Closed szhan closed 3 months ago
I consulted with @astheeggeggs. I checked the log-likelihood of the most likely paths from the two implementations, and they passed np.allclose
. So, it seems that in this instance, there exists at least two equally likely paths, and different paths were chosen during the runs.
While I was developing some tests, I encountered the following discrepancy in the pointer matrices from
forwards_viterbi_hap_naive_low_mem_rescaling
andforwards_viterbi_hap_low_mem_rescaling
.I got the following pointer matrices:
I don't quite follow the logic for assigning pointers in the low-memory implementation, so I'm not sure the source of this discrepancy. I think the pointer matrix from the naive implementation is right, but I'm not sure about the other implementation.