I found that changing the marathon_model sampling initializer to "advi+adapt_diag" fit the model far faster. ADVI took ~12,000 steps to converge and then NUTS took about 3 minutes (my computer is not very fast, so I suspect you and most others would see far faster times). I doubt this is new information for you, but I figured it would be useful since this model is probably the slowest fitting one in the notebooks (which is why you didn't fit it live in your lecture at PyData 2019).
I found that changing the
marathon_modelsampling initializer to"advi+adapt_diag"fit the model far faster. ADVI took ~12,000 steps to converge and then NUTS took about 3 minutes (my computer is not very fast, so I suspect you and most others would see far faster times). I doubt this is new information for you, but I figured it would be useful since this model is probably the slowest fitting one in the notebooks (which is why you didn't fit it live in your lecture at PyData 2019).