Use np.random's RandomState when seed is None

[ariddell]

• closes #13143
• tests added / passed
• passes git diff upstream/master | flake8 --diff
• whatsnew entry

The handle for numpy's current random state is
np.random.mtrand._rand.

Compare https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/utils/validation.py#L573

[shoyer]

Can we use something like this instead?

seed = np.random.randint(2 ** 32)
return np.random.RandomState(seed)

That way we don't need to use NumPy's private API.

[jorisvandenbossche]

I don't think that will do the same?

As this is to catch a previously called np.random.seed(), so feeding it another seed will not give the desired result I think?

[shoyer]

We want results to be reproducible after calling np.random.seed(). But we don't need to reuse exactly the same seed -- it's OK to also use a seed derived from NumPy's seed.

Using the exact same seed would only be necessary if we want to promise the sample makes the exact same choice as np.random.choice. But that's really an implementation detail.

(Clearly a comment in the code would also be in order if we use my suggested approach.)

[jorisvandenbossche]

Aha, I see.
For me it does not really matter which approach we take (sklearn also uses _rand, so although it is private it seems ok)

[shoyer]

This test doesn't look right to me. You need to call np.random.seed(test) before every call to sample, not just the first one.

[ariddell]

You're right. I fixed the test.

On 05/12, Stephan Hoyer wrote:

@@ -415,6 +415,10 @@ def test_sample(self):
o.sample(frac=0.7, random_state=np.random.RandomState(test)),
o.sample(frac=0.7, random_state=np.random.RandomState(test)))

•        np.random.seed(test)
  

•        self._compare(o.sample(n=4), o.sample(n=4))
  

This test doesn't look right to me. You need to call np.random.seed(test) before every call to sample, not just the first one.

---

You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub:
https://github.com/pydata/pandas/pull/13161/files/08b2cd1ae1e581e720d169cbf8d4888adcb07e67#r63099499

[jreback]

@ariddell pls also add a whatsnew entry. I would do it in API changes section.

[ariddell]

@jreback added. Thanks!

[shoyer]

@ariddell could you please fix the use of NumPy's private state as noted above? This is something simple that could avoid significant pain down the road.

[ariddell]

Fair enough. What should the seed be? Did you mean 2 ** 32 above?

[shoyer]

@ariddell Yes, I meant 2 ** 32 :).

[rkern]

Using np.random.mtrand._rand is the right approach. It's not going to disappear. This is an authorized use.

[shoyer]

Let's just return np.random from _random_state instead of np.random.mtrand._rand, using @rkern's suggestion of duck typing for RandomState objects from the related thread on numpy-discussion: https://mail.scipy.org/pipermail/numpy-discussion/2016-May/075489.html

[ariddell]

@shoyer duck typing it is.

[shoyer]

can you please update the docstring, too?

[ariddell]

Yes. Fixed.

On 05/17, Stephan Hoyer wrote:

@@ -2072,7 +2072,7 @@ def _random_state(state=None):
elif isinstance(state, np.random.RandomState):
return state
elif state is None:

•    return np.random.RandomState()
  

•    return np.random
  

can you please update the docstring, too?

---

You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub:
https://github.com/pydata/pandas/pull/13161/files/c8a38f99637ad44dad0db9118b17fd4e3c8643f3#r63603099

[jreback]

you have a failing test

[jreback]

@ariddell can you rebase / update

[ariddell]

OK. I think I've got it.

[jreback]

thanks @ariddell

[pydolan]

I'm not getting the behavior I'd expect from this fix in Pandas 0.19.2. Example:

>>> import pandas as pd
>>> import numpy as np
>>> np.random.seed(12345678)
>>> df = pd.DataFrame(np.random.randn(10, 5), columns=['a', 'b', 'c', 'd', 'e'])
>>> df.sample(n=2)
          a         b         c         d         e
8  0.365786 -0.855795 -0.511062 -0.116734  1.133541
0  1.281455 -0.423852  0.011236 -0.945194 -0.088124
>>> df.sample(n=2)
          a         b         c         d         e
2  0.341751 -1.158737  0.313814  1.827552 -0.351045
9  1.355003  0.532651  0.250463 -0.281751 -0.342741
>>> df.sample(n=2, random_state=12345678)
          a         b         c         d         e
8  0.365786 -0.855795 -0.511062 -0.116734  1.133541
6 -0.401077 -0.047115 -0.410951 -1.608354  0.290594
>>> df.sample(n=2, random_state=12345678)
          a         b         c         d         e
8  0.365786 -0.855795 -0.511062 -0.116734  1.133541
6 -0.401077 -0.047115 -0.410951 -1.608354  0.290594
>>> pd.__version__
u'0.19.2'
>>> np.__version__
'1.12.0'

Any thoughts on why df.sample(n=2) doesn't match df.sample(n=2, random_state=12345678)?

[shoyer]

@pydolan is there a missing np.random.seed() call somewhere in your example?

[pydolan]

@shoyer -- sorry about that; will update example with my actual code (i.e., with the one missing line)

[shoyer]

@pydolan It looks like you are calling np.random.randn() after setting the seed before df.sample()

[pydolan]

If I move the seed() call after instantiating the dataframe, I still get inconsistent behavior with calls to sample(), except when I provide the random_state arg. Example:

>>> import pandas as pd
>>> import numpy as np
>>> df = pd.DataFrame(np.random.randn(10, 5), columns=['a', 'b', 'c', 'd', 'e'])
>>> np.random.seed(12345678)
>>> df.sample(n=2)
          a         b         c         d         e
8 -1.250204  0.551508  1.408080  0.397452  0.424326
6 -0.028298  0.203270  0.939094 -1.802227 -0.088679
>>> df.sample(n=2)
          a         b         c         d         e
4  0.895497  0.609853 -1.548664 -1.238415 -1.058904
5  0.196420  0.472877 -0.918205  1.019862 -0.631993
>>> df.sample(n=2, random_state=12345678)
          a         b         c         d         e
8 -1.250204  0.551508  1.408080  0.397452  0.424326
6 -0.028298  0.203270  0.939094 -1.802227 -0.088679
>>> df.sample(n=2, random_state=12345678)
          a         b         c         d         e
8 -1.250204  0.551508  1.408080  0.397452  0.424326
6 -0.028298  0.203270  0.939094 -1.802227 -0.088679

Note that this time, the first call to sample() uses the seed, but the second call does not use the seed. Is it expected that seed() needs to be called before every sample call? I thought it is supposed to be, "set once", and all future randomization-related calls should use (including my original example, where randn() is called after seed() and before sample()).

(Also note that I did verify that calling seed() before every call to sample() does indeed produce the same sampled rows.)

[shoyer]

Is it expected that seed() needs to be called before every sample call? I thought it is supposed to be, "set once", and all future randomization-related calls should use (including my original example, where randn() is called after seed() and before sample()).

This is working as intended. Try just sampling random numbers from numpy -- it does the exact same thing.

[pydolan]

Good point; I completely spaced on the nature of random sampling. Sorry for the waste of time!