From e8c2501608f6db53ddc78ad48b4aa7874e446b53 Mon Sep 17 00:00:00 2001 From: Sam Duffield <s@mduffield.com> Date: Tue, 5 Nov 2024 11:36:24 +0000 Subject: [PATCH] Add aux to Going Bayesian doc --- docs/log_posteriors.md | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/docs/log_posteriors.md b/docs/log_posteriors.md index 1d46ada1..df7d0bb2 100644 --- a/docs/log_posteriors.md +++ b/docs/log_posteriors.md @@ -98,9 +98,12 @@ either $N$ or $n$ increaase. log_prior = diag_normal_log_prob(params, sd=1., normalize=False) mean_log_lik = Categorical(logits=logits).log_prob(batch['labels']).mean() mean_log_post = log_prior / num_data + mean_log_lik - return mean_log_post + return mean_log_post, torch.tensor([]) ``` + See [auxiliary information](#auxiliary-information) for why we return an + additional empty tensor. + The issue with running Bayesian methods (such as VI or SGHMC) on this mean log posterior function is that naive application will result in approximating the tempered posterior