I have an array containing arrays of coordinates like this:
a = [[0,0,300,400],[1,1,15,59],[5,5,300,400]]
Now I want to get the overlap ratio of each rectangle to the other rectangles:
def bool_rect_intersect(A, B):
return not (B[0]>A[2] or B[2]<A[0] or B[3]<A[1] or B[1]>A[3])
def get_overlap_ratio(A, B):
in_ = bool_rect_intersect(A, B)
if not in_:
return 0
else:
left = max(A[0], B[0]);
top = max(A[1], B[1]);
right = min(A[2], B[2]);
bottom = min(A[3], B[3]);
intersection = [left, top, right, bottom];
surface_intersection = (intersection[2]-intersection[0])*(intersection[3]-intersection[1]);
surface_A = (A[2]- A[0])*(A[3]-A[1]) + 0.0;
return surface_intersection / surface_A
Now i'm looking for the fastest way to compute the grid of overlaps for arrays of size 2000+.
If I loop over it it takes more than a minute. I tried np.vectorize, but i don't think this is applicable in a multidimensional array
Approach #1 : Here's a vectorized approach -
def pairwise_overlaps(a):
r,c = np.triu_indices(a.shape[0],1)
lt = np.maximum(a[r,:2], a[c,:2])
tb = np.minimum(a[r,2:], a[c,2:])
si_vectorized = (tb[:,0] - lt[:,0]) * (tb[:,1] - lt[:,1])
slicedA_comps = ((a[:,2]- a[:,0])*(a[:,3]-a[:,1]) + 0.0)
sA_vectorized = np.take(slicedA_comps, r)
return si_vectorized/sA_vectorized
Sample run -
In [48]: a
Out[48]:
array([[ 0, 0, 300, 400],
[ 1, 1, 15, 59],
[ 5, 5, 300, 400]])
In [49]: print get_overlap_ratio(a[0], a[1]) # Looping thru pairs
...: print get_overlap_ratio(a[0], a[2])
...: print get_overlap_ratio(a[1], a[2])
...:
0.00676666666667
0.971041666667
0.665024630542
In [50]: pairwise_overlaps(a) # Proposed app to get all those in one-go
Out[50]: array([ 0.00676667, 0.97104167, 0.66502463])
Approach #2 : Upon close inspection, we will see that in the previous approach, the indexing with the r's and c's would be performance killers as they will make copies. We can improve on this, by performing computations for each element in a column against each of other elements in the same column, as listed in the implementation below -
def pairwise_overlaps_v2(a):
rl = np.minimum(a[:,2], a[:,None,2]) - np.maximum(a[:,0], a[:,None,0])
bt = np.minimum(a[:,3], a[:,None,3]) - np.maximum(a[:,1], a[:,None,1])
si_vectorized2D = rl*bt
slicedA_comps = ((a[:,2]- a[:,0])*(a[:,3]-a[:,1]) + 0.0)
overlaps2D = si_vectorized2D/slicedA_comps[:,None]
r = np.arange(a.shape[0])
tril_mask = r[:,None] < r
return overlaps2D[tril_mask]
Runtime test
In [238]: n = 1000
In [239]: a = np.hstack((np.random.randint(0,100,(n,2)), \
np.random.randint(300,500,(n,2))))
In [240]: np.allclose(pairwise_overlaps(a), pairwise_overlaps_v2(a))
Out[240]: True
In [241]: %timeit pairwise_overlaps(a)
10 loops, best of 3: 35.2 ms per loop
In [242]: %timeit pairwise_overlaps_v2(a)
100 loops, best of 3: 16 ms per loop
Let's add in the original approach as loop-comprehension -
In [244]: r,c = np.triu_indices(a.shape[0],1)
In [245]: %timeit [get_overlap_ratio(a[r[i]], a[c[i]]) for i in range(len(r))]
1 loops, best of 3: 2.85 s per loop
Around 180x speedup there with the second approach over the original one!
