Cosine Similarity On Large Sparse Matrix With Numpy

The code below causes my system to run out of memory before it completes. Can you suggest a more efficient means of computing the cosine similarity on a large matrix, such as the

Solution 1:

Same problem here. I've got a big, non-sparse matrix. It fits in memory just fine, but cosine_similarity crashes for whatever unknown reason, probably because they copy the matrix one time too many somewhere. So I made it compare small batches of rows "on the left" instead of the entire matrix:

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

defcosine_similarity_n_space(m1, m2, batch_size=100):
    assert m1.shape[1] == m2.shape[1]
    ret = np.ndarray((m1.shape[0], m2.shape[0]))
    for row_i inrange(0, int(m1.shape[0] / batch_size) + 1):
        start = row_i * batch_size
        end = min([(row_i + 1) * batch_size, m1.shape[0]])
        if end <= start:
            break# cause I'm too lazy to elegantly handle edge cases
        rows = m1[start: end]
        sim = cosine_similarity(rows, m2) # rows is O(1) size
        ret[start: end] = sim
    return ret

No crashes for me; YMMV. Try different batch sizes to make it faster. I used to only compare 1 row at a time, and it took about 30X as long on my machine.

Stupid yet effective sanity check:

import random
while True:
    m = np.random.rand(random.randint(1, 100), random.randint(1, 100))
    n = np.random.rand(random.randint(1, 100), m.shape[1])
    assert np.allclose(cosine_similarity(m, n), cosine_similarity_n_space(m, n))

Solution 2:

I would run it in chunks like this

from sklearn.metrics.pairwise import cosine_similarity

# Change chunk_size to control resource consumption and speed# Higher chunk_size means more memory/RAM needed but also faster 
chunk_size = 500 
matrix_len = your_matrix.shape[0] # Not sparse numpy.ndarraydefsimilarity_cosine_by_chunk(start, end):
    ifend > matrix_len:end = matrix_len
    return cosine_similarity(X=your_matrix[start:end], Y=your_matrix) # scikit-learn functionfor chunk_start in xrange(0, matrix_len, chunk_size):
    cosine_similarity_chunk = similarity_cosine_by_chunk(chunk_start, chunk_start+chunk_size)
    # Handle cosine_similarity_chunk  ( Write it to file_timestamp and close the file )# Do not open the same file again or you may end up with out of memory after few chunks 

Solution 3:

You're running out of memory because you're trying to store a 65000x65000 matrix. Note that the matrix you're constructing is not sparse at all. np.random.rand generates a random number between 0 and 1. So there aren't enough zeros for csr_matrix to actually compress your data. In fact, there are almost surely no zeros at all.

If you look closely at your MemoryError traceback, you can see that cosine_similarity tries to use the sparse dot product if possible:

MemoryError                  Traceback (most recent calllast)
    887         Y_normalized =normalize(Y, copy=True)
    888--> 889     K = safe_sparse_dot(X_normalized, Y_normalized.T, dense_output=dense_output)890891return K

So the problem isn't with cosine_similarity, it's with your matrix. Try initializing an actual sparse matrix (with 1% sparsity, for example) like this:

>>>a = np.zeros((65000, 10))>>>i = np.random.rand(a.size)>>>a.flat[i < 0.01] = 1# Select 1% of indices and set to 1>>>a = sparse.csr_matrix(a)

Then, on a machine with 32GB RAM (8GB RAM was not enough for me), the following runs with no memory error:

>>> b = cosine_similarity(a)
>>> b
array([[ 0.,  0.,  0., ...,  0.,  0.,  0.],
       [ 0.,  0.,  0., ...,  0.,  0.,  0.],
       [ 0.,  0.,  0., ...,  0.,  0.,  0.],
       ..., 
       [ 0.,  0.,  0., ...,  1.,  0.,  0.],
       [ 0.,  0.,  0., ...,  0.,  0.,  0.],
       [ 0.,  0.,  0., ...,  0.,  0.,  0.]])