Implementation of quantiles for messenger guides [WIP]

[vitkl]

This PR contributes the implementation of quantiles for messenger guides - which resembles how this is done in AutoNormal. I see certain errors that I don't know how to address (https://github.com/pyro-ppl/pyro/runs/4490764371?check_suite_focus=true#step:5:22410) - namely, the messenger guide tries to expand returned quantiles new.v = self.v.expand(batch_shape + self.event_shape). This can potentially be addressed by running a for loop over requested quantiles like this:

def quantiles(self, quantiles, *args, **kwargs):
        self._computing_quantiles = True
        try:
            for i, q in enumerate(quantiles):
                self._quantile_values = q
                _quantiles_1 = self(*args, **kwargs)
                if i == 0:
                    _quantiles = {k: [v] for k, v in _quantiles_1.items()}
                else:
                    _quantiles_1
                    # add new quantile
                     _quantiles = {k: _quantiles[k] + [_quantiles_1[k]] for k in _quantiles.keys()}
               
            return  {k: torch.tensor(v, dtype=_quantiles_1[k].dtype, device=_quantiles_1[k].device) for k, v in _quantiles.items()}
        finally:
            self._computing_quantiles = False

Please let me know what you think @fritzo

[fritzo]

Thanks for sketching this @vitkl. The approach seems reasonable, but I'm unsure why the shapes are erroring.

[vitkl]

Is it possible to restart the checks? Maybe the issue resolved itself with new changes to pyro/pytorch?

[fritzo]

not sure how to restart in github actions, but you could merge Pyro dev branch and push, that should rerun ci