I'm trying to compute a simple dot product but leave nonzero values from the original matrix unchanged. A toy example:
import numpy as np
A = np.array([[2, 1, 1, 2],
[0, 2, 1, 0],
[1, 0, 1, 1],
[2, 2, 1, 0]])
B = np.array([[ 0.54331039, 0.41018682, 0.1582158 , 0.3486124 ],
[ 0.68804647, 0.29520239, 0.40654206, 0.20473451],
[ 0.69857579, 0.38958572, 0.30361365, 0.32256483],
[ 0.46195299, 0.79863505, 0.22431876, 0.59054473]])
Desired outcome:
C = np.array([[ 2. , 1. , 1. , 2. ],
[ 2.07466874, 2. , 1. , 0.73203386],
[ 1. , 1.5984076 , 1. , 1. ],
[ 2. , 2. , 1. , 1.42925865]])
The actual matrices in question, however, are sparse and look more like this:
A = sparse.rand(250000, 1700, density=0.001, format='csr')
B = sparse.rand(1700, 1700, density=0.02, format='csr')
One simple way go would be just setting the values using mask index, like that:
mask = A != 0
C = A.dot(B)
C[mask] = A[mask]
However, my original arrays are sparse and quite large, so changing them via index assignment is painfully slow. Conversion to lil matrix helps a bit, but again, conversion itself takes a lot of time.
The other obvious approach, I guess, would be just resort to iteration and skip masked values, but I'd like not to throw away the benefits of numpy/scipy-optimized array multiplication.
Some clarifications: I'm actually interested in some kind of special case, where B is always square, and therefore, A and C are of the same shape. So if there's a solution that doesn't work on arbitrary arrays but fits in my case, that's fine.
UPDATE: Some attempts:
from scipy import sparse
import numpy as np
def naive(A, B):
mask = A != 0
out = A.dot(B).tolil()
out[mask] = A[mask]
return out.tocsr()
def proposed(A, B):
Az = A == 0
R, C = np.where(Az)
out = A.copy()
out[Az] = np.einsum('ij,ji->i', A[R], B[:, C])
return out
%timeit naive(A, B)
1 loops, best of 3: 4.04 s per loop
%timeit proposed(A, B)
/usr/local/lib/python2.7/dist-packages/scipy/sparse/compressed.py:215: SparseEfficiencyWarning: Comparing a sparse matrix with 0 using == is inefficient, try using != instead.
/usr/local/lib/python2.7/dist-packages/scipy/sparse/coo.pyc in __init__(self, arg1, shape, dtype, copy)
173 self.shape = M.shape
174
--> 175 self.row, self.col = M.nonzero()
176 self.data = M[self.row, self.col]
177 self.has_canonical_format = True
MemoryError:
ANOTHER UPDATE:
Couldn't make anything more or less useful out of Cython, at least without going too far away from Python. The idea was to leave the dot product to scipy and just try to set those original values as fast as possible, something like this:
cimport cython
#cython.cdivision(True)
#cython.boundscheck(False)
#cython.wraparound(False)
cpdef coo_replace(int [:] row1, int [:] col1, float [:] data1, int[:] row2, int[:] col2, float[:] data2):
cdef int N = row1.shape[0]
cdef int M = row2.shape[0]
cdef int i, j
cdef dict d = {}
for i in range(M):
d[(row2[i], col2[i])] = data2[i]
for j in range(N):
if (row1[j], col1[j]) in d:
data1[j] = d[(row1[j], col1[j])]
This was a bit better then my pre-first "naive" implementation (using .tolil()), but following hpaulj's approach, lil can be thrown out. Maybe replacing python dict with something like std::map would help.
A possibly cleaner and faster version of your naive code:
In [57]: r,c=A.nonzero() # this uses A.tocoo()
In [58]: C=A*B
In [59]: Cl=C.tolil()
In [60]: Cl[r,c]=A.tolil()[r,c]
In [61]: Cl.tocsr()
C[r,c]=A[r,c] gives an efficiency warning, but I think that's aimed more a people do that kind of assignment in loop.
