Run the grid-search with a dynamical model

[miguelgondu]

• [x] Confirm whether we can easily compute the metrics (coherence, log-perplexity and so on) using Blei's binary.
  • [x] If we use it, do we need to re-implement coherence and log-perplexity?
• [x] Re-implement the grid search using these metrics (or the ones from the seq. model).

[juanrloaiza]

Both gensim.models.LdaSeq and gensim.models.wrappers.dtmmodel implement this function for dtm_coherence. Seems that this is the input for gensim.models.coherencemodel and it seems easy to implement for the files we have. We should implement this function for our "wrapper" and test this.

[juanrloaiza]

We can compute model coherence using the same procedure mentioned above. This means we can use gensim.models.coherencemodel to compute this metric. However, we can only compute it for one time slice or for the averaged model. We should briefly check which alternative is better from a conceptual standpoint.

[miguelgondu]

• We can think about computing the mean coherence if computing the coherence per slice is not expensive.
• We can drop log-perplexity, just for convenience sake.

[miguelgondu]

@juanrloaiza did it, the legend.