Hello!

I was hoping to get some help understanding using to_event(1), particularly in the Gaussian case. Does calling to_event(1) on some Gaussian distribution make it act like a MultivariateGaussian still when we use the TraceMeanFieldELBO? This was my understanding based on the Pyro tutorials on dimensions.

I’m asking because I was having some HUGE discrepancies when I was extracting the KL term in my ELBO with the suggestions in (Extract the KL divergence term from loss. How?) compared to explicitly calculating it from my parameters.

my understanding is if I call .to_event(1) on a Gaussian that, at a higher level, this signal to Pyro to pretend “hey, this is actually a Multi-gaussian distribution” So my probabalistic functions are like this:

```
#necessary imports
def model(args**):
#doing over a minibatch of my args
poutine.plate("data"):
mus, sigmas = BlackBoxMagic(args**)
distr = Gaussian(mus, sigmas).toevent(1) #my pretend Multivariate Gauss
pyro.sample("z", distr)
#the rest of my fancy model
def guide(args**):
#doing over a minibatch of my args
poutine.plate("data"):
mus, sigmas = OtherBlackBoxMagic(args**)
distr = Gaussian(mus, sigmas).toevent(1) #my pretend Multivariate Gauss
pyro.sample("z", distr)
optim = Adam({"lr": 0.001})
svi = SVI(model, guide, optim, TraceMeanField_ELBO())
svi.step(batch_o_data) #my ELBO is presumably going to use the KL for 2 multivariates??
```

when I actually step through the Trace_MeanFieldELBO and look at my sites which use this trick, I notice that internally, it looks like PyTorch is treating this as the regular KL between to Normals.

Any thoughts are appreciated!