fungs / mglex Goto Github PK

Implement adaptive smoothing (aka Laplace smoothing) in all models so that unlikely classifications are always > 0.

We are lacking a submodel to encode information from assembly such as a graph or linked read information. Such a model will effectively transport scaffolding information from the assembler to MGLEX.

Model file compress with a factor of 10 to 20 with gzip, consider adding this directly to MGLEX or ensure streamability using, for instance, BASH syntax.

Logging support is incomplete/nonfunctional, this needs to be implemented using the python standard logging module.

The code is currently slow, we can do much better using

• removing assertions in non-debug mode
• scaling data at the beginning to avoid likelihood calculations
• speeding up vector arithmetics with more advanced python packages/methods
• avoiding the time for parsing the data over and over again

If all likelihoods are zero (log-likelihood is inf), the corresponding row becomes nan after posterior transformation. It should stay zero instead!

More details are needed.