>>> b
tensor([[ 6, 7, 12, 7, 8],
[ 0, 1, 6, 1, 2],
[ 0, 1, 6, 1, 2],
[ 2, 3, 8, 3, 4],
[ 2, 3, 8, 3, 4],
[ 2, 3, 8, 3, 4],
[10, 11, 16, 11, 12],
[-1, 0, 5, 0, 1],
[-2, -1, 4, -1, 0],
[ 2, 3, 8, 3, 4],
[ 1, 2, 7, 2, 3],
[ 1, 2, 7, 2, 3],
[ 2, 3, 8, 3, 4],
[ 5, 6, 11, 6, 7],
[-2, -1, 4, -1, 0],
[-3, -2, 3, -2, -1],
[-5, -4, 1, -4, -3],
[ 1, 2, 7, 2, 3],
[12, 13, 18, 13, 14],
[-3, -2, 3, -2, -1],
[ 2, 3, 8, 3, 4],
[ 3, 4, 9, 4, 5],
[10, 11, 16, 11, 12],
[-6, -5, 0, -5, -4],
[ 9, 10, 15, 10, 11],
[12, 13, 18, 13, 14],
[-3, -2, 3, -2, -1],
[-2, -1, 4, -1, 0],
[-4, -3, 2, -3, -2],
[-1, 0, 5, 0, 1],
[ 2, 3, 8, 3, 4],
[ 4, 5, 10, 5, 6],
[-1, 0, 5, 0, 1],
[ 5, 6, 11, 6, 7],
[ 7, 8, 13, 8, 9],
[ 3, 4, 9, 4, 5],
[ 2, 3, 8, 3, 4],
[ 4, 5, 10, 5, 6],
[-4, -3, 2, -3, -2],
[ 2, 3, 8, 3, 4],
[-1, 0, 5, 0, 1],
[ 2, 3, 8, 3, 4],
[ 4, 5, 10, 5, 6],
[ 9, 10, 15, 10, 11],
[-1, 0, 5, 0, 1],
[-4, -3, 2, -3, -2],
[ 0, 1, 6, 1, 2],
[ 4, 5, 10, 5, 6],
[ 6, 7, 12, 7, 8],
[-2, -1, 4, -1, 0]])
>>> torch.mode(b, 0)
torch.return_types.mode(
values=tensor([2, 3, 8, 3, 4]),
indices=tensor([20, 20, 20, 20, 20]))
i don't know why indeces is all equal to 20
the details of torch.mode description as below
https://pytorch.org/docs/stable/generated/torch.mode.html#torch.mode
torch.mode(input, dim=- 1, keepdim=False, *, out=None)
Returns a namedtuple (values, indices) where values is the mode value of each row of the input tensor in the given dimension dim, i.e. a value which appears most often in that row, and indices is the index location of each mode value found.
By default, dim is the last dimension of the input tensor.
If keepdim is True, the output tensors are of the same size as input except in the dimension dim where they are of size 1. Otherwise, dim is squeezed (see torch.squeeze()), resulting in the output tensors having 1 fewer dimension than input.
It is because of the way the tensor b is. The row [2, 3, 8, 3, 4] is repeated a lot, so in each column the modes are respectively [2, 3, 8, 3, 4] and more importantly, the mode indices will be equal precisely because the modes occur together; if you look at the row with index 20 (i.e., the 21st row), it is exactly [2, 3, 8, 3, 4].
I am assuming that you constructed b similar to the example in torch.mode which I believe is a poor choice for an example as it leads to confusion like the one you are having.
Instead, consider the following:
>>> b = torch.randint(4, (5, 7))
>>> b
tensor([[0, 0, 0, 2, 0, 0, 2],
[0, 3, 0, 0, 2, 0, 1],
[2, 2, 2, 0, 0, 0, 3],
[2, 2, 3, 0, 1, 1, 0],
[1, 1, 0, 0, 2, 0, 2]])
>>> torch.mode(b, 0)
torch.return_types.mode(
values=tensor([0, 2, 0, 0, 0, 0, 2]),
indices=tensor([1, 3, 4, 4, 2, 4, 4]))
In the above, b has different modes in each column which are respectively [0, 2, 0, 0, 0, 0, 2] and the indices returned by torch.mode are [1, 3, 4, 4, 2, 4, 4]. This makes sense because, for example, in the first column, 0 is the most common element and there is a 0 at index 1. Similarly, in the second column, 2 is the most common element and there is a 2 at index 3. This is true for all columns. If you want the modes of the rows instead, you would do torch.mode(b, 1).
