Computing `AB⁻¹` with `np.linalg.solve()` - python

I need to compute AB⁻¹ in Python / Numpy for two matrices A and B (B being square, of course).
I know that np.linalg.inv() would allow me to compute B⁻¹, which I can then multiply with A.
I also know that B⁻¹A is actually better computed with np.linalg.solve().
Inspired by that, I decided to rewrite AB⁻¹ in terms of np.linalg.solve().
I got to a formula, based on the identity (AB)ᵀ = BᵀAᵀ, which uses np.linalg.solve() and .transpose():
np.linalg.solve(a.transpose(), b.transpose()).transpose()
that seems to be doing the job:
import numpy as np
n, m = 4, 2
np.random.seed(0)
a = np.random.random((n, n))
b = np.random.random((m, n))
print(np.matmul(b, np.linalg.inv(a)))
# [[ 2.87169378 -0.04207382 -1.10553758 -0.83200471]
# [-1.08733434 1.00110176 0.79683577 0.67487591]]
print(np.linalg.solve(a.transpose(), b.transpose()).transpose())
# [[ 2.87169378 -0.04207382 -1.10553758 -0.83200471]
# [-1.08733434 1.00110176 0.79683577 0.67487591]]
print(np.all(np.isclose(np.matmul(b, np.linalg.inv(a)), np.linalg.solve(a.transpose(), b.transpose()).transpose())))
# True
and also comes up much faster for sufficiently large inputs:
n, m = 400, 200
np.random.seed(0)
a = np.random.random((n, n))
b = np.random.random((m, n))
print(np.all(np.isclose(np.matmul(b, np.linalg.inv(a)), np.linalg.solve(a.transpose(), b.transpose()).transpose())))
# True
%timeit np.matmul(b, np.linalg.inv(a))
# 100 loops, best of 3: 13.3 ms per loop
%timeit np.linalg.solve(a.transpose(), b.transpose()).transpose()
# 100 loops, best of 3: 7.71 ms per loop
My question is: does this identity always stand correct or there are some corner cases I am overlooking?

In general, np.linalg.solve(B, A) is equivalent to B-1A. The rest is just math.
In all cases, (AB)T = BTAT: https://math.stackexchange.com/q/1440305/295281.
Not necessary for this case, but for invertible matrices, (AB)-1 = B-1A-1: https://math.stackexchange.com/q/688339/295281.
For an invertible matrix, it is also the case that (A-1)T = (AT)-1: https://math.stackexchange.com/q/340233/295281.
From that it follows that (AB-1)T = (B-1)TAT = (BT)-1AT. As long as B is invertible, you should have no issues with the transformation you propose in any case.

Related

Why np.dot's perfomance in loop is better than without loop [duplicate]

I'm getting some efficiency test results that I can't explain.
I want to assemble a matrix B whose i-th entries B[i,:,:] = A[i,:,:].dot(x), where each A[i,:,:] is a 2D matrix, and so is x.
I can do this three ways, to test performance I make random (numpy.random.randn) matrices A = (10,1000,1000), x = (1000,1200). I get the following time results:
(1) single multidimensional dot product
B = A.dot(x)
total time: 102.361 s
(2) looping through i and performing 2D dot products
# initialize B = np.zeros([dim1, dim2, dim3])
for i in range(A.shape[0]):
B[i,:,:] = A[i,:,:].dot(x)
total time: 0.826 s
(3) numpy.einsum
B3 = np.einsum("ijk, kl -> ijl", A, x)
total time: 8.289 s
So, option (2) is the fastest by far. But, considering just (1) and (2), I don't see the big difference between them. How can looping through and doing 2D dot products be ~ 124 times faster? They both use numpy.dot. Any insights?
