search

vrenkens / nabu

Code for end-to-end ASR with neural networks, build with TensorFlow
MIT License
108 stars 43 forks source link

your local BeamSearchDecoder implementation #28

Closed riyijiye closed 6 years ago

riyijiye commented 6 years ago

Hi Vincent,

I have not read into code detail of your beam_search_decoder.py implementation. but I am just wondering, to achieve beam search decoding, what is the difference between using your BeamSearchDecoder and the one provided as tf.contrib.seq2seq.BeamSearchDecoder. with my rough understanding about tf.contrib.seq2seq module, It looks passing tf.contrib.seq2seq.BeamSearchDecoder into tf.contrib.seq2seq.dynamic_decode gives decoding results based on beam search algorithm.

thanks, Harry

vrenkens commented 6 years ago

Hi Harry,

2 reasons why I have my own implementation: 1) When I first implemented this BeamSerachDecoder a long time ago, the contrib BeamSearchDecoder did not exist yet :) 2) Now that I have my own and it works I can do some custom stuff, like visualizing the attention. I tried doing the same thing with the contrib one but failed :).

Cheers

riyijiye commented 6 years ago

thanks Vincent!

by the way, I saw you have a timit_fisher config, is it for combined timit and fisher English corpus. if it is, do you mind sharing your best results for this?

thanks, Harry

vrenkens commented 6 years ago

No it's not :s it was a project of a master student to do transfer learning with Elastic Weight Consolidation using the fisher information matrix, hence the name