Improving NUTS Runtime

[junpenglao]

@MrinankSharma

The model has a random walk in it; I would expect that this gives high correlations?

Not necessary - you can do x = pm.Normal(‘x’) and x_random_walk = tt.cumsum(x)

@MrinankSharma

Another question: is it possible to directly set the mass matrix, and/or estimate it from posterior samples?

Yes - for example Initialzing NUTS with covariance matrix as mass matrix