Hey guys,
I believe there is a bug in the GaussianHMM distribution - particularly with the handling of the transition_distribution when there is a time dimension to the distribution. Let’s pretend the batch shape is empty. If you have a transition distribution that is unique along the time dimension (let’s pretend there are 100 time steps, and hidden dim of 5, so you pass in a transition distribution with, say, batch_shape of (100,5), and event shape of 5), the the distribution returned has batch shape of 100, and event shape of (100,5), which is incorrect. (it should be batch shape of empty, and even shape of (100,5).
Mike