These are chat archives for ramin-git/rnn_traning

23rd
Apr 2015
ramin-git
@ramin-git
Apr 23 2015 04:36
without cudo there is need very more time. Did you test on BillionWords data without word tree?
If not, please test, it is interesting. Thanks
Nicholas Léonard
@nicholas-leonard
Apr 23 2015 13:54
Yeah, without softmaxtree, you need cuda. And yes I tested it with BillionWords. Without softmaxtree, even with cuda, it is extremely slow, and requires tons of memory.
ramin-git
@ramin-git
Apr 23 2015 19:23
i could train with small vacabulary on cuda
Nicholas Léonard
@nicholas-leonard
Apr 23 2015 19:31
yes that is recommended
You could start with the most frequent words, and group all remaining into <unk> word
ramin-git
@ramin-git
Apr 23 2015 19:34
yes i did that. Cut vacabulary by frequency word. And when generaty train, valid and test data word not exist in vacabulary wrote unk id