Hi all, I found my mean_field approximation using mean_field = pm.fit(50000, obj_optimizer=pm.adam(learning_rate=0.01), callbacks=[pm.callbacks.CheckParametersConvergence(diff='absolute')]) Now I want to do sampling using this mean_field approximation as the starting point. n…

Sorry, I misread your original post. You need to supply a single point as start, so a dict of {varname:number} something like this: tr = mean_field.sample(100) start = {var_name:tr[var_name].mean() for var_name in tr.varnames} with model: trace = pm.sample(start=start)

Feed the pm.fit results to NUTS sampler

Questions

nkaimcaudle July 31, 2020, 12:10am 2

You probably need to input the start dict once for each chain, so try
pm.sample(..., start=[start, start, start, start])

Topic		Replies	Views
`start` argument with multiple chains? Questions	1	544	November 26, 2018
Initialize Once for Multiple Runs? Questions	2	722	January 10, 2020
Differenc between PyMC 3.1 and PyMC 3.0 on NUTS Questions	1	1297	August 28, 2017
Pymc3 DEMetroplis Initialization Questions	1	598	May 17, 2019
Sample method of pymc3 Questions	4	685	July 15, 2018

Feed the pm.fit results to NUTS sampler

Related topics