Blind output with user-provided offsets

[SebastianBocquet]

Let the user provide a set of parameter offsets for the purpose of blinding the output. This makes looking at the chains safe!

This is activated by setting apply_blinding_offsets = True in the [output] section of the parameter file. The offsets are saved as a .npy file. Add blinding_offsets = /path/to/file to the parameter file.

Caveat: The user must set the offsets corresponding to output columns such as prior like post weight to 0 manually.

The modification is mostly copied from @MCostanzi.

[SebastianBocquet]

Sorry, should have run the tests offline first... Sampling works but the tests don't. I'll take a closer look.

[joezuntz]

It looks like this is because the blinding is not working for the in-memory output format.

[SebastianBocquet]

I applied the same modification to in_memory_output.py and the tests now pass. However, my changes to output_base will still break other output methods (which are not covered by the tests).

[SebastianBocquet]

Please let me know if you like the modification / blind scheme at all. If so, I can modify the other output options, too. If not, there's no point.

[SebastianBocquet]

It just occurred to me that offsetting the output makes resuming a chain from that output impossible. I guess cosmosis could unblind the chain internally and resume from there. I haven't looked at the code yet. We should probably fix that before merging.

[joezuntz]

Ouch, yes, this is difficult!

We could unblind on loading the chain in the resume method in the samplers that support resuming, but then we'd also have to check how that fits with the convergence diagnostics for the metropolis sampler. This might be a bit too complicated in terms of maintenance.

[SebastianBocquet]

I could add an error when trying to resume with blinding_offsets. Then the user can decide to unblind the chain if a resume is necessary. Not ideal, this only gets us half-way to where we want to be... What do you think?

[joezuntz]

It feels like we should do something a bit more organized, and modify the way that the chains are being resumed by adding a method to read a chain and detect if it has been blinded by looking at the header / metadata. Let me think about this!