He Jia

ENH: stats.gaussian_kde: add method that returns marginal distribution

10

Following the discussion at #9906, added `pdf_marginal` and `logpdf_marginal` to `scipy.stats.kde`, which are useful for visualizing high dimensional distributions.

limit the # of threads for jax

2

Hi, I know this is probably more of an issue on the jax side and has been discussed there, e.g. google/jax#743, google/jax#1539 and google/jax#6790, although I'm still wondering if you...

document that lower and upper bounds in Bisection need stop_gradient for differentiation

7

Hello, I'm trying to work with the following snippet, ``` import numpy as np import jax.numpy as jnp import jax from jaxopt import Bisection @jax.jit def _xy_c(r, phi, spin, theta_o):...

Use custom dataloader for training

5

Hello, is there a straightforward way to use custom pytorch `DataLoader` for training, instead of the `append_simulations` method? I didn't find this information in the documentation. This will allow us...

Seems that for now we'd better use numpy>=1.20 and numdifftools

set logp=-inf if the evaluation fails

1

Sometimes the target density is not defined in some regions of the parameter space. Maybe we should let BF automatically return logp=-inf for such cases?

Now the SIT transformer uses `use_parallel=False` by default. Need to review this at some point.

This should be easy, as we have already separated the pickle-able and (possible) non-pickle-able components.

1. to compute & extract additional parameters, like sigma8, using the true model 2. to allow additional importance weights from some other function (like the `post_BAO` in Planck), but maybe...

BF needs ~1s to import at the moment. Should study why it takes such a long time to import at some point.