random_seed in sample_vp does not fix the results

[aloctavodia]

I am observing a weird behavior with random_seed in variational.sample_vp

with pm.Model() as model:
    mu = pm.Normal('mu', 0, 1)
    sd = pm.Uniform('sd', 0, 1)
    v_params = pm.variational.advi(n=1000, random_seed=1)
    start = pm.variational.sample_vp(v_params, 1, progressbar=False, hide_transformed=False,
                                     random_seed=1)
    print(v_params.means, start[0])

v_params.means is always

{'mu': array(0.0020075358649095402), 'sd_interval_': array(0.009411317153529386)}

but start[0] switch between two sets of values

{'sd': 0.15213536777333242, 'mu': -0.30831562426121062, 'sd_interval_': -1.7179502892676166}

or

{'sd': 0.41065146913471795, 'mu': -1.4440759392611413, 'sd_interval_': -0.36127289410088403}

[bhargavvader]

Does #1655 fix this?

[aloctavodia]

@bhargavvader no, it doesn't. AFAIK, the problem is still there. @ColCarroll probably has more to say about this.

[ColCarroll]

Can't reproduce any more!

@ferrine I think this is the current equivalent:

with pm.Model() as model:
    mu = pm.Normal('mu', 0, 1)
    sd = pm.Uniform('sd', 0, 1)
    v_params = pm.fit(1000, method='advi', random_seed=1)
    start = pm.sample_approx(v_params, 1)
    print(start[0], v_params.mean.eval())