Initialization of Autoguide network parameters

If I use the AutoIAFNormal autoguide, what is the default initialization scheme of the autoregressive network parameters?
(I read the Initialization¶ docs, but I couldn’t understand how it reflects on networks).

Thanks :slight_smile:

cc @stefanwebb who knows that part of code.

Note I believe flowtorch has more thoughtful initialization.

Hi @stefanwebb, what is the default initialization scheme of weights and biases in autoregressive networks?
Does the initialization schemes in Initialization refer to weights and biases of the network or to samples that are used to calculate the initial KL divergence? if it is the latter, is the prior being the base distribution (standard normal)?