I include the code used for the above results just below:
import numpy as np
import numpy.random as npr
import time
dim1, dim2, dim3 = 10, 1000, 1200
A = npr.randn(dim1, dim2, dim2)
x = npr.randn(dim2, dim3)
# consider three ways of assembling the same matrix B: B1, B2, B3
t = time.time()
B1 = np.dot(A,x)
td1 = time.time() - t
print "a single dot product of A [shape = (%d, %d, %d)] with x [shape = (%d, %d)] completes in %.3f s" \
% (A.shape[0], A.shape[1], A.shape[2], x.shape[0], x.shape[1], td1)
B2 = np.zeros([A.shape[0], x.shape[0], x.shape[1]])
t = time.time()
for i in range(A.shape[0]):
B2[i,:,:] = np.dot(A[i,:,:], x)
td2 = time.time() - t
print "taking %d dot products of 2D dot products A[i,:,:] [shape = (%d, %d)] with x [shape = (%d, %d)] completes in %.3f s" \
% (A.shape[0], A.shape[1], A.shape[2], x.shape[0], x.shape[1], td2)
t = time.time()
B3 = np.einsum("ijk, kl -> ijl", A, x)
td3 = time.time() - t
print "using np.einsum, it completes in %.3f s" % td3
With smaller dims 10,100,200, I get a similar ranking
In [355]: %%timeit
.....: B=np.zeros((N,M,L))
.....: for i in range(N):
B[i,:,:]=np.dot(A[i,:,:],x)
.....:
10 loops, best of 3: 22.5 ms per loop
In [356]: timeit np.dot(A,x)
10 loops, best of 3: 44.2 ms per loop
In [357]: timeit np.einsum('ijk,km->ijm',A,x)
10 loops, best of 3: 29 ms per loop
In [367]: timeit np.dot(A.reshape(-1,M),x).reshape(N,M,L)
10 loops, best of 3: 22.1 ms per loop
In [375]: timeit np.tensordot(A,x,(2,0))
10 loops, best of 3: 22.2 ms per loop
the itererative is faster, though not by as much as in your case.
This is probably true as long as that iterating dimension is small compared to the other ones. In that case the overhead of iteration (function calls etc) is small compared to the calculation time. And doing all the values at once uses more memory.
I tried a dot variation where I reshaped A into 2d, thinking that dot does that kind of reshaping internally. I'm surprised that it is actually fastest. tensordot is probably doing the same reshaping (that code if Python readable).
einsum sets up a 'sum of products' iteration involving 4 variables, the i,j,k,m - that is dim1*dim2*dim2*dim3 steps with the C level nditer. So the more indices you have the larger the iteration space.
numpy.dot only delegates to a BLAS matrix multiply when the inputs each have dimension at most 2:
#if defined(HAVE_CBLAS)
if (PyArray_NDIM(ap1) <= 2 && PyArray_NDIM(ap2) <= 2 &&
(NPY_DOUBLE == typenum || NPY_CDOUBLE == typenum ||
NPY_FLOAT == typenum || NPY_CFLOAT == typenum)) {
return cblas_matrixproduct(typenum, ap1, ap2, out);
}
#endif
When you stick your whole 3-dimensional A array into dot, NumPy takes a slower path, going through an nditer object. It still tries to get some use out of BLAS in the slow path, but the way the slow path is designed, it can only use a vector-vector multiply rather than a matrix-matrix multiply, which doesn't give the BLAS anywhere near as much room to optimize.
I am not too familiar with numpy's C-API, and the numpy.dot is one such builtin function that used to be under _dotblas in earlier versions.
Nevertheless, here are my thoughts.
1) numpy.dot takes different paths for 2-dimensional arrays and n-dimensional arrays. From the numpy.dot's online documentation:
For 2-D arrays it is equivalent to matrix multiplication, and for 1-D arrays to inner product of vectors (without complex conjugation). For N dimensions it is a sum product over the last axis of a and the second-to-last of b
dot(a, b)[i,j,k,m] = sum(a[i,j,:] * b[k,:,m])
So for 2-D arrays you are always guaranteed to have one call to BLAS's dgemm, however for N-D arrays numpy might choose the multiplication axes for arrays which might not correspond to the fastest changing axis (as you can see from the excerpt I have posted), and as result the full power of dgemm could be missed out on.
