You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In a non-centered normal distribution, if the standard deviation is initialized to a very low value, it crashes to an extremely low value and never recovers. Manual non-centering does not encounter the same problem. This makes non-centering via offset/multiplier unusable for some models (see below).
This is pretty repeatable that the custom code doesn’t have a problem with inits but the built in does.
Encountering this issue "in the wild"
Some classes of model reliably pinch through very small standard deviations early in warmup. Here's an example--not extreme enough to hit the "sticky boundary", but enough to show why it can be an issue. Notice how in early warmup sigma pinches down to a very low value before recovering. This is consistent across seeds.
data{
int n;
real y[n];
}
parameters{
real<lower = 0> sigma;
real<multiplier=sigma> x[n];
}
model{
sigma ~ std_normal();
x ~ normal(0, sigma);
y ~ normal(x, .01);
}
It's not the end of the world, because manual non-centering still works fine, but this issue makes offset/multiplier noncentering unusable in some of the models I work with.
Current Version:
v2.26.1
The text was updated successfully, but these errors were encountered:
Summary:
In a non-centered normal distribution, if the standard deviation is initialized to a very low value, it crashes to an extremely low value and never recovers. Manual non-centering does not encounter the same problem. This makes non-centering via offset/multiplier unusable for some models (see below).
For more, see: https://discourse.mc-stan.org/t/offset-multiplier-initialization/20712
Description and Reprex
Here's @bbbales2 on the issue, pasted over from discourse (link above):
Encountering this issue "in the wild"
Some classes of model reliably pinch through very small standard deviations early in warmup. Here's an example--not extreme enough to hit the "sticky boundary", but enough to show why it can be an issue. Notice how in early warmup
sigma
pinches down to a very low value before recovering. This is consistent across seeds.Code to run:
It's not the end of the world, because manual non-centering still works fine, but this issue makes offset/multiplier noncentering unusable in some of the models I work with.
Current Version:
v2.26.1
The text was updated successfully, but these errors were encountered: