Juan M. Cruz-Martinez

I guess the solution to this issue then is to find the right `covmat` function then and use that for the fit and the replicas.

Those are only for `nnfit` I'm afraid (for vp they are either outdated like `bigexp` or are not used at all, like the `sampling` flag)

For the thcovmat that's cared for in #1528, @andreab1997 didn't have to reinvent (many) wheels, only implement all the sampling flags that were actually only used in `nnfit`. The reordering...

@RoyStegeman let's wait a few weeks for finishing this, once I finish with the optimization of the n3fit it should be more robust for the implementation. Also, we need to...

Yes, the tensorflow package for mac is set to 2.0 because the conda packagers are not able to make anything beyond 2.0 work in Mac :__ (so conda is not...

As mentioned in the phone call this can be a good compromise solution where other backends can easily be implemented without permeating to the default conda package. If people like...

I think the problem people had was having the dependence and not the runcard option.

Ok, this is ready. @RoyStegeman please have a look and check the docs if you want to add more stuff. @Zaharid since I'm going the [_responsible user_](https://github.com/realpython/python-guide/pull/524) way I would...

To have the option to select backend you have to have more than 4 lines, of course. The extra backend itself is literally 5 statements: 2 imports, 2 optimizers, 1...

> However I'd say a bunch of context dependent imports is bringing in way too many footguns. I think having more than one backend means having context dependent imports. So,...