DiffstarPop currently accepts a jax.random.key in order to generate a SFH population. It would be helpful for the API to include support for using a precomputed set of randoms, e.g., drawn in advance from a multivariate normal with the identity matrix for covariance (for the case of main sequence and quenched sequence) and uniform random (for the case of mc_is_q). This API modification would round out support within diffsky for fully pre-computed randoms, which allows for greater control over stochasticity in our gradient descent and HMCs, e.g., when running gradient descent in epochs, or eliminating MC noise in our posteriors.
DiffstarPop currently accepts a
jax.random.keyin order to generate a SFH population. It would be helpful for the API to include support for using a precomputed set of randoms, e.g., drawn in advance from a multivariate normal with the identity matrix for covariance (for the case of main sequence and quenched sequence) and uniform random (for the case ofmc_is_q). This API modification would round out support within diffsky for fully pre-computed randoms, which allows for greater control over stochasticity in our gradient descent and HMCs, e.g., when running gradient descent in epochs, or eliminating MC noise in our posteriors.