2) Your A array is too large to be loaded on to CPU cache. In your example, you use A with dimensions (10,1000,1000) which gives
In [1]: A.nbytes
80000000
In [2]: 80000000/1024
78125
That is almost 80MB, much larger than your cache size. So again you lose most of dgemm's power right there.
3) You are also timing the functions somewhat imprecisely. The time function in Python is known to be not precise. Use timeit instead.
So having all the above points in mind, let's try experimenting with arrays that can be loaded on to the cache
dim1, dim2, dim3 = 20, 20, 20
A = np.random.rand(dim1, dim2, dim2)
x = np.random.rand(dim2, dim3)
def for_dot1(A,x):
for i in range(A.shape[0]):
np.dot(A[i,:,:], x)
def for_dot2(A,x):
for i in range(A.shape[0]):
np.dot(A[:,i,:], x)
def for_dot3(A,x):
for i in range(A.shape[0]):
np.dot(A[:,:,i], x)
and here are the timings that I get (using numpy 1.9.2 built against OpenBLAS 0.2.14):
In [3]: %timeit np.dot(A,x)
10000 loops, best of 3: 174 µs per loop
In [4]: %timeit np.einsum("ijk, kl -> ijl", A, x)
10000 loops, best of 3: 108 µs per loop
In [5]: %timeit np.einsum("ijk, lk -> ijl", A, x)
10000 loops, best of 3: 97.1 µs per loop
In [6]: %timeit np.einsum("ikj, kl -> ijl", A, x)
1000 loops, best of 3: 238 µs per loop
In [7]: %timeit np.einsum("kij, kl -> ijl", A, x)
10000 loops, best of 3: 113 µs per loop
In [8]: %timeit for_dot1(A,x)
10000 loops, best of 3: 101 µs per loop
In [9]: %timeit for_dot2(A,x)
10000 loops, best of 3: 131 µs per loop
In [10]: %timeit for_dot3(A,x)
10000 loops, best of 3: 133 µs per loop
Notice that there is still a time difference, but not in orders of magnitude. Also note the importance of choosing the axis of multiplication. Now perhaps, a numpy developer can shed some light on what numpy.dot actually does under the hood for N-D arrays.

Segmentation using maximum likelihood algorithm on images using python

I would like to perform image segmentation using maximum likelihood algorithm implemented in python.
The mean vectors of the classes, and covariance matrices are known, and iterating over the images (which are quite big...5100X7020) we can calculate for each pixel the probability of being part of the given class.
Simply written in Python
import numpy as np
from numpy.linalg import inv
from numpy.linalg import det
...
probImage1 = []
probImage1Vector = []
norm = 1.0 / (np.power((2*np.pi), i/2) * np.sqrt(np.linalg.det(covMatrixClass1)))
covMatrixInverz = np.linalg.inv(covMatrixClass1)
for x in xrange(x_img):
for y in xrange(y_img):
X = realImage[x,y]
pixelValueDifference = X - meanVectorClass1
mult1 = np.multiply(-0.5,np.transpose(pixelValueDifference))
mult2 = np.dot(covMatrixInverz,pixelValueDifference)
multMult = np.dot(mult1,mult2)
expo = np.exp(multMult)
probImage1Vector.append(np.multiply(norm,expo))
probImage1.append(probImage1Vector)
probImage1Vector = []
The problem that this code is very slow when performing on large images...
The calculations like vector subtraction and multiplication consumes a lot of time, even though they are only 1X3 vectors.
Could you please give a hint how to speed up this code? I would really appreciate. Sorry, if I was not clear I am still beginner in python.
