Some minor changes

[aseyboldt]

• More info about tune argument in pm.sample docstring
• Cast metropolis scaling to double (ints lead to errors)
• Use mean of lengths of grads instead of mean of squared grads in advi+adapt_diag_grad.

[fonnesbeck]

Looks good. Merge when tests pass.

[aseyboldt]

The change of adapt_diag_grad improves it enough so that I get better eff samples / s with it than using the usual adapt_diag (in the hierarchical_partial_pooling notebook).

[junpenglao]

The change of adapt_diag_grad improves it enough so that I get better eff samples / s with it than using the usual adapt_diag (in the hierarchical_partial_pooling notebook).

We should put that into the benchmark models

[fonnesbeck]

Also, perhaps populate RELEASE-NOTES with a line of information about the changes, as we discussed at the project meeting.

[twiecki]

What's the reasoning behind the changes in adaption? Just works better empirically?

[aseyboldt]

@twiecki Yes, I tried a couple of variants and this worked best. Also a bit of intuition, taking the mean of the inverse variances is a bit strange, it puts a lot of weight on very large values.

[junpenglao]

Thanks @aseyboldt!