fix #2108

[junpenglao]

No description provided.

[junpenglao]

fix #2108 @twiecki @ferrine

[ferrine]

What is wrong with order? Is it the best place for fix?

[junpenglao]

@ferrine basically, the problem is that the somehow the bijection.map in approximation class is different than the model.bijection.map. So in the
https://github.com/pymc-devs/pymc3/blob/master/pymc3/sampling.py#L588-L589, the cov has a different order compare to start.
As a result, the pm.NUTS get the wrong scaling.

In the old advi init, it does cov = np.power(model.dict_to_array(v_params.stds), 2) to make sure the cov has the right order.
https://github.com/pymc-devs/pymc3/blob/bb25d7bb43aa0684fdbd7628546be044c873b27b/pymc3/sampling.py#L474

[twiecki]

Seems like there's a underlying issue which we should fix instead.

…

On Tue, May 2, 2017 at 10:23 PM, Junpeng Lao ***@***.***> wrote: @ferrine <https://github.com/ferrine> basically, the problem is that the somehow the bijection.map in approximation class is different than the model.bijection.map. So in the https://github.com/pymc-devs/pymc3/blob/master/pymc3/sampling.py#L588-L589, the cov has a different order compare to start. As a result, the pm.NUTS get the wrong scaling. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2110 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AApJmHpWUJyxJu5oDxyRLDHCjvAzuZWVks5r15CvgaJpZM4NOkyT> .

[junpenglao]

@twiecki I agree - something must happen during pm.fit, which somehow operate on model.
However, it might be useful to have this fix in place as a fail-safe

[ferrine]

that's more pretty

stds = mf.gbij.rmap((tt.diag(mf.cov).eval()))
cov = model.dict_to_array(stds)

[junpenglao]

thanks! I will go with this.
If I understand correctly the posterior cov of advi is a diagonal matrix right?

[ferrine]

Yes

[twiecki]

>           stds = approx.gbij.rmap((tt.diag(approx.cov).eval()))
E           NameError: global name 'tt' is not defined

[junpenglao]

Oops, thanks @twiecki