Cross-likelihood validation of syllable uniqueness | Current implementation in moseq2_model.train.util.get_crosslikes

[X4ndri]

This is regarding Cross-likelihood validation as in Figure 2D in the 2015 paper. Where it shows how well the AR parameters of each state are used to explain every the observed variable during each other state. It is already implemented in moseq2-app @ moseq2_model.train.util.get_crosslikes . I can call this method manually as it only requires the arhmm object. However, each instance in the arhmm.states_list, that is, an instance of pyhsmm.internals.hmm_states.HMMStatesEigen corresponding to each session has its cache cleared, meaning that it does not contain the PC data used during training in its .data attribute.

To fix that, I did re-assigned the .data attribute manually for each session's instance. Briefly, I read in the PC scores corresponding to each session, then whiten it using the whitening parameters used during training, then perform AR striding using the used nlags then set it to the HMMStatesEigen.data attribute corresponding to that session. I can then call the get_crosslikes method and get a Cross Likelihood matrix.
However, I am concerned about the resulting matrix, because of the resulting range, I am getting a range of max of 0 and a min of ~-800; See below.