In [63]: %%timeit C=A*B
...: C[r,c]=A[r,c]
...
The slowest run took 7.32 times longer than the fastest....
1000 loops, best of 3: 334 µs per loop
In [64]: %%timeit C=A*B
...: Cl=C.tolil()
...: Cl[r,c]=A.tolil()[r,c]
...: Cl.tocsr()
...:
100 loops, best of 3: 2.83 ms per loop
My A is small, only (250,100), but it looks like the round trip to lil isn't a time saver, despite the warning.
Masking with A==0 is bound to give problems when A is sparse
In [66]: Az=A==0
....SparseEfficiencyWarning...
In [67]: r1,c1=Az.nonzero()
Compared to the nonzero r for A, this r1 is much larger - the row index of all zeros in the sparse matrix; everything but the 25 nonzeros.
In [70]: r.shape
Out[70]: (25,)
In [71]: r1.shape
Out[71]: (24975,)
If I index A with that r1 I get a much larger array. In effect I am repeating each row by the number of zeros in it
In [72]: A[r1,:]
Out[72]:
<24975x100 sparse matrix of type '<class 'numpy.float64'>'
with 2473 stored elements in Compressed Sparse Row format>
In [73]: A
Out[73]:
<250x100 sparse matrix of type '<class 'numpy.float64'>'
with 25 stored elements in Compressed Sparse Row format>
I've increased the shape and number of nonzero elements by roughly 100 (the number of columns).
Defining foo, and copying Divakar's tests:
def foo(A,B):
r,c = A.nonzero()
C = A*B
C[r,c] = A[r,c]
return C
In [83]: timeit naive(A,B)
100 loops, best of 3: 2.53 ms per loop
In [84]: timeit proposed(A,B)
/...
SparseEfficiencyWarning)
100 loops, best of 3: 4.48 ms per loop
In [85]: timeit foo(A,B)
...
SparseEfficiencyWarning)
100 loops, best of 3: 2.13 ms per loop
So my version has a modest speed inprovement. As Divakar found out, changing sparsity changes the relative advantages. I expect size to also change them.
The fact that A.nonzero uses the coo format, suggests it might be feasible to construct the new array with that format. A lot of sparse code builds a new matrix via the coo values.
In [97]: Co=C.tocoo()
In [98]: Ao=A.tocoo()
In [99]: r=np.concatenate((Co.row,Ao.row))
In [100]: c=np.concatenate((Co.col,Ao.col))
In [101]: d=np.concatenate((Co.data,Ao.data))
In [102]: r.shape
Out[102]: (79,)
In [103]: C1=sparse.csr_matrix((d,(r,c)),shape=A.shape)
In [104]: C1
Out[104]:
<250x100 sparse matrix of type '<class 'numpy.float64'>'
with 78 stored elements in Compressed Sparse Row format>
This C1 has, I think, the same non-zero elements as the C constructed by other means. But I think one value is different because the r is longer. In this particular example, C and A share one nonzero element, and the coo style of input sums those, where as we'd prefer to have A values overwrite everything.
If you can tolerate this discrepancy, this is a faster way (at least for this test case):
def bar(A,B):
C=A*B
Co=C.tocoo()
Ao=A.tocoo()
r=np.concatenate((Co.row,Ao.row))
c=np.concatenate((Co.col,Ao.col))
d=np.concatenate((Co.data,Ao.data))
return sparse.csr_matrix((d,(r,c)),shape=A.shape)
In [107]: timeit bar(A,B)
1000 loops, best of 3: 1.03 ms per loop
Cracked it! Well, there's a lot of scipy stuffs specific to sparse matrices that I learnt along the way. Here's the implementation that I could muster -
# Find the indices in output array that are to be updated
R,C = ((A!=0).dot(B!=0)).nonzero()
mask = np.asarray(A[R,C]==0).ravel()
R,C = R[mask],C[mask]
# Make a copy of A and get the dot product through sliced rows and columns
# off A and B using the definition of matrix-multiplication
out = A.copy()
out[R,C] = (A[R].multiply(B[:,C].T).sum(1)).ravel()
The most expensive part seems to be element-wise multiplication and summing. On some quick timing tests, it seems that this would be good on a sparse matrices with a high degree of sparsity to beat the original dot-mask-based solution in terms of performance, which I think comes from its focus on memory efficiency.
