Trouble sampling from 1D MixtureGeneral

numpyro
21

Thank You so much for solving my problem! :two_hearts:

For me, in this concrete example:

  • both fixes are necessary: shifting the Pareto by an epsilon to the left, and choosing a different implementation of SkewNormal log prob
  • out of the three options for log(1 + erf(x)):
    1. the log1p_erf by @fehiepsi above
    2. jnp.log(jax.scipy.special.erfc(-x))
    3. dist.Normal().log_cdf(x * jnp.sqrt(2))
      the second runs about 30% faster than the other two.