rlouf/blog-benchmark-rwmetropolis

Code to reproduce the results of 👇

Rémi Louf / Twitter / Linkedin / Mastodon

I am a statistician (some say data scientist) and software developer. I am particularly interested in Bayesian statistics and generative modeling, MCMC samplers and symbolic programming.

I am the author of Blackjax and a core member of Aesara.

• ASK ME ANYTHING
• The typical set is an experiment in note taking and sharing (repo)

blog-benchmark-rwmetropolis's Issues

sign in logpdf

Dear contributors,

if I'm not mistaken, there is a typo in the sign of the mixture_logpdf(x) output. The last line of that function shouldn't have a minus, thus

`return logsumexp(np.log(weights) + log_probs, axis=0)`

Fortunately, the scale of your proposal distribution is so small that the acceptance rate is close to 1. Thus the sampler does not explode and your results shouldn't change much.

Best, AT

PS: in the blog to the notebook there is even a second minus sign in the same line, which shouldn't be there

Edit: As much as I see Colin Carroll doesn't have any of the minuses

mistake in the mixture log prob computation

The mixture log prob is evaluated on the whole observation, instead of each dimension:
instead of:

log_probs = np.array([dist_1(x[0]), dist_2(x[1]), dist_3(x[2]), dist_4(x[3])])

it should be

log_probs = np.array([dist_1(x), dist_2(x), dist_3(x), dist_4(x)])