How to append to each element in the tensor without a loop?-CodePudding

How to manipulate each element in a tensor?
For example, a = [[1,1,1], [2,2,2], [3,3,3]] and b = [0,0,0]. Here I want to connect b before each element of a, which will become [[0,0,0,1,1,1], [0,0,0,2,2,2], [0,0,0,3,3,3]], is there a way to implement this without a loop since using loop is pretty slow in python while dealing with large data.
Many thanks.

CodePudding user response：

import numpy as np
import torch

a = np.array([[1,1,1], [2,2,2], [3,3,3]])
b = np.array([0, 0, 0])

result = np.array([np.array((*b, *row)) for row in a])

# Converting to torch
result = torch.from_numpy(result)

Result:

array([[0, 0, 0, 1, 1, 1],
       [0, 0, 0, 2, 2, 2],
       [0, 0, 0, 3, 3, 3]])

CodePudding user response：

You can use map():

a = [[1,1,1], [2,2,2], [3,3,3]]
b = [0,0,0]

result = list(map(lambda x: b   x, a))

# Prints [[0, 0, 0, 1, 1, 1], [0, 0, 0, 2, 2, 2], [0, 0, 0, 3, 3, 3]]
print(result)

CodePudding user response：

In [185]: a = [[1, 1, 1], [2, 2, 2], [3, 3, 3]]
     ...: b = [0, 0, 0]

The obvious list comprenhensin:

In [186]: [b   x for x in a]
Out[186]: [[0, 0, 0, 1, 1, 1], [0, 0, 0, 2, 2, 2], [0, 0, 0, 3, 3, 3]]

And array approach:

In [187]: np.concatenate((np.zeros((len(a), len(a[0])), int), np.array(a)), axis=1)
Out[187]: 
array([[0, 0, 0, 1, 1, 1],
       [0, 0, 0, 2, 2, 2],
       [0, 0, 0, 3, 3, 3]])

for this small example the list approach is better:

In [188]: timeit [b   x for x in a]
528 ns ± 12.7 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
In [189]: timeit np.concatenate((np.zeros((len(a), len(a[0])), int), np.array(a)), axis=1)
9.77 µs ± 95.4 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

If A is already an array:

In [192]: %%timeit A = np.array(a)
     ...:  np.concatenate((np.zeros(A.shape,A.dtype), A), axis=1)
6.26 µs ± 223 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

The last will scalar much better.

And the map alternative:

In [194]: timeit list(map(lambda x: b   x, a))
856 ns ± 31.6 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)