Warning: scipy.stats.distributions rvs broadcasting
josef...@gmail.com
I just got bitten again by http://projects.scipy.org/scipy/ticket/1544
scale is scalar, mu_true is array of loc
good:
y_obs = self.rvs(mu_true, scale=scale, size=nobs)
bad:
y_obs = self.rvs(mu_true, scale=scale)
this has no noise, one random draw gets shifted by the loc array
If the array is for the shape parameter and loc and scale are scalars,
then it's fine.
Josef
Wes McKinney
Major pet peeve of mine. I've been bitten by this before too-- is it
an easy fix?
josef...@gmail.com
You didn't open a ticket. Major pet peeves are supposed to be
reported, maybe they are bugs. :)
Not in 5 minutes, but a warning would be good.
The general solution might require quite a bit of shape checking to
see that things broadcast or don't broadcast correctly. But I haven't
looked at the details yet.
I don't know if anyone intentionally uses the current behavior.
Josef
Warren Weckesser
I was just talking to Travis about this. He suspects is was just an oversight. We can fix this (I'll do it if no one beats me to it), but statsmodels will still have to work around the bug in the current releases of scipy.
Warren
Josef
josef...@gmail.com
Thanks Warren. I'm glad if I don't have to beat you to it. I'm still
writing tests (with size=...) and bug hunting in GAM, and then I would
like to hunt for my own bugs (or write new buggy code so I can look
for them:).
Josef
>
> Warren
>
>
>
>>
>> Josef
>
>
josef...@gmail.com
I found by chance how to create truncated normal random variables in
Greene, only to realize it is just the standard ppf transformation as
it is also used in scipy.stats.
After reinventing the wheel, I thought I also reinvent broadcasting
(so my time is not completely wasted).
http://groups.google.com/group/scipy-user/browse_thread/thread/f778f5ff4660e451?hl=en
(aside: this should be useful for simulating Probit models.)
Josef
>
> Josef
>
>>
>> Warren
>>
>>
>>
>>>
>>> Josef
>>
>>
>