Michael Deistler

Does multi-round SNPE work with RejectionPosterior as proposal?

1

Need to double check, have some worrying results that there might be a bug

How many copies of the dataset are created in our pipeline?

1

For very big datasets, it is undesirable to copy them. @tbmiller-astro suggested that some copies could be removed. Here are some candidates where we create local copies of the dataset:...

- should our tests be bimodal? - slow tests for all pyro samplers (hmc requires gradients...) - should we start thinking about a `v1.0.0` release? - progress bar for rejection...

Flow training in unconstrained space

5

This should happen only after we have moved to pyroflows

Preconfigured embedding_nets

10

1) We should provide a tutorial on how to use embedding nets. 2) We should have pre-configuered RNNs and CNNs. See also #162

Should we switch to pyro flows?

7

Currently, we are relying on `nflows` for our flows. I think that pyro's flows look pretty great, see e.g. [here](http://docs.pyro.ai/en/latest/distributions.html#spline) for NSF. They very nicely separate the transformer from the...

We might want to use `nsf` as default density estimator. It has almost always better performance at the cost of being more computationally expensive.

Deprecate the simple interface.

1

Currently, in our simple interface, we only allow very few kwargs. We could add two arguments `init_kwargs` and `call_kwargs`, which would both be dictionaries. Not sure if we want this...

Implement SNPE-B

7

SNPE-B is currently not implemented. When implementing it take care with the following: - we evaluate the importance weights always on the last posterior, no matter what round the samples...

`compose` creates empty figure

2

OS: Ubuntu 20.04 LTS python: 3.8 ### Setup I installed within a conda environment with: ``` sudo apt-get install libxml2-dev libxslt-dev pip3 install svgutils --user ``` ### Problem ```python #!/usr/bin/env...