Initialzing NUTS with covariance matrix as mass matrix

Questions
4

Oh right I see, yeah that’s a tricky one, as sigma is a bounded random variable, and PyMC3 samples everything in continuous space, so you need to get the sigma after tranformed. Maybe try something like:

df = pd.DataFrame()
for myname in trace.varnames:
  if myname in ["alpha", "sigma_log__"]:
    k = pd.DataFrame(trace[500:][myname])
    print(myname)
    df = pd.merge(df, k, left_index=True, right_index=True, how='outer')

# var/cov matrix
covmatrix = np.cov(df.values)