I have a mXn numpy array called a:
I would like to write a function which returns an array with size (3, mxn) which contains for each couple (x,y) in the first array the correspondant value.
import numpy as np
m=5
n=10
a = np.random.random((m, n))
x = np.random.random((m, 1)) # x coordinates
y = np.random.random((1, n)) # y coordinates
b = np.empty((3, m*n)) # array to store coordinates
k=0
for i in range (0,m):
for j in range (0,n):
b[0,k] = a[i,0]
b[1,k] = a[0,j]
b[2,k] = a[i,j]
k=k+1
This seems to run ok, but is there a faster or better coded way to do this?
Steps :
Initialize a 3D array, such that m and n are separate ones. This lets us broadcast values.
Index into the first three elements along the first axis of output with appropriate elements off a and make sure that those shapes are broadcastable.
Reshape the output back to 2D.
That's all the play is about here! Here's the vectorized implementation -
b_out = np.empty((3, m,n),dtype=a.dtype) # 1. Initialize
b_out[0] = a[:,0,None] # 2. Assign
b_out[1] = a[0]
b_out[2] = a
b_out.shape = (3,m*n) # 3. Reshape back to 2D
Runtime test
Approaches -
def loopy_app(a):
m,n = a.shape
b = np.empty((3, m*n),dtype=a.dtype)
k=0
for i in range (0,m):
for j in range (0,n):
b[0,k] = a[i,0]
b[1,k] = a[0,j]
b[2,k] = a[i,j]
k=k+1
return b
def vectorized_app(a):
b_out = np.empty((3, m,n),dtype=a.dtype)
b_out[0] = a[:,0,None]
b_out[1] = a[0]
b_out[2] = a
b_out.shape = (3,m*n)
return b_out
Timings -
In [194]: m=5
...: n=10
...: a = np.random.random((m, n))
...:
In [195]: %timeit loopy_app(a)
...: %timeit vectorized_app(a)
...:
10000 loops, best of 3: 28.2 µs per loop
100000 loops, best of 3: 2.48 µs per loop
In [196]: m=50
...: n=100
...: a = np.random.random((m, n))
...:
In [197]: %timeit loopy_app(a)
...: %timeit vectorized_app(a)
...:
100 loops, best of 3: 2.56 ms per loop
100000 loops, best of 3: 6.31 µs per loop
In [198]: 2560/6.31
Out[198]: 405.7052297939778
400x+ speedup on large datasets and more on larger ones!
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] )
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
I have a tensor U composed of n matrices of dimension (d,k) and a matrix V of dimension (k,n).
I would like to multiply them so that the result returns a matrix of dimension (d,n) in which column j is the result of the matrix multiplication between the matrix j of U and the column j of V.
One possible way to obtain this is:
for j in range(n):
res[:,j] = U[:,:,j] * V[:,j]
I am wondering if there is a faster approach using numpy library. In particular I'm thinking of the np.tensordot() function.
This small snippet allows me to multiply a single matrix by a scalar, but the obvious generalization to a vector is not returning what I was hoping for.
a = np.array(range(1, 17))
a.shape = (4,4)
b = np.array((1,2,3,4,5,6,7))
r1 = np.tensordot(b,a, axes=0)
Any suggestion?
There are a couple of ways you could do this. The first thing that comes to mind is np.einsum:
# some fake data
gen = np.random.RandomState(0)
ni, nj, nk = 10, 20, 100
U = gen.randn(ni, nj, nk)
V = gen.randn(nj, nk)
res1 = np.zeros((ni, nk))
for k in range(nk):
res1[:,k] = U[:,:,k].dot(V[:,k])
res2 = np.einsum('ijk,jk->ik', U, V)
print(np.allclose(res1, res2))
# True
np.einsum uses Einstein notation to express tensor contractions. In the expression 'ijk,jk->ik' above, i,j and k are subscripts that correspond to the different dimensions of U and V. Each comma-separated grouping corresponds to one of the operands passed to np.einsum (in this case U has dimensions ijk and V has dimensions jk). The '->ik' part specifies the dimensions of the output array. Any dimensions with subscripts that aren't present in the output string are summed over.
np.einsum is incredibly useful for performing complex tensor contractions, but it can take a while to fully wrap your head around how it works. You should take a look at the examples in the documentation (linked above).
