I'm trying to calculate this type of calculation:
arr = np.arange(4)
# array([0, 1, 2, 3])
arr_t =arr.reshape((-1,1))
# array([[0],
# [1],
# [2],
# [3]])
mult_arr = np.multiply(arr,arr_t) # <<< the multiplication
# array([[0, 0, 0, 0],
# [0, 1, 2, 3],
# [0, 2, 4, 6],
# [0, 3, 6, 9]])
to eventually perform it in a bigger matrix index of single row, and to sum all the matrices that are reproduced by the calculation:
arr = np.random.random((600,150))
arr_t =arr.reshape((-1,arr.shape[1],1))
mult = np.multiply(arr[:,None],arr_t)
summed = np.sum(mult,axis=0)
summed
Till now its all pure awesomeness, the problem starts when I try to covert on a bigger dataset, for example this array instead :
arr = np.random.random((6000,1500))
I get the following error - MemoryError: Unable to allocate 101. GiB for an array with shape (6000, 1500, 1500) and data type float64
which make sense, but my question is:
can I get around this anyhow without being forced to use loops that slow down the process entirely ??
my question is mainly about performance and solution that require long running tasks more then 30 secs is not an option.
Looks like you are simply trying to perform a dot product:
arr.T#arr
or
arr.T.dot(arr)
checking this is what you want
arr = np.random.random((600,150))
arr_t =arr.reshape((-1,arr.shape[1],1))
mult = np.multiply(arr[:,None],arr_t)
summed = np.sum(mult,axis=0)
np.allclose((arr.T#arr), summed)
# True
Let's say I have a two-dimensional array
import numpy as np
a = np.array([[1, 1, 1], [2,2,2], [3,3,3]])
and I would like to replace the third vector (in the second dimension) with zeros. I would do
a[:, 2] = np.array([0, 0, 0])
But what if I would like to be able to do that programmatically? I mean, let's say that variable x = 1 contained the dimension on which I wanted to do the replacing. How would the function replace(arr, dimension, value, arr_to_be_replaced) have to look if I wanted to call it as replace(a, x, 2, np.array([0, 0, 0])?
numpy has a similar function, insert. However, it doesn't replace at dimension i, it returns a copy with an additional vector.
All solutions are welcome, but I do prefer a solution that doesn't recreate the array as to save memory.
arr[:, 1]
is basically shorthand for
arr[(slice(None), 1)]
that is, a tuple with slice elements and integers.
Knowing that, you can construct a tuple of slice objects manually, adjust the values depending on an axis parameter and use that as your index. So for
import numpy as np
arr = np.array([[1, 1, 1], [2, 2, 2], [3, 3, 3]])
axis = 1
idx = 2
arr[:, idx] = np.array([0, 0, 0])
# ^- axis position
you can use
slices = [slice(None)] * arr.ndim
slices[axis] = idx
arr[tuple(slices)] = np.array([0, 0, 0])
I have 2d numpy array (think greyscale image). I want to assign certain value to a list of coordinates to this array, such that:
img = np.zeros((5, 5))
coords = np.array([[0, 1], [1, 2], [2, 3], [3, 4]])
def bad_use_of_numpy(img, coords):
for i, coord in enumerate(coords):
img[coord[0], coord[1]] = 255
return img
bad_use_of_numpy(img, coords)
This works, but I feel like I can take advantage of numpy functionality to make it faster. I also might have a use case later to to something like following:
img = np.zeros((5, 5))
coords = np.array([[0, 1], [1, 2], [2, 3], [3, 4]])
vals = np.array([1, 2, 3, 4])
def bad_use_of_numpy(img, coords, vals):
for coord in coords:
img[coord[0], coord[1]] = vals[i]
return img
bad_use_of_numpy(img, coords, vals)
Is there a more vectorized way of doing that?
We can unpack each row of coords as row, col indices for indexing into img and then assign.
Now, since the question is tagged : Python 3.x, on it we can simply unpack with [*coords.T] and then assign -
img[[*coords.T]] = 255
Generically, we can use tuple to unpack -
img[tuple(coords.T)] = 255
We can also compute the linear indices and then assign with np.put -
np.put(img, np.ravel_multi_index(coords.T, img.shape), 255)
I want to get the intersecting (common) rows across two 2D numpy arrays. E.g., if the following arrays are passed as inputs:
array([[1, 4],
[2, 5],
[3, 6]])
array([[1, 4],
[3, 6],
[7, 8]])
the output should be:
array([[1, 4],
[3, 6])
I know how to do this with loops. I'm looking at a Pythonic/Numpy way to do this.