Taking a closer look at :
mult1 = np.multiply(-0.5,np.transpose(pixelValueDifference))
mult2 = np.dot(covMatrixInverz,pixelValueDifference)
multMult = np.dot(mult1,mult2)
We see that the operation is basically :
A.T (d) C (d) A # where `(d)` is the dot-product
Those three steps could be easily expressed as one string notation in np.einsum, like so -
np.einsum('k,lk,l->',pA,covMatrixInverz,-0.5*pA)
Performing this across both iterators i(=x) and j(=y), we would have a fully vectorized expression -
np.einsum('ijk,lk,ijl->ij',pA,covMatrixInverz,-0.5*pA))
Alternatively, we could perform the first part of sume-reduction with np.tensordot -
mult2_vectorized = np.tensordot(pA, covMatrixInverz, axes=([2],[1]))
output = np.einsum('ijk,ijk->ij',-0.5*pA, mult2_vectorized)
Benchmarking
Listing all approaches as functions -
# Original code posted by OP to return array
def org_app(meanVectorClass1, realImage, covMatrixInverz, norm):
probImage1 = []
probImage1Vector = []
x_img, y_img = realImage.shape[:2]
for x in xrange(x_img):
for y in xrange(y_img):
X = realImage[x,y]
pixelValueDifference = X - meanVectorClass1
mult1 = np.multiply(-0.5,np.transpose(pixelValueDifference))
mult2 = np.dot(covMatrixInverz,pixelValueDifference)
multMult = np.dot(mult1,mult2)
expo = np.exp(multMult)
probImage1Vector.append(np.multiply(norm,expo))
probImage1.append(probImage1Vector)
probImage1Vector = []
return np.asarray(probImage1).reshape(x_img,y_img)
def vectorized(meanVectorClass1, realImage, covMatrixInverz, norm):
pA = realImage - meanVectorClass1
mult2_vectorized = np.tensordot(pA, covMatrixInverz, axes=([2],[1]))
return np.exp(np.einsum('ijk,ijk->ij',-0.5*pA, mult2_vectorized))*norm
def vectorized2(meanVectorClass1, realImage, covMatrixInverz, norm):
pA = realImage - meanVectorClass1
return np.exp(np.einsum('ijk,lk,ijl->ij',pA,covMatrixInverz,-0.5*pA))*norm
Timings -
In [19]: # Setup inputs
...: meanVectorClass1 = np.array([23.96000000, 58.159999, 61.5399])
...:
...: covMatrixClass1 = np.array([[ 514.20040404, 461.68323232, 364.35515152],
...: [ 461.68323232, 519.63070707, 446.48848485],
...: [ 364.35515152, 446.48848485, 476.37212121]])
...: covMatrixInverz = np.linalg.inv(covMatrixClass1)
...:
...: norm = 0.234 # Random float number
...: realImage = np.random.rand(1000,2000,3)
...:
In [20]: out1 = org_app(meanVectorClass1, realImage, covMatrixInverz, norm )
...: out2 = vectorized(meanVectorClass1, realImage, covMatrixInverz, norm )
...: out3 = vectorized2(meanVectorClass1, realImage, covMatrixInverz, norm )
...: print np.allclose(out1, out2)
...: print np.allclose(out1, out3)
...:
True
True
In [21]: %timeit org_app(meanVectorClass1, realImage, covMatrixInverz, norm )
1 loops, best of 3: 27.8 s per loop
In [22]: %timeit vectorized(meanVectorClass1, realImage, covMatrixInverz, norm )
1 loops, best of 3: 182 ms per loop
In [23]: %timeit vectorized2(meanVectorClass1, realImage, covMatrixInverz, norm )
1 loops, best of 3: 275 ms per loop
Looks like the fully vectorized einsum + tensordot hybrid solution is doing pretty good!
For further performance boost, one can also look into numexpr module to speedup the exponential computations on large arrays.
As a first step, I would get rid of unnecessary function calls like transpose, dot, and multiply. These are all simple calculations which you should be doing inline. When you can actually see what you are doing, instead of hiding things inside of functions, it will be easier to understand the performance problems.
The fundamental issue here is that this appears to be at least a quartic complexity operation. You might want to simply multiply out how many operations you are doing in all of your loops. Is it 500 million, 2 billion, 350 billion? How many?
To get control of performance you need to understand how many instructions you are doing. A modern computer can do about 1 billion instructions per second, but if memory movements are involved, it can be substantially slower.

