jean-eric

But then it seems that the muti-rounds is not so good (here is for 10 rounds) 

Yes, I anderstand... I have increased to round trip number of simul 200->1000 and the final sampling I use 10_000 samples. And I use the observation as the red dots...

And finally using 100 obs around the true parameters, leads to the shrinkage towards these true parameters in the right amount (here the red dot is the true parameter location)...

Hi, a small add: `NeuralPosteriorEnsemble`is not found in `sbi.utils` but in `sbi.utils.posterior_ensemble`. May be missing in `__init__.py`?