with pm.Model():
...
approx = pm.fit(method='advi')
After
with pm.Model():
...
approx = pm.fit(method='dadvi')
Context for the issue:
The deterministic ADVI appears to be a clear improvement on the original ADVI algorithm, both in terms of speed and accuracy. Having a look at the paper there should not be any obvious impediment to adding it to PyMC (though we may want it tied more closely to JAX than the current ADVI is).
Before
After
Context for the issue:
The deterministic ADVI appears to be a clear improvement on the original ADVI algorithm, both in terms of speed and accuracy. Having a look at the paper there should not be any obvious impediment to adding it to PyMC (though we may want it tied more closely to JAX than the current ADVI is).