Posterior predictive check (plot_ppc) with dataframe encoding.

Hello everyone!

I have a model which has two cases enconded in df[‘membrane_enc’], so this should give two indepently modelled likelyhoods and posteriors. I train the model in this fashion:

shape = df['membrane_enc'].unique().shape
with pm.Model() as my_model:
    tau = pm.HalfNormal('tau', sigma=1.,shape =shape)
    tau_0 = pm.TruncatedNormal('tau_0', mu = min_val, sigma = 1, lower = 0, upper = min_val, shape =shape)
    like = DensityDist('like', logp_exp_lag(tau_0[df['membrane_enc']],tau[df['membrane_enc']]), observed=df['tpore'], random=custom_random, shape = (2,180))
with my_model:
    my_model_trace = pm.sample(4000, tune=3000,  random_seed=RANDOM_SEED)
    ppc = pm.sample_posterior_predictive(my_model_trace, var_names = ['like'],  random_seed=RANDOM_SEED)
    my_model_trace = az.from_pymc3(trace=my_model_trace, model=my_model, posterior_predictive=ppc)
    my_model_trace.to_netcdf(model_path)

The model converges nicely and gives reasonable posteriors for the two encodings. The problem comes when I try to plot the posterior predictive. I cannot find the way to plot the two distinct ppc’s in two graphs:

az.plot_ppc(my_model_trace, random_seed=RANDOM_SEED, var_names = ['like'], flatten = [])

This code fails and I am not sure which dimension I should flatten since there is only one in my_model_trace, like_dim_0 and it doesn’t work. All this code works nicely if you only work with one dimension.

Can someone please help me make this plotting work?

By the way, I am using arviz 0.10.0.

Thank you very much in advance! Best, Sergio

[sperezconesa]

Thanks for the pointers, now the code looks like:

my_model_trace = my_model_trace.assign_coords(like_dim_0=df['membrane_enc'].to_numpy(),
                                              groups='observed_data')
my_model_trace.posterior_predictive = my_model_trace.posterior_predictive.rename({'like_dim_1':'membrane_enc'})
my_model_trace.observed_data = my_model_trace.observed_data.rename({'like_dim_0':'membrane_enc'})
az.plot_ppc(my_model_trace,  flatten=['draw', 'chain', 'like_dim_0']) # flatten all dimension except the one we are interested in

I have added labels to `like_dim_1’ and renamed the coordinates that represent the encoding. Now I get the result I was looking for: Thanks a lot for your help!

[OriolAbril]

Glad I could be of help!