Faster Python Cosine dissimilarity between Scipy CSR "vectors"

It's a classic question, but I believe many people still searching for answers.
This question is a different than this one, since my question is operation between two sparse vectors (not a matrix).
I wrote a blog post about how Cosine Scipy Spatial Distance (SSD) is getting slower when the dimension of the data is getting higher (because it works on dense vectors). The post is in Indonesian language, but the code, my experiment settings & results should be easily understandable regardless of the language (at the bottom of the post).
Currently this solution is more than 70 times faster for high dimension data (compared to SSD) & more memory efficient:
import numpy as np
def fCosine(u,v): # u,v CSR vectors, Cosine Dissimilarity
uData = u.data; vData = v.data
denominator = np.sqrt(np.sum(uData**2)) * np.sqrt(np.sum(vData**2))
if denominator>0:
uCol = u.indices; vCol = v.indices # np array
intersection = set(np.intersect1d(uCol,vCol))
uI = np.array([u1 for i,u1 in enumerate(uData) if uCol[i] in intersection])
vI = np.array([v2 for j,v2 in enumerate(vData) if vCol[j] in intersection])
return 1-np.dot(uI,vI)/denominator
else:
return float("inf")
Is it possible to even further improve that function (Pythonic or via JIT/Cython???).
Here is an alternative, alt_fCosine, which (on my machine) is about 3x faster for CSR vectors of size 10**5 and 10**4 non-zero elements:
import scipy.sparse as sparse
import numpy as np
import math
def fCosine(u,v): # u,v CSR vectors, Cosine Dissimilarity
uData = u.data; vData = v.data
denominator = np.sqrt(np.sum(uData**2)) * np.sqrt(np.sum(vData**2))
if denominator>0:
uCol = u.indices; vCol = v.indices # np array
intersection = set(np.intersect1d(uCol,vCol))
uI = np.array([u1 for i,u1 in enumerate(uData) if uCol[i] in intersection])
vI = np.array([v2 for j,v2 in enumerate(vData) if vCol[j] in intersection])
return 1-np.dot(uI,vI)/denominator
else:
return float("inf")
def alt_fCosine(u,v):
uData, vData = u.data, v.data
denominator = math.sqrt(np.sum(uData**2) * np.sum(vData**2))
if denominator>0:
uCol, vCol = u.indices, v.indices
uI = uData[np.in1d(uCol, vCol)]
vI = vData[np.in1d(vCol, uCol)]
return 1-np.dot(uI,vI)/denominator
else:
return float("inf")
# Check that they return the same result
N = 10**5
u = np.round(10*sparse.random(1, N, density=0.1, format='csr'))
v = np.round(10*sparse.random(1, N, density=0.1, format='csr'))
assert np.allclose(fCosine(u, v), alt_fCosine(u, v))
alt_fCosine replaces two list comprehensions, a call to np.intersection1d
and the formation of a Python set with two calls to np.in1d and advanced
indexing.
For N = 10**5:
In [322]: %timeit fCosine(u, v)
100 loops, best of 3: 5.73 ms per loop
In [323]: %timeit alt_fCosine(u, v)
1000 loops, best of 3: 1.62 ms per loop
In [324]: 5.73/1.62
Out[324]: 3.537037037037037

Why is B = numpy.dot(A,x) so much slower looping through doing B[i,:,:] = numpy.dot(A[i,:,:],x) )?

I'm getting some efficiency test results that I can't explain.
I want to assemble a matrix B whose i-th entries B[i,:,:] = A[i,:,:].dot(x), where each A[i,:,:] is a 2D matrix, and so is x.
I can do this three ways, to test performance I make random (numpy.random.randn) matrices A = (10,1000,1000), x = (1000,1200). I get the following time results:
(1) single multidimensional dot product
B = A.dot(x)
total time: 102.361 s
(2) looping through i and performing 2D dot products
# initialize B = np.zeros([dim1, dim2, dim3])
for i in range(A.shape[0]):
B[i,:,:] = A[i,:,:].dot(x)
total time: 0.826 s
(3) numpy.einsum
B3 = np.einsum("ijk, kl -> ijl", A, x)
total time: 8.289 s
So, option (2) is the fastest by far. But, considering just (1) and (2), I don't see the big difference between them. How can looping through and doing 2D dot products be ~ 124 times faster? They both use numpy.dot. Any insights?