For short arrays, using sets is probably the clearest and most readable way to do it.
Another way is to use numpy.intersect1d. You'll have to trick it into treating the rows as a single value, though... This makes things a bit less readable...
import numpy as np
A = np.array([[1,4],[2,5],[3,6]])
B = np.array([[1,4],[3,6],[7,8]])
nrows, ncols = A.shape
dtype={'names':['f{}'.format(i) for i in range(ncols)],
'formats':ncols * [A.dtype]}
C = np.intersect1d(A.view(dtype), B.view(dtype))
# This last bit is optional if you're okay with "C" being a structured array...
C = C.view(A.dtype).reshape(-1, ncols)
For large arrays, this should be considerably faster than using sets.
You could use Python's sets:
>>> import numpy as np
>>> A = np.array([[1,4],[2,5],[3,6]])
>>> B = np.array([[1,4],[3,6],[7,8]])
>>> aset = set([tuple(x) for x in A])
>>> bset = set([tuple(x) for x in B])
>>> np.array([x for x in aset & bset])
array([[1, 4],
[3, 6]])
As Rob Cowie points out, this can be done more concisely as
np.array([x for x in set(tuple(x) for x in A) & set(tuple(x) for x in B)])
There's probably a way to do this without all the going back and forth from arrays to tuples, but it's not coming to me right now.
I could not understand why there is no suggested pure numpy way to get this working. So I found one, that uses numpy broadcast. The basic idea is to transform one of the arrays to 3d by axes swapping. Let's construct 2 arrays:
a=np.random.randint(10, size=(5, 3))
b=np.zeros_like(a)
b[:4,:]=a[np.random.randint(a.shape[0], size=4), :]
With my run it gave:
a=array([[5, 6, 3],
[8, 1, 0],
[2, 1, 4],
[8, 0, 6],
[6, 7, 6]])
b=array([[2, 1, 4],
[2, 1, 4],
[6, 7, 6],
[5, 6, 3],
[0, 0, 0]])
The steps are (arrays can be interchanged) :
#a is nxm and b is kxm
c = np.swapaxes(a[:,:,None],1,2)==b #transform a to nx1xm
# c has nxkxm dimensions due to comparison broadcast
# each nxixj slice holds comparison matrix between a[j,:] and b[i,:]
# Decrease dimension to nxk with product:
c = np.prod(c,axis=2)
#To get around duplicates://
# Calculate cumulative sum in k-th dimension
c= c*np.cumsum(c,axis=0)
# compare with 1, so that to get only one 'True' statement by row
c=c==1
#//
# sum in k-th dimension, so that a nx1 vector is produced
c=np.sum(c,axis=1).astype(bool)
# The intersection between a and b is a[c]
result=a[c]
In a function with 2 lines for used memory reduction (correct me if wrong):
def array_row_intersection(a,b):
tmp=np.prod(np.swapaxes(a[:,:,None],1,2)==b,axis=2)
return a[np.sum(np.cumsum(tmp,axis=0)*tmp==1,axis=1).astype(bool)]
which gave result for my example:
result=array([[5, 6, 3],
[2, 1, 4],
[6, 7, 6]])
This is faster than set solutions, as it makes use only of simple numpy operations, while it reduces constantly dimensions, and is ideal for two big matrices. I guess I might have made mistakes in my comments, as I got the answer by experimentation and instinct. The equivalent for column intersection can either be found by transposing the arrays or by changing the steps a little. Also, if duplicates are wanted, then the steps inside "//" have to be skipped. The function can be edited to return only the boolean array of the indices, which came handy to me ,while trying to get different arrays indices with the same vector. Benchmark for the voted answer and mine (number of elements in each dimension plays role on what to choose):
Code:
def voted_answer(A,B):
nrows, ncols = A.shape
dtype={'names':['f{}'.format(i) for i in range(ncols)],
'formats':ncols * [A.dtype]}
C = np.intersect1d(A.view(dtype), B.view(dtype))
return C.view(A.dtype).reshape(-1, ncols)
a_small=np.random.randint(10, size=(10, 10))
b_small=np.zeros_like(a_small)
b_small=a_small[np.random.randint(a_small.shape[0],size=[a_small.shape[0]]),:]
a_big_row=np.random.randint(10, size=(10, 1000))