Related
I have the indices of a 2D array. I want to partition the indices such that the corresponding entries form blocks (block size is given as input m and n).
For example, if the indices are as given below
(array([0, 0, 0, 0, 1, 1, 1, 1, 6, 6, 6, 6, 7, 7, 7, 7 ]), array([0, 1, 7, 8, 0,1,7,8, 0,1,7,8, 0, 1, 7, 8]))
for the original matrix (from which the indices are generated)
array([[3, 4, 2, 0, 1, 1, 0, 2, 4],
[1, 3, 2, 0, 0, 1, 0, 4, 0],
[1, 0, 0, 1, 1, 0, 1, 1, 3],
[0, 0, 0, 3, 3, 0, 4, 0, 4],
[4, 3, 4, 2, 1, 1, 0, 0, 4],
[0, 1, 0, 4, 4, 2, 2, 2, 1],
[2, 4, 0, 1, 1, 0, 0, 2, 1],
[0, 4, 1, 3, 3, 2, 3, 2, 4]])
and if the block size is (2,2), then the blocks should be
[[3, 4],
[1, 3]]
[[2, 4]
[4, 0]]
[[2, 4]
[0, 4]]
[[2, 1]
[2, 4]]
Any help to do it efficiently?
Does this help? A is your matrix.
row_size = 2
col_size = 3
for i in range(A.shape[0] // row_size):
for j in range(A.shape[1] // col_size):
print(A[row_size*i:row_size*i + row_size, col_size*j:col_size*j + col_size])
Let's say I have defined a function to generate a list of 6 random integers ranging from 0 to 10
import random
def func():
randomlist = random.sample(range(11), 6)
return randomlist
Run:
func()
Output:
[3, 7, 10, 9, 4, 1]
Now I want to define another function which calls func() inside with a parameter n to generate a multidimensional matrix, where each element will be replaced by a newly created list generated by func(), and n is the times of the completed replacement (so that the total number of integers in the matrix would be 6^n) -- for example --
when n=1, expected result:
[3, 7, 10, 9, 4, 1]
when n=2, expected result:
[[6, 2, 9, 1, 4, 0],
[7, 8, 1, 9, 4, 1],
[1, 0, 4, 6, 3, 1],
[9, 4, 3, 8, 6, 7],
[2, 4, 3, 9, 5, 6],
[4, 7, 2, 0, 1, 8]]
when n=3, expected result:
[[[4, 7, 3, 0, 2, 8],[1, 5, 6, 5, 4, 8],[8, 9, 6, 5, 10, 4],[7, 8, 6, 6, 4, 10],[7, 8, 1, 0, 2, 3],[4, 5, 8, 5, 4, 6]],
[[1, 7, 2, 0, 2, 8],[4, 5, 8, 8, 4, 5],[9, 5, 6, 2, 1, 3],[5, 4, 1, 2, 6, 10],[7, 5, 4, 1, 1, 4],[9, 6, 5, 2, 2, 1]],
[[8, 2, 7, 10, 2, 7],[8, 9, 5, 4, 5, 5],[5, 8, 7, 7, 4, 6],[9, 5, 9, 10, 5, 4],[1, 4, 5, 6, 5, 7],[9, 8, 7, 6, 5, 1]],
[[0, 7, 4, 0, 1, 9],[4, 7, 3, 0, 2, 8],[8, 9, 6, 5, 10, 4],[8, 2, 7, 10, 2, 7],[[7, 5, 4, 1, 1, 4],[7, 8, 1, 9, 4, 1]],
[[9, 5, 3, 9, 2, 8],[8, 9, 6, 5, 10, 4],[9, 4, 3, 8, 6, 7],[3, 7, 10, 9, 4, 1],[4, 7, 3, 0, 2, 8],[9, 4, 3, 8, 6, 7]],
[[5, 3, 4, 5, 2, 10],[[7, 5, 4, 1, 1, 4],[4, 7, 3, 0, 2, 8],[4, 5, 8, 8, 4, 5],[7, 8, 1, 9, 4, 1],[8, 2, 7, 10, 2, 7]]]
I think I'd need a recursive function, but I'm totally out of a clue. Any insights? Thanks a lot.
