Fix sample_ppc

[aseyboldt]

This should hopefully fix #1824 and #2307.

[twiecki]

brief doc-string would be helpful.

[aseyboldt]

coming up soon

[twiecki]

why is that required?

[aseyboldt]

draw_values used to return a list if it was passed more than one value and a single value otherwise. I didn't like that and change it to return always a list. I am playing around with sample_ppc a bit, and there it could be that the list of variable to draw is dynamic, which would make the previous behaviour somewhat strange. But I can change it back if you disagree.

[twiecki]

I see, I don't disagree.

[junpenglao]

is it effect the speed? Since you commented in #2296 that draw_values might also cause slowdown in random method.

[aseyboldt]

I think so

[aseyboldt]

@hvasbath Can you verify that the speed issue is gone with this PR? It runs faster in a simple example I tried.

[hvasbath]

I am not experienced enough with git to checkout this PR. Can you please post some command lines I would need to follow to try? Thanks!

[fonnesbeck]

@hvasbath The pattern is like this:

git fetch origin pull/<pr_number>/head:<local_name>
git checkout <local_name>

where local_name is whatever you want to call the PR on your machine.

[hvasbath]

Thanks @fonnesbeck!
With the test in #1824
Time on master:
('time for initialising population', 26.08741807937622)
Time with this PR:
('time for initialising population', 1.317948818206787)

So yeah almost as good as before! Good job @aseyboldt !
@junpenglao Now it doesnt compile a theano function anymore for each sample, which is why it is much faster again. It compiles the functions if there are theano.tensors involved- thats why it affects the speed.

[twiecki]

Awesome, thanks for checking @hvasbath and sorry you went down that other rabbit hole. Will wait for the doc string and then merge.

[aseyboldt]

@twiecki I wrote a docstring and also made that function private, there really isn't any reason why someone should be using that instead of draw_values.

@hvasbath Good to hear and thanks for checking.

[twiecki]

Awesome, can merge once tests pass.

[hvasbath]

Awesome that this is merged! It heavily affected SMC initialisation! @twiecki at least it was useful to draw @aseyboldt s attention. So that it is fixed in a much better and cleaner way now! No one can know everything right ;) . Thanks again for fixing @aseyboldt !