I include the code used for the above results just below:
import numpy as np
import numpy.random as npr
import time
dim1, dim2, dim3 = 10, 1000, 1200
A = npr.randn(dim1, dim2, dim2)
x = npr.randn(dim2, dim3)
# consider three ways of assembling the same matrix B: B1, B2, B3
t = time.time()
B1 = np.dot(A,x)
td1 = time.time() - t
print "a single dot product of A [shape = (%d, %d, %d)] with x [shape = (%d, %d)] completes in %.3f s" \
% (A.shape[0], A.shape[1], A.shape[2], x.shape[0], x.shape[1], td1)
B2 = np.zeros([A.shape[0], x.shape[0], x.shape[1]])
t = time.time()
for i in range(A.shape[0]):
B2[i,:,:] = np.dot(A[i,:,:], x)
td2 = time.time() - t
print "taking %d dot products of 2D dot products A[i,:,:] [shape = (%d, %d)] with x [shape = (%d, %d)] completes in %.3f s" \
% (A.shape[0], A.shape[1], A.shape[2], x.shape[0], x.shape[1], td2)
t = time.time()
B3 = np.einsum("ijk, kl -> ijl", A, x)
td3 = time.time() - t
print "using np.einsum, it completes in %.3f s" % td3
With smaller dims 10,100,200, I get a similar ranking
In [355]: %%timeit
.....: B=np.zeros((N,M,L))
.....: for i in range(N):
B[i,:,:]=np.dot(A[i,:,:],x)
.....:
10 loops, best of 3: 22.5 ms per loop
In [356]: timeit np.dot(A,x)
10 loops, best of 3: 44.2 ms per loop
In [357]: timeit np.einsum('ijk,km->ijm',A,x)
10 loops, best of 3: 29 ms per loop
In [367]: timeit np.dot(A.reshape(-1,M),x).reshape(N,M,L)
10 loops, best of 3: 22.1 ms per loop
In [375]: timeit np.tensordot(A,x,(2,0))
10 loops, best of 3: 22.2 ms per loop
the itererative is faster, though not by as much as in your case.
This is probably true as long as that iterating dimension is small compared to the other ones. In that case the overhead of iteration (function calls etc) is small compared to the calculation time. And doing all the values at once uses more memory.
I tried a dot variation where I reshaped A into 2d, thinking that dot does that kind of reshaping internally. I'm surprised that it is actually fastest. tensordot is probably doing the same reshaping (that code if Python readable).
einsum sets up a 'sum of products' iteration involving 4 variables, the i,j,k,m - that is dim1*dim2*dim2*dim3 steps with the C level nditer. So the more indices you have the larger the iteration space.
numpy.dot only delegates to a BLAS matrix multiply when the inputs each have dimension at most 2:
#if defined(HAVE_CBLAS)
if (PyArray_NDIM(ap1) <= 2 && PyArray_NDIM(ap2) <= 2 &&
(NPY_DOUBLE == typenum || NPY_CDOUBLE == typenum ||
NPY_FLOAT == typenum || NPY_CFLOAT == typenum)) {
return cblas_matrixproduct(typenum, ap1, ap2, out);
}
#endif
When you stick your whole 3-dimensional A array into dot, NumPy takes a slower path, going through an nditer object. It still tries to get some use out of BLAS in the slow path, but the way the slow path is designed, it can only use a vector-vector multiply rather than a matrix-matrix multiply, which doesn't give the BLAS anywhere near as much room to optimize.
I am not too familiar with numpy's C-API, and the numpy.dot is one such builtin function that used to be under _dotblas in earlier versions.
Nevertheless, here are my thoughts.