b_big_row=a_big_row[np.random.randint(a_big_row.shape[0],size=[a_big_row.shape[0]]),:]
a_big_col=np.random.randint(10, size=(1000, 10))
b_big_col=a_big_col[np.random.randint(a_big_col.shape[0],size=[a_big_col.shape[0]]),:]
a_big_all=np.random.randint(10, size=(100,100))
b_big_all=a_big_all[np.random.randint(a_big_all.shape[0],size=[a_big_all.shape[0]]),:]
print 'Small arrays:'
print '\t Voted answer:',timeit.timeit(lambda:voted_answer(a_small,b_small),number=100)/100
print '\t Proposed answer:',timeit.timeit(lambda:array_row_intersection(a_small,b_small),number=100)/100
print 'Big column arrays:'
print '\t Voted answer:',timeit.timeit(lambda:voted_answer(a_big_col,b_big_col),number=100)/100
print '\t Proposed answer:',timeit.timeit(lambda:array_row_intersection(a_big_col,b_big_col),number=100)/100
print 'Big row arrays:'
print '\t Voted answer:',timeit.timeit(lambda:voted_answer(a_big_row,b_big_row),number=100)/100
print '\t Proposed answer:',timeit.timeit(lambda:array_row_intersection(a_big_row,b_big_row),number=100)/100
print 'Big arrays:'
print '\t Voted answer:',timeit.timeit(lambda:voted_answer(a_big_all,b_big_all),number=100)/100
print '\t Proposed answer:',timeit.timeit(lambda:array_row_intersection(a_big_all,b_big_all),number=100)/100
with results:
Small arrays:
Voted answer: 7.47108459473e-05
Proposed answer: 2.47001647949e-05
Big column arrays:
Voted answer: 0.00198730945587
Proposed answer: 0.0560171294212
Big row arrays:
Voted answer: 0.00500325918198
Proposed answer: 0.000308241844177
Big arrays:
Voted answer: 0.000864889621735
Proposed answer: 0.00257176160812
Following verdict is that if you have to compare 2 big 2d arrays of 2d points then use voted answer. If you have big matrices in all dimensions, voted answer is the best one by all means. So, it depends on what you choose each time.
Numpy broadcasting
We can create a boolean mask using broadcasting which can be then used to filter the rows in array A which are also present in array B
A = np.array([[1,4],[2,5],[3,6]])
B = np.array([[1,4],[3,6],[7,8]])
m = (A[:, None] == B).all(-1).any(1)
>>> A[m]
array([[1, 4],
[3, 6]])
Another way to achieve this using structured array:
>>> a = np.array([[3, 1, 2], [5, 8, 9], [7, 4, 3]])
>>> b = np.array([[2, 3, 0], [3, 1, 2], [7, 4, 3]])
>>> av = a.view([('', a.dtype)] * a.shape[1]).ravel()
>>> bv = b.view([('', b.dtype)] * b.shape[1]).ravel()
>>> np.intersect1d(av, bv).view(a.dtype).reshape(-1, a.shape[1])
array([[3, 1, 2],
[7, 4, 3]])
Just for clarity, the structured view looks like this:
>>> a.view([('', a.dtype)] * a.shape[1])
array([[(3, 1, 2)],
[(5, 8, 9)],
[(7, 4, 3)]],
dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8')])
np.array(set(map(tuple, b)).difference(set(map(tuple, a))))
This could also work
Without Index
Visit https://gist.github.com/RashidLadj/971c7235ce796836853fcf55b4876f3c
def intersect2D(Array_A, Array_B):
"""
Find row intersection between 2D numpy arrays, a and b.
"""
# ''' Using Tuple ''' #
intersectionList = list(set([tuple(x) for x in Array_A for y in Array_B if(tuple(x) == tuple(y))]))
print ("intersectionList = \n",intersectionList)
# ''' Using Numpy function "array_equal" ''' #
""" This method is valid for an ndarray """
intersectionList = list(set([tuple(x) for x in Array_A for y in Array_B if(np.array_equal(x, y))]))
print ("intersectionList = \n",intersectionList)
# ''' Using set and bitwise and '''
intersectionList = [list(y) for y in (set([tuple(x) for x in Array_A]) & set([tuple(x) for x in Array_B]))]
print ("intersectionList = \n",intersectionList)
return intersectionList
With Index
Visit https://gist.github.com/RashidLadj/bac71f3d3380064de2f9abe0ae43c19e
def intersect2D(Array_A, Array_B):
"""
Find row intersection between 2D numpy arrays, a and b.
