The Infinite Mixture Models tutorial error

[patwa67]

It is not possible to use:

k = map(t -> length(unique(chain[:z].value[t,:,:])), 1:iterations);

it throws the error:
ERROR: ArgumentError: index z not found

[devmotion]

This example uses an old version of MCMCChains. With newer versions you have to use

zvars = namesingroup(chain, :z)
k = map(t -> length(unique(chain[t, zvars, :])), 1:iterations)

(or something similar based on namesingroup, the code is untested). The issue will be fixed by the reorganization and updates in TuringLang/docs#113.

[patwa67]

No it doesn't work. The following works, but it just produces a vector of ones:

k = map(t -> length(unique(get_params(chain[t,zvars,:]))), 1:iterations);

[devmotion]

I assume the problem is that you use get_params: it returns a NamedTuple of parameters (see https://turinglang.github.io/MCMCChains.jl/dev/chains/#MCMCChains.get_params-Tuple{Chains}). I guess

k = map(t -> length(unique(Array(chain[t, zvars, :]))), 1:iterations)

is better (I assumed, apparently incorrectly since it failed (?), that chain[t, zvars, :] returns already an array).

[patwa67]

That works better, but what is the logic behind this array construction?

[devmotion]

Array(chain) returns an array of the values in the chain, where as default the different chains are appended (i.e., the output has #iterations * #chains rows and #parameters columns): https://turinglang.github.io/MCMCChains.jl/dev/getting-started/#Exporting-Chains

[patwa67]

OK, fine. I have run this example with some different settings and samplers. One strange thing occurs when I change from the SMC() (which mixes badly) to the PG() sampler. When I use SMC() the output in zvars and consequently the k variable corresponds well with the output in the locations variable muvars = namesingroup(chain, :μ), i.e. there is one location for each cluster. However, when the PG() sampler is used, this is no longer the case, i.e. there are more locations than clusters. However, the number of clusters seems to be correct.