I have two methods for calculating the unit vector of an array, both of which handle sparse arrays. One of them is very much a 'manual' computation, whereas the other is more 'formal' (from the gensim.matutils source code).
This is the manual method:
def manual_unitvec(vec):
vec = vec.tocsr()
if sparse.issparse(vec):
vec_sum_of_squares = vec.multiply(vec)
unit = 1. / np.sqrt(vec_sum_of_squares.sum())
return vec.multiply(unit)
elif not sparse.issparse(vec):
sum_vec_squared = np.sum(vec ** 2)
vec /= np.sqrt(sum_vec_squared)
return vec
This is the modified gensim method, where the method explicitly computing the unit vector is unitvec:
import numpy as np
from scipy import sparse
from gensim.matutils import ret_normalized_vec, blas
import scipy.sparse
blas_nrm2 = blas('nrm2', np.array([], dtype=float))
blas_scal = blas('scal', np.array([], dtype=float))
def unitvec(vec, norm='l2'):
"""Scale a vector to unit length.
Parameters
----------
vec : {numpy.ndarray, scipy.sparse, list of (int, float)}
Input vector in any format
norm : {'l1', 'l2'}, optional
Normalization that will be used.
Returns
-------
{numpy.ndarray, scipy.sparse, list of (int, float)}
Normalized vector in same format as `vec`.
Notes
-----
Zero-vector will be unchanged.
"""
if norm not in ('l1', 'l2'):
raise ValueError("'%s' is not a supported norm. Currently supported norms are 'l1' and 'l2'." % norm)
if scipy.sparse.issparse(vec):
print("INSIDE SPARSE HANDLING")
vec = vec.tocsr()
if norm == 'l1':
veclen = np.sum(np.abs(vec.data))
if norm == 'l2':
veclen = np.sqrt(np.sum(vec.data ** 2))
if veclen > 0.0:
if np.issubdtype(vec.dtype, np.int) == True:
vec = vec.astype(np.float)
return vec / veclen
else:
vec /= veclen
return vec.astype(vec.dtype)
else:
return vec
if isinstance(vec, np.ndarray):
print("INSIDE NORMAL VEC HANDLING")
vec = np.asarray(vec, dtype=vec.dtype)
if norm == 'l1':
veclen = np.sum(np.abs(vec))
if norm == 'l2':
veclen = blas_nrm2(vec)
if veclen > 0.0:
if np.issubdtype(vec.dtype, np.int) == True:
vec = vec.astype(np.float)
return blas_scal(1.0 / veclen, vec).astype(vec.dtype)
else:
return blas_scal(1.0 / veclen, vec).astype(vec.dtype)
else:
return vec
try:
first = next(iter(vec)) # is there at least one element?
except StopIteration:
return vec
if isinstance(first, (tuple, list)) and len(first) == 2: # gensim sparse format
print("INSIDE GENSIM SPARSE FORMAT HANDLING")
if norm == 'l1':
length = float(sum(abs(val) for _, val in vec))
if norm == 'l2':
length = 1.0 * math.sqrt(sum(val ** 2 for _, val in vec))
assert length > 0.0, "sparse documents must not contain any explicit zero entries"
return ret_normalized_vec(vec, length)
else:
raise ValueError("unknown input type")
When running tests, I want to check that the output from each of these methods is the same. Below is a snippet of example code:
vec = sparse.csr_matrix([[1, 2, 3], [4, 5, 6], [7, 8, 9]]).astype(np.float32)
output1 = manual_unitvec(vec)
output2 = unitvec(vec)
print(output1)
print(' ')
print(output2)
print(np.array_equal(output1, output2))
print(type(output1) == type(output2))
So what I want to check is assertTrue(output1, output2). You can't do this because the truth value of arrays is ambiguous, so I use assertTrue(np.array_equal(output1, output2)).
Now the issue is that array_equal does not view output1 and output2 as being the same, even though I can see from printing them out that they are identical.
Running all of the code above gives the following output:
MacBook-Air:matutils.unitvec Olly$ python try.py
INSIDE SPARSE HANDLING
try.py:80: FutureWarning: Conversion of the second argument of issubdtype from `int` to `np.signedinteger` is deprecated. In future, it will be treated as `np.int64 == np.dtype(int).type`.
if np.issubdtype(vec.dtype, np.int) == True:
(0, 0) 0.059234887
(0, 1) 0.118469775
(0, 2) 0.17770466
(1, 0) 0.23693955
(1, 1) 0.29617444
(1, 2) 0.35540932
(2, 0) 0.4146442
(2, 1) 0.4738791
(2, 2) 0.53311396
(0, 0) 0.059234887
(0, 1) 0.118469775
(0, 2) 0.17770466
(1, 0) 0.23693955
(1, 1) 0.29617444
(1, 2) 0.35540932
(2, 0) 0.4146442
(2, 1) 0.4738791
(2, 2) 0.53311396
/Users/Olly/anaconda2/lib/python2.7/site-packages/scipy/sparse/compressed.py:226: SparseEfficiencyWarning: Comparing sparse matrices using == is inefficient, try using != instead.
" != instead.", SparseEfficiencyWarning)
False
True
I thought that the issue might have come from the sparse array types, but as you can see, they are equal. You can also visually see that the elements are exactly the same.
So why is array_equal returning false? How can I change it?
In your first function you do:
vec = vec.tocsr()
if sparse.issparse(vec):
I don't think the issparse test does anything for you. If the input argument is a sparse matrix, it has a tocsr method, and the result is sparse matrix. If vec is a ndarray is does not have a tocsr method, and the first line will throw an error.
In the rest of that function, a sparse matrix has a multiply method, and a sum method. The result of sum is dense, so np.sqrt works fine on it. Actually np.sqrt(M) also works on sparse matrix because M.sqrt exists.
In the second function you work with the data attribute, which is a 1d ndarray.
np.sum(np.abs(vec.data))
That's fine. But notice that M.__abs__ for a sparse matrix is
self._with_data(abs(self._deduped_data()))
In a slightly more round about way, funtions/methods like abs, sqrt work with the .data attribute as well. Only they return a new sparse matrix.
As for the testing look at np.array_equal
return bool(asarray(a1 == a2).all())
If I try to use it on output1 (I won't try your gensim solution
In [106]: np.array_equal(output1, output1)
/usr/local/lib/python3.5/dist-packages/scipy/sparse/compressed.py:226: SparseEfficiencyWarning: Comparing sparse matrices using == is inefficient, try using != instead.
" != instead.", SparseEfficiencyWarning)
Out[106]: False
It doesn't like taking == on sparse matrices. Usually these are large with many 0s. That means the result will be True for all those, and hence no longer sparse.
Your output1 is a sparse matrix, but, at least for these inputs, not sparse:
In [107]: output1.A
Out[107]:
array([[0.05923489, 0.11846977, 0.17770466],
[0.23693955, 0.29617444, 0.35540932],
[0.4146442 , 0.4738791 , 0.53311396]], dtype=float32)
But even if you get around the sparsity bit, np.array_equal(output1.A, output2.A) could fail due to a comparison of floats
allclose on the dense versions is probably the simplest test:
In [113]: np.allclose(output1.A, output1.A)
Out[113]: True
You could also compare the data (assuming the sparsity is the same):
In [114]: np.allclose(output1.data, output1.data)
Out[114]: True
A fuller sparse test would need to check shape, nnz, and indices attributes.
Actually I'm not sure where np.array_equal is failing. Notice that it starts with a1=asarray(a1), which produces a 0d object dtype array. This is one of those numpy functions that insists on treating its inputs as arrays. It is not sparse aware.