Returns another numpy array with shared rows and index of items in A & B arrays
"""
# [[IDX], [IDY], [value]] where Equal
# ''' Using Tuple ''' #
IndexEqual = np.asarray([(i, j, x) for i,x in enumerate(Array_A) for j, y in enumerate (Array_B) if(tuple(x) == tuple(y))]).T
# ''' Using Numpy array_equal ''' #
IndexEqual = np.asarray([(i, j, x) for i,x in enumerate(Array_A) for j, y in enumerate (Array_B) if(np.array_equal(x, y))]).T
idx, idy, intersectionList = (IndexEqual[0], IndexEqual[1], IndexEqual[2]) if len(IndexEqual) != 0 else ([], [], [])
return intersectionList, idx, idy
A = np.array([[1,4],[2,5],[3,6]])
B = np.array([[1,4],[3,6],[7,8]])
def matching_rows(A,B):
matches=[i for i in range(B.shape[0]) if np.any(np.all(A==B[i],axis=1))]
if len(matches)==0:
return B[matches]
return np.unique(B[matches],axis=0)
>>> matching_rows(A,B)
array([[1, 4],
[3, 6]])
This of course assumes the rows are all the same length.
import numpy as np
A=np.array([[1, 4],
[2, 5],
[3, 6]])
B=np.array([[1, 4],
[3, 6],
[7, 8]])
intersetingRows=[(B==irow).all(axis=1).any() for irow in A]
print(A[intersetingRows])
An example
Suppose I have a tensor values with shape (2,2,2)
values = [[[0, 1],[2, 3]],[[4, 5],[6, 7]]]
And a tensor indicies with shape (2,2) which describes what values to be selected in the innermost dimension
indicies = [[1,0],[0,0]]
Then the result will be a (2,2) matrix with these values
result = [[1,2],[4,6]]
What is this operation called in tensorflow and how to do it?
General
Note that the above shape (2,2,2) is only an example, it can be any dimension. Some conditions for this operation:
ndim(values) -1 = ndim(indicies)
values.shape[:-1] == indicies.shape == result.shape
indicies.max() < values.shape[-1] -1
I think you can emulate this with tf.gather_nd. You will just have to convert "your" indices to a representation that is suitable for tf.gather_nd. The following example here is tied to your specific example, i.e. input tensors of shape (2, 2, 2) but I think this gives you an idea how you could write the conversion for input tensors with arbitrary shape, although I am not sure how easy it would be to implement this (haven't thought about it too long). Also, I'm not claiming that this is the easiest possible solution.
import tensorflow as tf
import numpy as np
values = np.array([[[0, 1], [2, 3]], [[4, 5], [6, 7]]])
values_tf = tf.constant(values)
indices = np.array([[1, 0], [0, 0]])
converted_idx = []
for k in range(values.shape[0]):
outer = []
for l in range(values.shape[1]):
inds = [k, l, indices[k][l]]
outer.append(inds)
print(inds)
converted_idx.append(outer)
with tf.Session() as sess:
result = tf.gather_nd(values_tf, converted_idx)
print(sess.run(result))
This prints
[[1 2]
[4 6]]
Edit: To handle arbitrary shapes here is a recursive solution that should work (only tested on your example):
def convert_idx(last_dim_vals, ori_indices, access_to_ori, depth):
if depth == len(last_dim_vals.shape) - 1:
inds = access_to_ori + [ori_indices[tuple(access_to_ori)]]
return inds
outer = []
for k in range(ori_indices.shape[depth]):
inds = convert_idx(last_dim_vals, ori_indices, access_to_ori + [k], depth + 1)
outer.append(inds)
return outer
You can use this together with the original code I posted like so:
...
converted_idx = convert_idx(values, indices, [], 0)
with tf.Session() as sess:
result = tf.gather_nd(values_tf, converted_idx)
print(sess.run(result))