1) numpy.dot takes different paths for 2-dimensional arrays and n-dimensional arrays. From the numpy.dot's online documentation:
For 2-D arrays it is equivalent to matrix multiplication, and for 1-D arrays to inner product of vectors (without complex conjugation). For N dimensions it is a sum product over the last axis of a and the second-to-last of b
dot(a, b)[i,j,k,m] = sum(a[i,j,:] * b[k,:,m])
So for 2-D arrays you are always guaranteed to have one call to BLAS's dgemm, however for N-D arrays numpy might choose the multiplication axes for arrays which might not correspond to the fastest changing axis (as you can see from the excerpt I have posted), and as result the full power of dgemm could be missed out on.
2) Your A array is too large to be loaded on to CPU cache. In your example, you use A with dimensions (10,1000,1000) which gives
In [1]: A.nbytes
80000000
In [2]: 80000000/1024
78125
That is almost 80MB, much larger than your cache size. So again you lose most of dgemm's power right there.
3) You are also timing the functions somewhat imprecisely. The time function in Python is known to be not precise. Use timeit instead.
So having all the above points in mind, let's try experimenting with arrays that can be loaded on to the cache
dim1, dim2, dim3 = 20, 20, 20
A = np.random.rand(dim1, dim2, dim2)
x = np.random.rand(dim2, dim3)
def for_dot1(A,x):
for i in range(A.shape[0]):
np.dot(A[i,:,:], x)
def for_dot2(A,x):
for i in range(A.shape[0]):
np.dot(A[:,i,:], x)
def for_dot3(A,x):
for i in range(A.shape[0]):
np.dot(A[:,:,i], x)
and here are the timings that I get (using numpy 1.9.2 built against OpenBLAS 0.2.14):
In [3]: %timeit np.dot(A,x)
10000 loops, best of 3: 174 µs per loop
In [4]: %timeit np.einsum("ijk, kl -> ijl", A, x)
10000 loops, best of 3: 108 µs per loop
In [5]: %timeit np.einsum("ijk, lk -> ijl", A, x)
10000 loops, best of 3: 97.1 µs per loop
In [6]: %timeit np.einsum("ikj, kl -> ijl", A, x)
1000 loops, best of 3: 238 µs per loop
In [7]: %timeit np.einsum("kij, kl -> ijl", A, x)
10000 loops, best of 3: 113 µs per loop
In [8]: %timeit for_dot1(A,x)
10000 loops, best of 3: 101 µs per loop
In [9]: %timeit for_dot2(A,x)
10000 loops, best of 3: 131 µs per loop
In [10]: %timeit for_dot3(A,x)
10000 loops, best of 3: 133 µs per loop
Notice that there is still a time difference, but not in orders of magnitude. Also note the importance of choosing the axis of multiplication. Now perhaps, a numpy developer can shed some light on what numpy.dot actually does under the hood for N-D arrays.

Pairwise operations (distance) on two lists in numpy

I have two lists of coordinates:
l1 = [[x,y,z],[x,y,z],[x,y,z],[x,y,z],[x,y,z]]
l2 = [[x,y,z],[x,y,z],[x,y,z]]
I want to find the shortest pairwise distance between l1 and l2. Distance between two coordinates is simply:
numpy.linalg.norm(l1_element - l2_element)
So how do I use numpy to efficiently apply this operation to each pair of elements?
Here is a quick performance analysis of the four methods presented so far:
import numpy
import scipy
from itertools import product
from scipy.spatial.distance import cdist
from scipy.spatial import cKDTree as KDTree
n = 100
l1 = numpy.random.randint(0, 100, size=(n,3))
l2 = numpy.random.randint(0, 100, size=(n,3))
# by #Phillip
def a(l1,l2):
return min(numpy.linalg.norm(l1_element - l2_element) for l1_element,l2_element in product(l1,l2))
# by #Kasra
def b(l1,l2):
return numpy.min(numpy.apply_along_axis(
numpy.linalg.norm,
2,
l1[:, None, :] - l2[None, :, :]
))
# mine
def c(l1,l2):
return numpy.min(scipy.spatial.distance.cdist(l1,l2))
