HMM labeled training issue

[AKuederle]

@nilsroth I think there was a bug in the training code for labeled training of the combined model

In this code snippet, you are attempting to set the only the edges that are not there yet.
However, you assume that the e.g. the entry for state transition 15 to 16 is at position (15, 16) in the transition matrix. However, because of the "sorting" bug in pomegranate this is not the case.
In result, some edges are not written and some edges are overwritten, eventhough they exist already.

gaitmap/gaitmap/future/hmm/segmentation_model.py

Lines 281 to 285 in ff2a333

	# Add missing transitions which will "connect" transition-hmm and stride-hmm
	for trans in transitions:
	# if edge already exists, skip
	if not model_untrained.dense_transition_matrix()[trans[0], trans[1]]:
	add_transition(model_untrained, ["s%d" % (trans[0]), "s%d" % (trans[1])], 0.1)

I fixed that by getting the existing transitions based on the actual statenames and checking based on that.

Could you check, if the same issue exists in your old training code?

With my fix, the train-hmm example now produces an actual usable HMM :)

[AKuederle]

Here is the updated code:

gaitmap/gaitmap_mad/gaitmap_mad/stride_segmentation/hmm/_segmentation_model.py

Lines 585 to 591 in e4449da

	existing_transitions = {(start.name, end.name) for start, end in new_model.graph.edges()}
	missing_transitions = transitions - existing_transitions
	# Add missing transitions which will "connect" transition-hmm and stride-hmm
	# We initialize with a very small probability, so that the model can learn the correct values in the next
	# step.
	for trans in missing_transitions:
	add_transition(new_model, trans, 0.1)

I also changed extract_transitions_starts_stops_from_hidden_state_sequence to return a set with the actual state names and not the state indices to make that work.

Still a little bit scary that we depend on this default naming of pg... But it is what it is...