Well base case of 1 is just to call your func, otherwise build a list of the recursion call on n-1 6^(n-1) times:
def recur_6(n):
if n <= 1:
return func()
return [recur_6(n-1) for _ in range(6**(n-1))]
Live example
Currently I have a 2D tensor, for each row, I want to insert a new element e before the first index of a specified value v. Additional information: cannot guarantee each row could have a such value. If there isn't, just append the element
Example: Supporse e is 0, v is 10, Given a tensor
[[9, 6, 5, 4, 10],
[8, 7, 3, 5, 5],
[4, 9, 10, 10, 10]]
I want to get
[[9, 6, 5, 4, 0, 10],
[8, 7, 3, 5, 5, 0],
[4, 9, 0, 10, 10, 10]]
Are there some Torch-style ways to do this? The worst case I can treat this as a trivial Python problem but I think the corresponding solution is a little time-consuming.
I haven't yet found a full PyTorch solution. I'll keep looking, but here is somewhere to start:
>>> v, e = 10, 0
>>> v, e = torch.tensor([v]), torch.tensor([e])
>>> x = torch.tensor([[ 9, 6, 5, 4, 10],
[ 8, 7, 3, 5, 5],
[ 4, 9, 10, 10, 10],
[10, 9, 7, 10, 2]])
To deal with the edge case where v is not found in one of the rows you can add a temporary column to x. This will ensure every row has a value v in it. We will use x_ as a helper tensor:
>>> x_ = torch.cat([x, v.repeat(x.size(0))[:, None]], axis=1)
tensor([[ 9, 6, 5, 4, 10, 10],
[ 8, 7, 3, 5, 5, 10],
[ 4, 9, 10, 10, 10, 10],
[10, 9, 7, 10, 2, 10]])
Find the indices of the first value v on each row:
>>> bp = (x_ == v).int().argmax(axis=1)
tensor([4, 5, 2, 0])
Finally, the easiest way to insert values at different positions in each row is with a list comprehension:
>>> torch.stack([torch.cat([xi[:bpi], e, xi[bpi:]]) for xi, bpi in zip(x, bp)])
tensor([[ 9, 6, 5, 4, 0, 10],
[ 8, 7, 3, 5, 5, 0],
[ 4, 9, 0, 10, 10, 10],
[ 0, 10, 9, 7, 10, 2]])
Edit - If v cannot occur in the first position, then no need for x_:
>>> x
tensor([[ 9, 6, 5, 4, 10],
[ 8, 7, 3, 5, 5],
[ 4, 9, 10, 10, 10]])
>>> bp = (x == v).int().argmax(axis=1) - 1
>>> torch.stack([torch.cat([xi[:bpi], e, xi[bpi:]]) for xi, bpi in zip(x, bp)])
tensor([[ 9, 6, 5, 0, 4, 10],
[ 8, 7, 3, 5, 0, 5],
[ 4, 0, 9, 10, 10, 10]])
I'm hoping to calculate the distances between two points in a (Nx1) numpy array, i.e.:
a = [2, 5, 5, 12, 5, 3, 10, 8, 1, 3, 1]
I'm hoping to get a square matrix with the (normed) distances between each point:
sq = [[0, |2-5|, |2-5|, |2-12|, |2-5|, ...],
[|5-2|, 0, ...], ...]
So far, what I have doesn't work, giving wrong values for the square distance matrix. Is there a way to (I'm not sure if it is the correct term?) vectorise my method too, but am unfamiliar with the advanced indexing.
What I currently have is the following:
sq = np.zero((len(a), len(a))
for i in a:
for j in len(a+1):
sq[i,j] = np.abs(a[:,0] - a[:,0])
Would appreciate any help!
