Bug #829
pYIN explodes when presented with a longish file
| Status: | New | Start date: | 2013-12-07 | |
|---|---|---|---|---|
| Priority: | High | Due date: | ||
| Assignee: | - | % Done: | 0% | |
| Category: | - | |||
| Target version: | - |
Description
Just tested Tony with a few audio files on Windows, and found that it blows up during the getRemainingFeatures call ("Windows Runtime has requested" etc, so probably an exception) when asked to extract features from a file of around 10 minutes duration. It was fine with files of 3 minutes or so.
(The longer file was "Nothin But Time" by Cat Power, in case it turns out that something other than duration is the cause)
History
#1
Updated by Matthias Mauch over 10 years ago
I couldn't really reproduce the exception, but I see that in getRemainingFeatures, memory usage is very high (>1.4GB).
Now it's not super-straight-forward to get around that if we keep the HMM as it is because we store two matrices (one float, one int) that have nFrame x nState elements. That's 960 or so states, and approx. 100,000 frames in 10 minutes, i.e. 96 million elements -- is that around 360MB in floats just for the one? Well, anyway, lots.
Two ways around that:
- calculate observation probability on the fly -- that still leaves us with the int matrix ("psi" in the viterbi algorithm)
- do a local Viterbi (fixed-lag smoothing)
--- things to think about.