Joshua G Albert
Joshua G Albert
Hi, I'm getting an error that I think could be fixed by replacing, `if self.values == None` with `if self.values is None`. ``` python Traceback (most recent call last): File...
Truncated normal gives wrong values sometimes. Seems to be when the scale is relatively small, but in surprising situations where you'd expect it to work like TruncatedNormal(1, 0.1, 0, 10)....
Adding the poisson quantile would be useful. As a use case, JAXNS uses quantiles to reparametrise. Below is a bisection approach accurate to rate of 1e4. ```python from functools import...
In LM method, the `max(diag(JTJ))` is used to set the damping factor. As per option 2 in https://github.com/google/jax/issues/19711 it can be made more efficient than currrently implemented. I discovered this...
# Description The `LevenbergMarquardt` implementation does not accept PyTree parameters, giving `TypeError: primal and tangent arguments to jax.jvp must have the same tree structure` at levenberg_marquardt.py, line 534. # MVCE...
The documentation says a different thing than code. Specifically, the `-` is inconsistent with the `+` in docstring at this part `(tree_l2_norm(params) - self.xtol)`. Docstring says: ``` the convergence is...
There are several works that introduce L-BFGS for the mini-batch setting. 1. [This one is used in astronomical radio interferometry calibration](https://arxiv.org/pdf/1904.05619.pdf). 2. [This is similar](https://ieeexplore.ieee.org/document/8626766). 3. [This one from the...
It's now 2024 and numerous science projects rely on the FFT. This repo is the first thing that pops up when Google'ing FFT performance. I think it would be very...
@JohannesBuchner doing cross comparison with mutliple NS implementations I found that upgrading to 4.1.6 from 3.5 decreased performance, and introduced strange performance scaling, e.g. `max_num_improvement_loops=0` sometimes more efficient that `max_num_improvement_loops=-1`,...
@JohannesBuchner we really need you to add ESS to the static implementation results, when `num_live_points>0` for cross comparison in a paper. Can you quickly add this, or point me how...