Runtime test
Function definitions -
def naive(A, B):
mask = A != 0
out = A.dot(B).tolil()
out[mask] = A[mask]
return out.tocsr()
def proposed(A, B):
R,C = ((A!=0).dot(B!=0)).nonzero()
mask = np.asarray(A[R,C]==0).ravel()
R,C = R[mask],C[mask]
out = A.copy()
out[R,C] = (A[R].multiply(B[:,C].T).sum(1)).ravel()
return out
Timings -
In [57]: # Input matrices
...: M,N = 25000, 170
...: A = sparse.rand(M, N, density=0.001, format='csr')
...: B = sparse.rand(N, N, density=0.02, format='csr')
...:
In [58]: %timeit naive(A, B)
10 loops, best of 3: 92.2 ms per loop
In [59]: %timeit proposed(A, B)
10 loops, best of 3: 132 ms per loop
In [60]: # Input matrices with increased sparse-ness
...: M,N = 25000, 170
...: A = sparse.rand(M, N, density=0.0001, format='csr')
...: B = sparse.rand(N, N, density=0.002, format='csr')
...:
In [61]: %timeit naive(A, B)
10 loops, best of 3: 78.1 ms per loop
In [62]: %timeit proposed(A, B)
100 loops, best of 3: 8.03 ms per loop
Python isn't my main language, but I thought this was an interesting problem and I wanted to give this a stab :)
Preliminaries:
import numpy
import scipy.sparse
# example matrices and sparse versions
A = numpy.array([[1, 2, 0, 1], [1, 0, 1, 2], [0, 1, 2 ,1], [1, 2, 1, 0]])
B = numpy.array([[1,2,3,4],[1,2,3,4],[1,2,3,4],[1,2,3,4]])
A_s = scipy.sparse.lil_matrix(A)
B_s = scipy.sparse.lil_matrix(B)
So you want to convert the original problem of:
C = A.dot(B)
C[A.nonzero()] = A[A.nonzero()]
To something sparse-y.
Just to get this out of the way, the direct "sparse" translation of the above is:
C_s = A_s.dot(B_s)
C_s[A_s.nonzero()] = A_s[A_s.nonzero()]
But it sounds like you're not happy about this, as it calculates all the dot products first, which you worry might be inefficient.
So, your question is, if you find the zeros first, and only evaluate dot products on those elements, will that be faster? I.e. for a dense matrix this could be something like:
Xs, Ys = numpy.nonzero(A==0)
D = A[:]
D[Xs, Ys] = map ( lambda x,y: A[x,:].dot(B[:,y]), Xs, Ys)
Let's translate this to a sparse matrix. My main stumbling block here was finding the "Zero" indices; since A_s==0 doesn't make sense for sparse matrices, I found them this way:
Xmax, Ymax = A_s.shape
DenseSize = Xmax * Ymax
Xgrid, Ygrid = numpy.mgrid[0:Xmax, 0:Ymax]
Ygrid = Ygrid.reshape([DenseSize,1])[:,0]
Xgrid = Xgrid.reshape([DenseSize,1])[:,0]
AllIndices = numpy.array([Xgrid, Ygrid])
NonzeroIndices = numpy.array(A_s.nonzero())
ZeroIndices = numpy.array([x for x in AllIndices.T.tolist() if x not in NonzeroIndices.T.tolist()]).T
If you know of a better / faster way, by all means try it. Once we have the Zero indices, we can do a similar mapping as before:
D_s = A_s[:]
D_s[ZeroIndices[0], ZeroIndices[1]] = map ( lambda x, y : A_s[x,:].dot(B[:,y])[0], ZeroIndices[0], ZeroIndices[1] )
which gives you your sparse matrix result.
