Michael Deistler

Would be great to have a tutorial that uses the ImportanceSamplingPosterior to perform asymptotically correct inference if the likelihood is available. ```python from torch import ones, eye import torch from...

Currently, the `VIPosterior` requires to be retrained for every `x_o`. We should have an option to make the `VIPosterior` a conditional density estimator that does not require retraining.

Training interface

1

In order to give more flexibility, we could expose the entire training loop to the user: ```python from sbi.inference import nre_loss, ratio_estimator_based_potential, MCMCPosterior net = classifier_nn("mdn") data_loader = ... for...

resume_training no longer working

1

AssertionError: You have already trained this neural network. After you had trained the network,

As raised in #1066 the `DensityEstimator`s currently have the following limitations: Not **all** `DensityEstimator`s can... - ...handle `batch_shapes` (but only `batch_dims`, i.e. scalar values) - ...evaluate one datapoint under multiple...

As all neural networks in `sbi`, the `MixedDensityEstimator` should also be allowed to have an `embedding_net`. This is especially critical if we intent to use the `MixedDensityEstimator` also to estimate...

Currently, the `build_categoricalmassestimator` does [not take arguments for z-scoring](https://github.com/sbi-dev/sbi/blob/0f0ee6e9f6c35a6d030281427409bcfacc0e4764/sbi/neural_nets/categorial.py#L12-L16). Instead, since it is only actively used in MNLE, `build_mnle` takes care of z-scoring by passing a standarization net to the...

The [README of the tutorials](https://github.com/sbi-dev/sbi/blob/main/tutorials/README.md) separates tutorials between users and contributors in a too strong way. In fact, all tutorials are for users, not for contributors. In addition, the README...

In #1066, I did not update the `log_prob_iid` method. Instead, MNLE currently uses the standard `log_prob()` to sample [here](https://github.com/sbi-dev/sbi/blob/0f0ee6e9f6c35a6d030281427409bcfacc0e4764/sbi/inference/potentials/likelihood_based_potential.py#L217-L221). For improved sampling speed we should revert to using `log_prob_iid()`