Hey, great effort, would love to see Bayesian approaches be adopted and such workflow tools bridge the practice gap.
A comment on the example. A Logistic Regression model is presented:
\begin{align}
y_i | f_i &\sim \text{Bernoulli}(\text{logit}=f_i), \
f_i | x_i $= w^\top x_i, \
w &\sim \mathcal{N}(0, \sigma^2), \
\sigma &\sim \dots
\end{align}
Method predict_dist seems to sample the posterior $f|D$ instead of $y|D$ (which would be $\in{0, 1}$). This is a bit problematic, as MAP logistic regression already returns a point estimate for $f$, which defines a Bernoulli distribution for $y$. In the Bayesian case, $y|D \sim \text{Bernoulli}(p=\mathbb{E}[f|D])$.
Similarly, predict seems to give the median of $f|D$ (which incidentally is not necessarily a useful quantity in the Bayesian paradigm).
Did I misunderstand function or intent?
Hey, great effort, would love to see Bayesian approaches be adopted and such workflow tools bridge the practice gap.
A comment on the example. A Logistic Regression model is presented:
\begin{align}
y_i | f_i &\sim \text{Bernoulli}(\text{logit}=f_i), \
f_i | x_i $= w^\top x_i, \
w &\sim \mathcal{N}(0, \sigma^2), \
\sigma &\sim \dots
\end{align}
Method$f|D$ instead of $y|D$ (which would be $\in{0, 1}$ ). This is a bit problematic, as MAP logistic regression already returns a point estimate for $f$ , which defines a Bernoulli distribution for $y$ . In the Bayesian case, $y|D \sim \text{Bernoulli}(p=\mathbb{E}[f|D])$ .
predict_distseems to sample the posteriorSimilarly,$f|D$ (which incidentally is not necessarily a useful quantity in the Bayesian paradigm).
predictseems to give the median ofDid I misunderstand function or intent?