BUG: fix issue #8250 when np.array gets called on an invalid sequence

[evanlimanto]

Fixes the issue in #8250. We should let np.random.shuffle should handle the intricacies of the different types passed in.

[stefanv]

Thanks, @evanlimanto. Could you please add a test to verify this behavior?

[evanlimanto]

Added tests. I know I could've made a helper function (in duplicating the tests for shuffle), but I didn't want to disrupt the structure of the code.

[eric-wieser]

Looks good to me. I'll merge in a few days if there are no other objections

[evanlimanto]

Can this be merged soon?

[eric-wieser]

Hang on, this means that x is now modified in place, rather than copied first. This isn't ok. Perhaps a copy.copy here instead?

[seberg]

Its also not good, that the tests did not pick it up.

[evanlimanto]

Fixed, and added a test to check for mutation.

[eric-wieser]

Can this go at the top with the other imports?

[eric-wieser]

This might be clearer immediately under the permutation call

[evanlimanto]

Fixed, thanks!

[eric-wieser]

Looks good to me, as do the tests. The source of array_copy shows that np.copy(arr) is basically equivalent to copy.copy(arr), so I think we're all good here.

[eric-wieser]

Thanks @evanlimanto

[jreback]

This broke one of our tests downstream because this is now not returning an array (but a list). I don't think this is what you want?

You are testing this input, but not that it actually returns an ndarray (just that it is equal).

https://travis-ci.org/pandas-dev/pandas/jobs/210272568

In [5]: np.__version__
Out[5]: '1.13.0.dev0+cf349c9'

In [6]: np.random.permutation([(0, 1), (1, 0)])
Out[6]: [(0, 1), (1, 0)]

In [1]: np.__version__
Out[1]: '1.12.0'

In [2]: np.random.permutation([(0, 1), (1, 0)])
Out[2]: 
array([[1, 0],
       [0, 1]])

[seberg]

Hmmm, maybe we should just revert and just change the documentation, its a safer bet. Is there much gain in allowing general sequences here?

[charris]

I think reversion would be the quick fix. I'm not sure there is a better fix.

[eric-wieser]

How about the following?

try:
    arr = np.copy(arr)
except:
    arr = copy.copy(arr)

[charris]

I think the problem is that an array should be returned by default, as previously happened. So at a minimum we still need arr = array(arr). Note that the example given does work if dtype=object is used in the array conversion.

In [7]: %paste
N = 4
A = np.arange(N)[:,None]
A = np.concatenate((A,A,A,A), axis = 1)
B = range(N)
c = list(zip(B,A))

## -- End pasted text --

In [8]: np.random.permutation(array(c, dtype=object))
Out[8]: 
array([[0, array([0, 0, 0, 0])],
       [3, array([3, 3, 3, 3])],
       [2, array([2, 2, 2, 2])],
       [1, array([1, 1, 1, 1])]], dtype=object)

So perhaps use that as the fallback, although that may also lead to hard to interpret errors.