WFSA training with Sautrela

In order to be trainable, a WFSA must implement a simple interface. Once implemented, the Trainer module is responsible of the rescoring of transition probabilities, whereas the model is responsible of translating the rescored probabilities to its internal representation. The Trainable interface consist of four methods used to initialize, do the transmission of the training counts and dump the resulting parameters. The counts are externally obtained, whereas the model is just responsible of translating those counts to its internal representation.

Training.png

All designed models (LMM inclusive) do implement this interface, thus, they can be trained using the Trainer module (there is no need of external training). Well known training criteria can be used:

  • Maximum Likelihood (Baum Welch)
  • Best Path Maximum Likelihood (Viterbi)

  • Maximum Mutual Information

This topic: Sautrela > WebHome > TrainingWithSautrela
History: r4 - 30 May 2007 - 12:48:55 - MikelPenagarikano
 
Copyright Mikel PenagarikanoValid XHTML 1.0 Transitional