[np.random] shuffle is not compatible with pytorch tensor

When shuffling a pytorch tensor with numpy.random.shuffle, incorrect result will be produced with 100% (so far) chance.

Reproducing code example:

import torch as th
import numpy as np

N = 10
x = th.arange(N)
print(x)
np.random.shuffle(x)
print(x)
assert(len(set(x.tolist())) == N)

Error message:

tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
tensor([0, 0, 2, 1, 3, 4, 6, 3, 3, 9])
Traceback (most recent call last):
  File "x.py", line 9, in <module>
    assert(len(set(x.tolist())) == N)
AssertionError

In the shuffled result, 0 and 3 occured twice, which is clearly problematic.

NumPy/Python version information:

numpy ‘1.18.5’ through anaconda.

See also: https://github.com/pytorch/pytorch/issues/50880

[rkern]

I’ve added some comments on pytorch/pytorch#50880

I suspect that we’ll need input from both communities to properly resolve this, so I apologize to all for the bifurcated discussion.

[rgommers]

Thanks. Fix in gh-18211.