JuliaTrustworthyAI / LaplaceRedux.jl

Effortless Bayesian Deep Learning through Laplace Approximation for Flux.jl neural networks.
https://www.taija.org/LaplaceRedux.jl/
MIT License
39 stars 3 forks source link

Fixed tutorials code snippets #94

Closed pasq-cat closed 3 months ago

pasq-cat commented 4 months ago

there was a missing plot in the logit example page and the code snippets missed the listing the libraries needed to run. part of gsoc 24

pasq-cat commented 4 months ago

Thanks, Pasquale, but the Markdown documents are rendered through Quarto (**.qmd), so any changes should be made there. Will explain during our meeting this morning.

@pat-alt the page " a note on the priors" is incomplete and the code is hidden. Should i leave it as it is now or i can set the code to be visible?

ps. have you tried some other distribution for the priors? i have read that the normal distribution is not optimal and the student-t distribution gives better results. maybe we can test it later.

pat-alt commented 4 months ago

Feel free to unhide the code, if that makes it clearer.

ps. have you tried some other distribution for the priors? i have read that the normal distribution is not optimal and the student-t distribution gives better results. maybe we can test it later.

I'm not sure how this would work, since we always assume a multivariate Gaussian posterior (see Murphy (2022) chapter 4.6.8.2). But if you have an idea, happy to discuss.

pasq-cat commented 4 months ago

Feel free to unhide the code, if that makes it clearer.

ps. have you tried some other distribution for the priors? i have read that the normal distribution is not optimal and the student-t distribution gives better results. maybe we can test it later.

I'm not sure how this would work, since we always assume a multivariate Gaussian posterior (see Murphy (2022) chapter 4.6.8.2). But if you have an idea, happy to discuss.

yeah i had forgotten the laplace approx is a plugin approximation over a standard nn, rather than a fully bayesian nn.

pasq-cat commented 4 months ago

sorry for the delay but there was an installation issue with vs code jupyter and quarto and the render kept failing