Fix docs build

[gdalle]

Hi! I'm trying to include your package in a benchmark but the documentation does not seem up-to-date, perhaps because the recent builds have failed?

[lucasondel]

Yes, this package is most likely unstable and documentation not updated. Yet, the forward-backward implementation was quite efficient when running on GPU so I guess it would make sense to include it in your benchmark.

In a nutshell:

• create a fsm by specifying the initial weights, transition, final weights and state labels using the FSM constructor
• compile the fsm using the compile function. You need to provide a sparse matrix C mapping states to pdf (I allow states to share the same pdf)
• call pdfposteriors to run the forward backward

Things to keep in mind:

• my approach was designed for speech related problems with big markov chains (at least tens of thousand of states and more than 200k transitions)
• a big benefit of my implementation was to be able to compute the Forward-backward on multiple sequences with varying length
• I've never optimized the CPU version, my main pipeline was GPU-based implementation
• Looking at your code, I see that you use the rescaled version. I was running the forward-backward in the log-domain because the rescaled implementation is not very stable and I couln't ignore sequences giving me NaN. On CPU, the rescaled version is much faster than the log-domain one. However, the difference on GPU is not so significant (Nvidia GPU have dedicated instructions for log/exp function).

Hope that helps !