Model fails to train on v0.6.0 with default settings having trained on v0.4.0

[joshnevin]

I have found that upgrading to v0.6.0 causes the following warning and then a breaking error, which I do not get using v0.4.0:

warning:
/home/zcicjne/dbpgpe/labgit/LabCode/gpenv/lib64/python3.6/site-packages/scipy/stats/_distn_infrastructure.py:1844: RuntimeWarning: divide by zero encountered in double_scalars
x = np.asarray((x - loc)/scale, dtype=dtyp)
Prior solver failed to converge

error:
File "/home/zcicjne/dbpgpe/labgit/LabCode/gpenv/lib64/python3.6/site-packages/scipy/stats/_distn_infrastructure.py", line 1844, in cdf                                   
    x = np.asarray((x - loc)/scale, dtype=dtyp)
FloatingPointError: divide by zero encountered in double_scalars

steps to reproduce:
kern = 'SquaredExponential'
gp_m = mogp_emulator.GaussianProcess(good_points_m, good_targets_m, kernel=kern)
gp_m = mogp_emulator.fit_GP_MAP(gp_m)

Data (attached):
inputs: good_points_m_issue.pkl
targets: good_targets_m_issue.pkl
data.zip

Environment results of pip list:
Package Version

---

mogp-emulator 0.6.0
numpy 1.19.5
patsy 0.5.2
pickle-mixin 1.0.2
pip 21.3.1
scipy 1.5.4
setuptools 39.2.0
six 1.16.0
Python 3.6.8

As a workaround, specifying the priors like this recovers the old behaviour from v0.4.0:
priors=mogp_emulator.Priors.GPPriors(nugget_type="adaptive",n_corr=3)
gp_m = mogp_emulator.GaussianProcess(good_points_m, good_targets_m, kernel=kern, priors=priors)
gp_m = mogp_emulator.fit_GP_MAP(gp_m)
i.e. the model fits fine when this is added.