Now I don't know if this is faster or not. I mostly took a stab because it was an interesting problem, and to see if I could do it in python. In fact I suspect it might not be faster than direct whole-matrix dotproduct, because it uses listcomprehensions and mapping on a large dataset (like you say, you expect a lot of zeros). But it is an answer to your question of "how can I only calculate dot products for the zero values without doing multiplying the matrices as a whole". I'd be interested to see if you do try this how it compares in terms of speed on your datasets.
EDIT: I'm putting below an example "block processing" version based on the above, which I think should allow you to process your large dataset without problems. Let me know if it works.
import numpy
import scipy.sparse
# example matrices and sparse versions
A = numpy.array([[1, 2, 0, 1], [1, 0, 1, 2], [0, 1, 2 ,1], [1, 2, 1, 0]])
B = numpy.array([[1,2,3,4],[1,2,3,4],[1,2,3,4],[1,2,3,4]])
A_s = scipy.sparse.lil_matrix(A)
B_s = scipy.sparse.lil_matrix(B)
# Choose a grid division (i.e. how many processing blocks you want to create)
BlockGrid = numpy.array([2,2])
D_s = A_s[:] # initialise from A
Xmax, Ymax = A_s.shape
BaseBSiz = numpy.array([Xmax, Ymax]) / BlockGrid
for BIndX in range(0, Xmax, BlockGrid[0]):
for BIndY in range(0, Ymax, BlockGrid[1]):
BSizX, BSizY = D_s[ BIndX : BIndX + BaseBSiz[0], BIndY : BIndY + BaseBSiz[1] ].shape
Xgrid, Ygrid = numpy.mgrid[BIndX : BIndX + BSizX, BIndY : BIndY + BSizY]
Xgrid = Xgrid.reshape([BSizX*BSizY,1])[:,0]
Ygrid = Ygrid.reshape([BSizX*BSizY,1])[:,0]
AllInd = numpy.array([Xgrid, Ygrid]).T
NZeroInd = numpy.array(A_s[Xgrid, Ygrid].reshape((BSizX,BSizY)).nonzero()).T + numpy.array([[BIndX],[BIndY]]).T
ZeroInd = numpy.array([x for x in AllInd.tolist() if x not in NZeroInd.tolist()]).T
#
# Replace zero-values in current block
D_s[ZeroInd[0], ZeroInd[1]] = map ( lambda x, y : A_s[x,:].dot(B[:,y])[0], ZeroInd[0], ZeroInd[1] )
I'm trying to compute the pairwise np.vdot of a complex 2D array x with itself. So the behaviour I want is:
X = np.empty((x.shape[0], x.shape[0]), dtype='complex128')
for i in range(x.shape[0]):
for j in range(x.shape[0]):
X[i, j] = np.vdot(x[i], x[j])
Is there a way to do this without the explicit loops? I tried using pairwise_kernel from sklearn but it assumes the input arrays are real numbers. I also tried broadcasting, but vdot flattens its inputs.
X = np.einsum('ik,jk->ij', np.conj(x), x)
is equivalent to
X = np.empty((x.shape[0], x.shape[0]), dtype='complex128')
for i in range(x.shape[0]):
for j in range(x.shape[0]):
X[i, j] = np.vdot(x[i], x[j])
np.einsum
takes a sum of products. The subscript 'ik,jk->ij' tells np.einsum that the second argument,
np.conj(x) is an array with subscripts ik and the third argument, x has
subscripts jk. Thus, the product np.conj(x)[i,k]*x[j,k] is computed for all
i,j,k. The sum is taken over the repeated subscript, k, and since that
leaves i and j remaining, they become the subscripts of the resultant array.
