Exposing the gradient of the log-likelihood

General
1

Hello all,

Is it possible to directly expose the gradient of the log-likelihood in Stan for every transition and to define it as a variable in the model?
I saw that the last mention of this was more than 4 years ago:

Thanks!!

1 Like
2

No

1 Like
3

Thanks for the answer! So how would one access the gradient or calculate (and store) it on the fly? Is that even possible?

4

You would have to either do it analytically and / or write your own C++ (very carefully).

1 Like
5

+1.

If you do it analytically or numerically, you can do it in Stan.

If you’re willing to get to C++, anything is possible.

6

ok thanks! Is there an example how to analytically or numerically calculate the gradient in Stan?
C++ is over my head atm (unless it is easier to implement there).

Thanks again!

7

and based on this discussion:

is it still possible? Because I didn’t see the gradient info for each iteration as one of the outputs in the diagnostic file from CmdStan.

8

It is in the columns that start with grad, although that is the gradient with respect to the unconstrained parameters, which may not be what you want.

9

I checked my *diagnostic.csv again and I have columns that start with g_* for different parameters. Is this the gradient in the unconstrained space? (I don’t see columns that start with grad.)
Also, what does p_* for different parameters mean?

10

yes

momentum

11

Ok, thanks! That’s super helpful. Now is there a way to access the gradient values for each iteration as I sample my model? I don’t need the values for each leapfrog step, so the unconstrained space and per-iteration gradient value is exactly what I need.

12
1 Like
13

I see :) thanks!!