On Fri, Jul 18, 2014 at 12:53 PM, Nathaniel Smith <[email protected]> wrote:

> On Fri, Jul 18, 2014 at 12:38 PM,  <[email protected]> wrote:
> >
> > On Thu, Jul 17, 2014 at 4:07 PM, <[email protected]> wrote:
> >
> >> If you mean by this to add atol=1e-8 as default, then I'm against it.
> >>
> >> At least it will change the meaning of many of our tests in statsmodels.
> >>
> >> I'm using rtol to check for correct 1e-15 or 1e-30, which would be
> >> completely swamped if you change the default atol=0.
> >> Adding atol=0 to all assert_allclose that currently use only rtol is a
> lot
> >> of work.
> >> I think I almost never use a default rtol, but I often leave atol at the
> >> default = 0.
> >>
> >> If we have zeros, then I don't think it's too much work to decide
> whether
> >> this should be atol=1e-20, or 1e-8.
> >
> >
> > copied from
> > http://mail.scipy.org/pipermail/numpy-discussion/2014-July/070639.html
> > since I didn't get any messages here
> >
> > This is a compelling use-case, but there are also lots of compelling
> > usecases that want some non-zero atol (i.e., comparing stuff to 0).
> > Saying that allclose is for one of those use cases and assert_allclose
> > is for the other is... not a very felicitious API design, I think. So
> > we really should do *something*.
> >
> > Are there really any cases where you want non-zero atol= that don't
> > involve comparing something against a 'desired' value of zero? It's a
> > little wacky, but I'm wondering if we ought to change the rule (for
> > all versions of allclose) to
> >
> > if desired == 0:
> >     tol = atol
> > else:
> >     tol = rtol * desired
> >
> > In particular, means that np.allclose(x, 1e-30) would reject x values
> > of 0 or 2e-30, but np.allclose(x, 0) will accept x == 1e-30 or 2e-30.
> >
> > -n
> >
> >
> > That's much too confusing.
> > I don't know what the usecases for np.allclose are since I don't have
> any.
>
> I wrote allclose because it's shorter, but my point is that
> assert_allclose and allclose should use the same criterion, and was
> making a suggestion for what that shared criterion might be.
>
> > assert_allclose is one of our (statsmodels) most frequently used numpy
> > function
> >
> > this is not informative:
> >
> > `np.allclose(x, 1e-30)`
> >
> >
> > since there are keywords
> > either np.assert_allclose(x, atol=1e-30)
>
> I think we might be talking past each other here -- 1e-30 here is my
> "gold" p-value that I'm hoping x will match, not a tolerance argument.
>

my mistake



>
> > if I want to be "close" to zero
> > or
> >
> > np.assert_allclose(x, rtol=1e-11, atol=1e-25)
> >
> > if we have a mix of large numbers and "zeros" in an array.
> >
> > Making the behavior of assert_allclose depending on whether desired is
> > exactly zero or 1e-20 looks too difficult to remember, and which desired
> I
> > use would depend on what I get out of R or Stata.
>
> I thought your whole point here was that 1e-20 and zero are
> qualitatively different values that you would not want to accidentally
> confuse? Surely R and Stata aren't returning exact zeros for small
> non-zero values like probability tails?
>
> > atol=1e-8 is not close to zero in most cases in my experience.
>
> If I understand correctly (Tony?) the problem here is that another
> common use case for assert_allclose is in cases like
>
> assert_allclose(np.sin(some * complex ** calculation / (that - should
> - be * zero)), 0)
>
> For cases like this, you need *some* non-zero atol or the thing just
> doesn't work, and one could quibble over the exact value as long as
> it's larger than "normal" floating point error. These calculations
> usually involve "normal" sized numbers, so atol should be comparable
> to eps * these values.  eps is 2e-16, so atol=1e-8 works for values up
> to around 1e8, which is a plausible upper bound for where people might
> expect assert_allclose to just work. I'm trying to figure out some way
> to support your use cases while also supporting other use cases.
>

my problem is that there is no "normal" floating point error.
If I have units in 1000 or units in 0.0001 depends on the example and
dataset that we use for testing.

this test two different functions/methods that calculate the same thing

(Pdb) pval
array([  3.01270184e-42,   5.90847367e-02,   3.00066946e-12])
(Pdb) res2.pvalues
array([  3.01270184e-42,   5.90847367e-02,   3.00066946e-12])
(Pdb) assert_allclose(pval, res2.pvalues, rtol=5 * rtol, atol=1e-25)

I don't care about errors that are smaller that 1e-25

for example testing p-values against Stata

(Pdb) tt.pvalue
array([  5.70315140e-30,   6.24662551e-02,   5.86024090e-11])
(Pdb) res2.pvalues
array([  5.70315140e-30,   6.24662551e-02,   5.86024090e-11])
(Pdb) tt.pvalue - res2.pvalues
array([  2.16612016e-40,   2.51187959e-15,   4.30027936e-21])
(Pdb) tt.pvalue / res2.pvalues - 1
array([  3.79811738e-11,   4.01900735e-14,   7.33806349e-11])
(Pdb) rtol
1e-10
(Pdb) assert_allclose(tt.pvalue, res2.pvalues, rtol=5 * rtol)


I could find a lot more and maybe nicer examples, since I spend quite a bit
of time fine tuning unit tests.

Of course you can change it.

But the testing functions are code and very popular code.

And if you break backwards compatibility, then I wouldn't mind reviewing a
pull request for statsmodels that adds 300 to 400 `atol=0` to the unit
tests. :)

Josef


>
> -n
>
> --
> Nathaniel J. Smith
> Postdoctoral researcher - Informatics - University of Edinburgh
> http://vorpus.org
> _______________________________________________
> NumPy-Discussion mailing list
> [email protected]
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>
_______________________________________________
NumPy-Discussion mailing list
[email protected]
http://mail.scipy.org/mailman/listinfo/numpy-discussion

Reply via email to