# just checking that numpy.min is indeed faster.
def c2(l1,l2):
return min(scipy.spatial.distance.cdist(l1,l2).reshape(-1))
# by #BrianLarsen
def d(l1,l2):
# make KDTrees for both sets of points
t1 = KDTree(l1)
t2 = KDTree(l2)
# we need a distance to not look beyond, if you have real knowledge use it, otherwise guess
maxD = numpy.linalg.norm(l1[0] - l2[0]) # this could be closest but anyhting further is certainly not
# get a sparce matrix of all the distances
ans = t1.sparse_distance_matrix(t2, maxD)
# get the minimum distance and points involved
minD = min(ans.values())
return minD
for x in (a,b,c,c2,d):
print("Timing variant", x.__name__, ':', flush=True)
print(x(l1,l2), flush=True)
%timeit x(l1,l2)
print(flush=True)
For n=100
Timing variant a :
2.2360679775
10 loops, best of 3: 90.3 ms per loop
Timing variant b :
2.2360679775
10 loops, best of 3: 151 ms per loop
Timing variant c :
2.2360679775
10000 loops, best of 3: 136 µs per loop
Timing variant c2 :
2.2360679775
1000 loops, best of 3: 844 µs per loop
Timing variant d :
2.2360679775
100 loops, best of 3: 3.62 ms per loop
For n=1000
Timing variant a :
0.0
1 loops, best of 3: 9.16 s per loop
Timing variant b :
0.0
1 loops, best of 3: 14.9 s per loop
Timing variant c :
0.0
100 loops, best of 3: 11 ms per loop
Timing variant c2 :
0.0
10 loops, best of 3: 80.3 ms per loop
Timing variant d :
0.0
1 loops, best of 3: 933 ms per loop
Using newaxis and broadcasting, l1[:, None, :] - l2[None, :, :] is an array of the pairwise difference vectors. You can reduce this array to an array of norms using apply_along_axis and then take the min:
numpy.min(numpy.apply_along_axis(
numpy.linalg.norm,
2,
l1[:, None, :] - l2[None, :, :]
))
Of course, this only works if l1 and l2 are numpy arrays, so if your lists in the question weren't pseudo-code, you'll have to add l1 = numpy.array(l1); l2 = numpy.array(l2).
You can use itertools.product to get the all combinations the use min :
l1 = [[x,y,z],[x,y,z],[x,y,z],[x,y,z],[x,y,z]]
l2 = [[x,y,z],[x,y,z],[x,y,z]]
from itertools import product
min(numpy.linalg.norm(l1_element - l2_element) for l1_element,l2_element in product(l1,l2))
If you have many, many, many points this is a great use for a KDTree. Totally overkill for this example but a good learning experience and really fast for a certain class of problems, and can give a bit more information on number of points within a certain distance.
import numpy as np
from scipy.spatial import cKDTree as KDTree
#sample data
l1 = [[0,0,0], [4,5,6], [7,6,7], [4,5,6]]
l2 = [[100,3,4], [1,0,0], [10,15,16], [17,16,17], [14,15,16], [-34, 5, 6]]
# make them arrays
l1 = np.asarray(l1)
l2 = np.asarray(l2)
# make KDTrees for both sets of points
t1 = KDTree(l1)
t2 = KDTree(l2)
# we need a distance to not look beyond, if you have real knowledge use it, otherwise guess
maxD = np.linalg.norm(l1[-1] - l2[-1]) # this could be closest but anyhting further is certainly not
# get a sparce matrix of all the distances
ans = t1.sparse_distance_matrix(t2, maxD)
# get the minimum distance and points involved
minA = min([(i,k) for k, i in ans.iteritems()])
print("Minimun distance is {0} between l1={1} and l2={2}".format(minA[0], l1[minA[1][0]], l2[minA[1][2]] ))
What this does is make a KDTree for the the sets of points then find all the distances for points within the guess distance and give back the distance and closest point. This post has a writeup of how a KDTree works.

Categories