Unique values in PyTorch tensor

Question

I'm tying to find distinct values in a PyTorch tensor. Is there an efficient analogue of Tensorflow's unique op?

Answer 1

There is a torch.unique() method in 0.4.0

In torch <= 0.3.1 you can try:

import torch
import numpy as np

x = torch.rand((3,3)) * 10
np.unique(x.round().numpy())

Answer 2

You can convert to a numpy array and make use of numpy's built-in unique function:

def unique(tensor1d):
    t, idx = np.unique(tensor1d.numpy(), return_inverse=True)
    return torch.from_numpy(t), torch.from_numpy(idx)  

Example:

t, idx = unique(torch.LongTensor([1, 1, 2, 4, 4, 4, 7, 8, 8]))  
# t --> [1, 2, 4, 7, 8]
# idx --> [0, 0, 1, 2, 2, 2, 3, 4, 4]

Answer 3

1. get common items between two tensors with torch.eq()
2. fetch indices and concatenate tensors
3. finally get common items via torch.unique:

import torch as pt

a = pt.tensor([1,2,3,2,3,4,3,4,5,6])
b = pt.tensor([7,2,3,2,7,4,9,4,9,8])

equal_data = pt.eq(a, b)
pt.unique(pt.cat([a[equal_data],b[equal_data]]))