
    for i in range(batches):
        d[i*rows:(i+1)*rows] = (np.einsum('ij,ij->i', h5f.root.carray[i*rows:(i+1)*rows],
        h5f.root.carray[i*rows:(i+1)*rows]) 
        + np.dot(vec, vec)
        -2 * np.dot(h5f.root.carray[i*rows:(i+1)*rows], vec))
    print (time.time()-t0)
    ndx = d.argsort()
    print ndx[:k]

    h5f.close()

Using numexpr: seems numexpr don't understand h5f.root.carray[i*rows:(i+1)*rows] it must be reassigned?

import numexpr as ne

def test_bruteforce_knn():
    h5f = tables.open_file(fileName)

    t0= time.time()
    d = np.empty((rows*batches,))
    for i in range(batches):
        ne.evaluate("sum((h5f.root.carray[i*rows:(i+1)*rows] - vec) ** 2, axis=1)")
    print (time.time()-t0)
    ndx = d.argsort()
    print ndx[:k]

    h5f.close()

Upvotes: 1

Views: 843

Answers (1)

Fred Foo
Fred Foo

Reputation: 363787

There's a potentially fast way (for very large arrays) using just NumPy, which is used in scikit-learn:

def squared_row_norms(X):
    # From http://stackoverflow.com/q/19094441/166749
    return np.einsum('ij,ij->i', X, X)

def squared_euclidean_distances(data, vec):
    data2 = squared_row_norms(data)
    vec2 = squared_row_norms(vec)
    d = np.dot(data, vec.T).ravel()
    d *= -2
    d += data2
    d += vec2
    return d

This is based on the fact that (x - y)² = x² + y² - 2xy, even for vectors.

Test:

>>> data = np.random.randn(10, 40)
>>> vec = np.random.randn(1, 40)
>>> ((data - vec) ** 2).sum(axis=1)
array([  96.75712686,   69.45894306,  100.71998244,   80.97797154,
         84.8832107 ,   82.28910021,   67.48309433,   81.94813371,
         64.68162331,   77.43265692])
>>> squared_euclidean_distances(data, vec)
array([  96.75712686,   69.45894306,  100.71998244,   80.97797154,
         84.8832107 ,   82.28910021,   67.48309433,   81.94813371,
         64.68162331,   77.43265692])
>>> from sklearn.metrics.pairwise import euclidean_distances
>>> euclidean_distances(data, vec, squared=True).ravel()
array([  96.75712686,   69.45894306,  100.71998244,   80.97797154,
         84.8832107 ,   82.28910021,   67.48309433,   81.94813371,
         64.68162331,   77.43265692])

Profile:

>>> data = np.random.randn(1000, 40)
>>> vec = np.random.randn(1, 40)
>>> %timeit ((data - vec)**2).sum(axis=1)
10000 loops, best of 3: 114 us per loop
>>> %timeit squared_euclidean_distances(data, vec)
10000 loops, best of 3: 52.5 us per loop

Using numexpr is also possible, but it doesn't seem to give any speedup for 1000 points (and at 10000, it isn't much better):

>>> %timeit ne.evaluate("sum((data - vec) ** 2, axis=1)")
10000 loops, best of 3: 142 us per loop

Upvotes: 4

Related Questions