Sampling from a tensor in Torch

1

I wanted to know whether we could sample elements of a tensor given a probability distribution of the tensor.
In numpy we would do something like this:

a = np.array([1,2,3,4])
b = np.random.choice(a, p=np.array([0.1, 0.1, 0.1, 0.7]))

In torch I would like to have the array a and p to be of torch.tensor. The numpy function works well with CPU torch tensors translated to numpy arrays, but unfortunately, using GPUs the trick fails.

2

Might be helpful? Torch equivalent of numpy.random.choice?

3

Have seen, did not do the trick.

4

Hi !
If you want to do :

a_numpy = np.array([1, 2, 3, 4])
p_numpy = np.array([0.1, 0.1, 0.1, 0.7])
b_numpy = np.random.choice(a_numpy, p=p_numpy)

You can use torch.multinomial :

a = torch.tensor([1, 2, 3, 4])
p = torch.tensor([0.1, 0.1, 0.1, 0.7])
index = p.multinomial(num_samples=1, replacement=True)
b = a[index]

Here I used replacement=True because it is the default behavior of numpy.choice, even though is makes no difference when num_samples=1

1 Like