Scatter_sample like operation?

1

i have a tensor x = [2, 3, 2, 1, 2, 3, 2]
I want to randomly sample a number from all same number, e.g. randomly sample the 2 from all 2’s index, and get index may be 0, namely sampled the first 2; randomly sample the 3, and get index may be 5. and randomly sample the 1, get index 3 since only single 1 exists. so the return index may like [0, 5, 3] or mask: [1, 0, 0, 1, 0, 1, 0].
I dnt know how to implement this logic by pytorch? please kindly help me.

Thank you so much.

2

You could get all occurrences of the current number you would like to draw and randomly permute these indices:

x = torch.tensor([2, 3, 2, 1, 2, 3, 2])
to_sample = torch.as_tensor(2)
idx_all = (x == to_sample).nonzero()
idx = idx_all[torch.randperm(len(idx))[0]]
1 Like
3

@ptrblck Thanks.
However, if i want to iterate over all unique number, i have to use a python for loop:

x = torch.tensor([2, 3, 2, 1, 2, 3, 2])
unique = torch.unique(x)
for u in unique: #!!! time consuming!
    to_sample = u
    idx_all = (x == to_sample).nonzero()
    idx = idx_all[torch.randperm(len(idx))[0]]

any concise way to finish it without python for loop?
Thank you.

4

using numpy might be faster; why iterating over your list?
Something like this,

import numpy as np
values = np.array([1,2,3,1,2,4,5,6,3,2,1])
searchval = 3
ii = np.where(values == searchval)[0]
5

In that case, I would just sample the values using torch.multinomial:

x = torch.tensor([2, 3, 2, 1, 2, 3, 2])
unique = torch.unique(x)
weight = x.unsqueeze(0) == unique.unsqueeze(1)
torch.multinomial(weight.float(), num_samples=1)
1 Like