Some other options:
Element-wise multiplication with broadcasting, followed by summation:
res3 = (U * V[None, ...]).sum(1)
inner1d with a load of transposing:
from numpy.core.umath_tests import inner1d
res4 = inner1d(U.transpose(0, 2, 1), V.T)
Some benchmarks:
In [1]: ni, nj, nk = 100, 200, 1000
In [2]: %%timeit U = gen.randn(ni, nj, nk); V = gen.randn(nj, nk)
....: np.einsum('ijk,jk->ik', U, V)
....:
10 loops, best of 3: 23.4 ms per loop
In [3]: %%timeit U = gen.randn(ni, nj, nk); V = gen.randn(nj, nk)
(U * V[None, ...]).sum(1)
....:
10 loops, best of 3: 59.7 ms per loop
In [4]: %%timeit U = gen.randn(ni, nj, nk); V = gen.randn(nj, nk)
inner1d(U.transpose(0, 2, 1), V.T)
....:
10 loops, best of 3: 45.9 ms per loop
I have a 2D matrix and I want to take norm of each row. But when I use numpy.linalg.norm(X) directly, it takes the norm of the whole matrix.
I can take norm of each row by using a for loop and then taking norm of each X[i], but it takes a huge time since I have 30k rows.
Any suggestions to find a quicker way? Or is it possible to apply np.linalg.norm to each row of a matrix?
For numpy 1.9+
Note that, as perimosocordiae shows, as of NumPy version 1.9, np.linalg.norm(x, axis=1) is the fastest way to compute the L2-norm.
For numpy < 1.9
If you are computing an L2-norm, you could compute it directly (using the axis=-1 argument to sum along rows):
np.sum(np.abs(x)**2,axis=-1)**(1./2)
Lp-norms can be computed similarly of course.
It is considerably faster than np.apply_along_axis, though perhaps not as convenient:
In [48]: %timeit np.apply_along_axis(np.linalg.norm, 1, x)
1000 loops, best of 3: 208 us per loop
In [49]: %timeit np.sum(np.abs(x)**2,axis=-1)**(1./2)
100000 loops, best of 3: 18.3 us per loop
Other ord forms of norm can be computed directly too (with similar speedups):
In [55]: %timeit np.apply_along_axis(lambda row:np.linalg.norm(row,ord=1), 1, x)
1000 loops, best of 3: 203 us per loop
In [54]: %timeit np.sum(abs(x), axis=-1)
100000 loops, best of 3: 10.9 us per loop
Resurrecting an old question due to a numpy update. As of the 1.9 release, numpy.linalg.norm now accepts an axis argument. [code, documentation]
This is the new fastest method in town:
In [10]: x = np.random.random((500,500))
In [11]: %timeit np.apply_along_axis(np.linalg.norm, 1, x)
10 loops, best of 3: 21 ms per loop
In [12]: %timeit np.sum(np.abs(x)**2,axis=-1)**(1./2)
100 loops, best of 3: 2.6 ms per loop
In [13]: %timeit np.linalg.norm(x, axis=1)
1000 loops, best of 3: 1.4 ms per loop
And to prove it's calculating the same thing:
In [14]: np.allclose(np.linalg.norm(x, axis=1), np.sum(np.abs(x)**2,axis=-1)**(1./2))
Out[14]: True
Much faster than the accepted answer is using NumPy's einsum,
numpy.sqrt(numpy.einsum('ij,ij->i', a, a))
And even faster than that is arranging the data such that the norms are computed across all columns,
numpy.sqrt(numpy.einsum('ij,ij->j', aT, aT))
Note the log-scale:
Code to reproduce the plot:
import numpy as np
import perfplot
rng = np.random.default_rng(0)
def setup(n):
x = rng.random((n, 3))
xt = np.ascontiguousarray(x.T)
return x, xt
def sum_sqrt(a, _):
return np.sqrt(np.sum(np.abs(a) ** 2, axis=-1))
def apply_norm_along_axis(a, _):
return np.apply_along_axis(np.linalg.norm, 1, a)
def norm_axis(a, _):
return np.linalg.norm(a, axis=1)
def einsum_sqrt(a, _):
return np.sqrt(np.einsum("ij,ij->i", a, a))
def einsum_sqrt_columns(_, aT):
return np.sqrt(np.einsum("ij,ij->j", aT, aT))
b = perfplot.bench(
setup=setup,
kernels=[
sum_sqrt,
apply_norm_along_axis,
norm_axis,
einsum_sqrt,
einsum_sqrt_columns,
],
n_range=[2**k for k in range(20)],
xlabel="len(a)",
)
b.show()
b.save("out.png")
Try the following:
In [16]: numpy.apply_along_axis(numpy.linalg.norm, 1, a)
Out[16]: array([ 5.38516481, 1.41421356, 5.38516481])
where a is your 2D array.
The above computes the L2 norm. For a different norm, you could use something like:
In [22]: numpy.apply_along_axis(lambda row:numpy.linalg.norm(row,ord=1), 1, a)
Out[22]: array([9, 2, 9])