I think that by exploiting numpy broadcasting, this is the faster solution:
a = [2, 5, 5, 12, 5, 3, 10, 8, 1, 3, 1]
a = np.array(a).reshape(-1,1)
sq = np.abs(a.T-a)
sq
array([[ 0, 3, 3, 10, 3, 1, 8, 6, 1, 1, 1],
[ 3, 0, 0, 7, 0, 2, 5, 3, 4, 2, 4],
[ 3, 0, 0, 7, 0, 2, 5, 3, 4, 2, 4],
[10, 7, 7, 0, 7, 9, 2, 4, 11, 9, 11],
[ 3, 0, 0, 7, 0, 2, 5, 3, 4, 2, 4],
[ 1, 2, 2, 9, 2, 0, 7, 5, 2, 0, 2],
[ 8, 5, 5, 2, 5, 7, 0, 2, 9, 7, 9],
[ 6, 3, 3, 4, 3, 5, 2, 0, 7, 5, 7],
[ 1, 4, 4, 11, 4, 2, 9, 7, 0, 2, 0],
[ 1, 2, 2, 9, 2, 0, 7, 5, 2, 0, 2],
[ 1, 4, 4, 11, 4, 2, 9, 7, 0, 2, 0]])
With numpy the following line might be the shortest to your result:
import numpy as np
a = np.array([2, 5, 5, 12, 5, 3, 10, 8, 1, 3, 1])
sq = np.array([np.array([(np.abs(i - j)) for j in a]) for i in a])
print(sq)
The following would give you the desired result without numpy.
a = [2, 5, 5, 12, 5, 3, 10, 8, 1, 3, 1]
sq = []
for i in a:
distances = []
for j in a:
distances.append(abs(i-j))
sq.append(distances)
print(sq)
With both, the result comes as:
[[0, 3, 3, 10, 3, 1, 8, 6, 1, 1, 1], [3, 0, 0, 7, 0, 2, 5, 3, 4, 2, 4], [3, 0, 0, 7, 0, 2, 5, 3, 4, 2, 4], [10, 7, 7, 0, 7, 9, 2, 4, 11, 9, 11], [3, 0, 0, 7, 0, 2, 5, 3, 4, 2, 4], [1, 2, 2, 9, 2, 0, 7, 5, 2, 0, 2], [8, 5, 5, 2, 5, 7, 0, 2, 9, 7, 9], [6, 3, 3, 4, 3, 5, 2, 0, 7, 5, 7], [1, 4, 4, 11, 4, 2, 9, 7, 0, 2, 0], [1, 2, 2, 9, 2, 0, 7, 5, 2, 0, 2], [1, 4, 4, 11, 4, 2, 9, 7, 0, 2, 0]]
There may be more than one way to do this but one way is to only use numpy operations instead of loops because internally python does lots of optimizations for numpy arrays.
One way to do only using array operations is to create an NxN matrix by repeating the original matrix (a) N times.
This will create a matrix N times.
E.g:
a = [1, 2, 3]
b = [[1 , 2, 3], [1 , 2, 3], [1 , 2, 3]]
Then you can do a matrix, array operation of
ans = abs(b - a)
Assuming a is numpy array, you can do:
b = np.repeat(a,a.shape).reshape((a.shape[0],a.shape[0]))
ans = np.abs(b - a)
I want to ask how numpy remove columns in batch by list.
The value in the list corresponds to the batch is different from each other.
I know this problem can use the for loop to solve, but it is too slow ...
Can anyone give me some idea to speed up?
array (batch size = 3):
[[0, 1, 2, 3, 4, 5, 6], [0, 1, 2, 3, 4, 5, 6], [0, 1, 2, 3, 4, 5, 6]]
remove index in the list (batch size = 3)
[[2, 3, 4], [1, 2, 6], [0, 1, 5]]
output:
[[0, 1, 5, 6], [0, 3, 4, 5], [2, 3, 4, 6]]
Assuming the array is 2d, and the indexing removes equal number of elements per row, we can remove items with a boolean mask:
In [289]: arr = np.array([[0, 1, 2, 3, 4, 5, 6], [0, 1, 2, 3, 4, 5, 6], [0, 1, 2, 3, 4, 5, 6]]
...: )
In [290]: idx = np.array([[2, 3, 4], [1, 2, 6], [0, 1, 5]])
In [291]: mask = np.ones_like(arr, dtype=bool)
In [292]: mask[np.arange(3)[:,None], idx] = False
In [293]: arr[mask]
Out[293]: array([0, 1, 5, 6, 0, 3, 4, 5, 2, 3, 4, 6])
In [294]: arr[mask].reshape(3,-1)
Out[294]:
array([[0, 1, 5, 6],
[0, 3, 4, 5],
[2, 3, 4, 6]])