recski/py-wfsa-learner

A pure python learner of weighted finite state automata

py-wfsa-learner

A pure python learner of weighted finite state automata

py-wfsa-learner's Issues

make factor depend on actual temperature in learning

Now factor is only considered when accepting a change that ends up in a higher temperature, but not in actual learning steps

States should emit more symbols

States should be able to emit more than one symbols of an alphabet even if they are more than 1-length each

Quantizer with a not-power-of-two number of levels

Maybe this will be not so hard

learning steps based on quantizer

When quantizer is given, the learning step should be at least an interval given by the quantizer, or on higher temperatures, n * interval

Two-step learning with two quantizers using different "resolution"

Needs to be discussed. If the learning step enhancement (#4) is implemented, maybe this is done. But if so, maybe a dumping option is needed on changing the temperature

floating precision errors and dict keys

LogLinQuantizer calls my_round() but it doesn't return the representers but a very close number to one of them. Something needs to be figured out, until that, linear searching is being done when looking for a representer.

Reverting while learning

Changing and edge, and reverting the change if it didn't work out causes errors now since using quantizers, because after rounding/quantizing and renormalizing, and changing back, sometimes we don't get the original weights. Now the whole automaton.m dictionary gets copied again as in old code, but should be replaced with something else, at least with copying only those edges that got changed

almighty Automaton

Way too long class, some methods to different files