Weird behaviour in terms of training the normalizing flow.

[kazewong]

There is a report on the following breaking mode:

When analyzing GW170817 as shown in the example, if the prior of cos(iota) is changed to [-1,-0.9], then choose a number of chain = 800 will give non-sensible result, such as nan in the log_prob.

Further investigation shows number of chains = 601, 801 will also give the same behavior. But 100, 500, 501 performs normally.

This is very weird, which I think it has to do with the training of the normalizing flow and thinning.