For example,
import numpy as np
N, M = 10, 20
a = np.random.random((N,M))
b = np.random.random((N,M))
x = a + b*1j
def orig(x):
X = np.empty((x.shape[0], x.shape[0]), dtype='complex128')
for i in range(x.shape[0]):
for j in range(x.shape[0]):
X[i, j] = np.vdot(x[i], x[j])
return X
def alt(x):
return np.einsum('ik,jk->ij', np.conj(x), x)
assert np.allclose(orig(x), alt(x))
In [307]: %timeit orig(x)
10000 loops, best of 3: 143 µs per loop
In [308]: %timeit alt(x)
100000 loops, best of 3: 8.63 µs per loop
To extend the np.vdot to all rows, you can use np.tensordot and I am borrowing the conjugate idea straight off #unutbu's solution , like so -
np.tensordot(np.conj(x),x,axes=(1,1))
Basically with np.tensordot, we specify the axes to be reduced, which in this case is the last axis for the conjugate version of x and the array itself, when applied on those two.
Runtime test -
Let's time #unutbu's solution with np.einsum and the proposed solution in this post -
In [27]: import numpy as np # From #unutbu's` solution again
...:
...: N, M = 1000, 1000
...: a = np.random.random((N,M))
...: b = np.random.random((N,M))
...: x = a + b*1j
...:
In [28]: %timeit np.einsum('ik,jk->ij', np.conj(x), x) # #unutbu's` solution
1 loops, best of 3: 4.45 s per loop
In [29]: %timeit np.tensordot(np.conj(x),x,axes=(1,1))
1 loops, best of 3: 3.76 s per loop
Suppose I have an N-dimensional numpy array x and an (N-1)-dimensional index array m (for example, m = x.argmax(axis=-1)). I'd like to construct (N-1) dimensional array y such that y[i_1, ..., i_N-1] = x[i_1, ..., i_N-1, m[i_1, ..., i_N-1]] (for the argmax example above it would be equivalent to y = x.max(axis=-1)).
For N=3 I could achieve what I want by
y = x[np.arange(x.shape[0])[:, np.newaxis], np.arange(x.shape[1]), m]
The question is, how do I do this for an arbitrary N?
you can use indices :
firstdims=np.indices(x.shape[:-1])
And add yours :
ind=tuple(firstdims)+(m,)
Then x[ind] is what you want.
In [228]: allclose(x.max(-1),x[ind])
Out[228]: True
Here's one approach using reshaping and linear indexing to handle multi-dimensional arrays of arbitrary dimensions -
shp = x.shape[:-1]
n_ele = np.prod(shp)
y_out = x.reshape(n_ele,-1)[np.arange(n_ele),m.ravel()].reshape(shp)
Let's take a sample case with a ndarray of 6 dimensions and let's say we are using m = x.argmax(axis=-1) to index into the last dimension. So, the output would be x.max(-1). Let's verify this for the proposed solution -
In [121]: x = np.random.randint(0,9,(4,5,3,3,2,4))
In [122]: m = x.argmax(axis=-1)
In [123]: shp = x.shape[:-1]
...: n_ele = np.prod(shp)
...: y_out = x.reshape(n_ele,-1)[np.arange(n_ele),m.ravel()].reshape(shp)
...:
In [124]: np.allclose(x.max(-1),y_out)
Out[124]: True
I liked #B. M.'s solution for its elegance. So, here's a runtime test to benchmark these two -
def reshape_based(x,m):
shp = x.shape[:-1]
n_ele = np.prod(shp)
return x.reshape(n_ele,-1)[np.arange(n_ele),m.ravel()].reshape(shp)
def indices_based(x,m): ## #B. M.'s solution
firstdims=np.indices(x.shape[:-1])
ind=tuple(firstdims)+(m,)
return x[ind]
Timings -
In [152]: x = np.random.randint(0,9,(4,5,3,3,4,3,6,2,4,2,5))
...: m = x.argmax(axis=-1)
...:
In [153]: %timeit indices_based(x,m)
10 loops, best of 3: 30.2 ms per loop
In [154]: %timeit reshape_based(x,m)
100 loops, best of 3: 5.14 ms per loop