From 49caeba74bd90f359e3d0831df5f4e3613b03e99 Mon Sep 17 00:00:00 2001 From: Determinant Date: Tue, 25 Aug 2015 11:10:47 +0800 Subject: allow proceeding global transf without frm ext --- speech_utils/init.lua | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'speech_utils') diff --git a/speech_utils/init.lua b/speech_utils/init.lua index 3500eb4..a2d1dc9 100644 --- a/speech_utils/init.lua +++ b/speech_utils/init.lua @@ -4,7 +4,7 @@ function nerv.speech_utils.global_transf(feat_utter, global_transf, frm_ext, frm_trim, gconf) local rearranged feat_utter = gconf.cumat_type.new_from_host(feat_utter) - if frm_ext > 0 then + if frm_ext ~= nil and frm_ext > 0 then local step = frm_ext * 2 + 1 -- expand the feature local expanded = gconf.cumat_type(feat_utter:nrow(), feat_utter:ncol() * step) -- cgit v1.2.3-70-g09d2 From 6495879e76dc0a2efe7b214b58d1d73a4ed3a931 Mon Sep 17 00:00:00 2001 From: Determinant Date: Wed, 26 Aug 2015 14:25:56 +0800 Subject: remove global_transf in IO modules; see asr_trainer.lua --- htk_io/init.lua | 7 ------- kaldi_io/init.lua | 15 ++++++--------- speech_utils/init.lua | 6 +++--- 3 files changed, 9 insertions(+), 19 deletions(-) (limited to 'speech_utils') diff --git a/htk_io/init.lua b/htk_io/init.lua index c4dfff9..b836e15 100644 --- a/htk_io/init.lua +++ b/htk_io/init.lua @@ -6,7 +6,6 @@ function TNetReader:__init(global_conf, reader_conf) self.feat_id = reader_conf.id self.frm_ext = reader_conf.frm_ext self.gconf = global_conf - self.global_transf = reader_conf.global_transf self.debug = global_conf.debug if self.debug == nil then self.debug = false @@ -31,12 +30,6 @@ function TNetReader:get_data() local res = {} -- read HTK feature local feat_utter = self.feat_repo:cur_utter(self.debug) - -- global transf - if self.global_transf ~= nil then - feat_utter = nerv.speech_utils.global_transf(feat_utter, - self.global_transf, self.frm_ext, - self.frm_ext, self.gconf) - end res[self.feat_id] = feat_utter -- add corresponding labels for id, repo in pairs(self.lab_repo) do diff --git a/kaldi_io/init.lua b/kaldi_io/init.lua index a1496c4..930705e 100644 --- a/kaldi_io/init.lua +++ b/kaldi_io/init.lua @@ -6,7 +6,6 @@ function KaldiReader:__init(global_conf, reader_conf) self.feat_id = reader_conf.id self.frm_ext = reader_conf.frm_ext self.gconf = global_conf - self.global_transf = reader_conf.global_transf self.debug = global_conf.debug if self.debug == nil then self.debug = false @@ -39,28 +38,26 @@ function KaldiReader:get_data() -- read Kaldi feature local feat_utter = self.feat_repo:cur_utter(self.debug) -- global transf - local transformed = nerv.speech_utils.global_transf(feat_utter, - self.global_transf, self.frm_ext, 0, self.gconf) - res[self.feat_id] = transformed + res[self.feat_id] = feat_utter -- add corresponding labels for id, repo in pairs(self.lab_repo) do local lab_utter = repo:get_utter(self.feat_repo, - transformed:nrow(), + feat_utter:nrow(), self.debug) res[id] = lab_utter end -- add corresponding lookup features for id, repo in pairs(self.lookup_repo) do local lookup_utter = repo:get_utter(self.feat_repo, - transformed:nrow(), + feat_utter:nrow(), self.debug) local nrow = lookup_utter:nrow() - if nrow < transformed:nrow() then + if nrow < feat_utter:nrow() then -- repeat the last frame - local nlu = lookup_utter:create(transformed:nrow()) + local nlu = lookup_utter:create(feat_utter:nrow()) nlu:copy_from(lookup_utter, 0, nrow) row = nlu[nrow - 1] - for i = 0, transformed:nrow() - nrow - 1 do + for i = 0, feat_utter:nrow() - nrow - 1 do nlu:copy_from(nlu, nrow - 1, nrow, nrow + i) end lookup_utter = nlu diff --git a/speech_utils/init.lua b/speech_utils/init.lua index a2d1dc9..7d5651f 100644 --- a/speech_utils/init.lua +++ b/speech_utils/init.lua @@ -3,7 +3,7 @@ nerv.speech_utils = {} function nerv.speech_utils.global_transf(feat_utter, global_transf, frm_ext, frm_trim, gconf) local rearranged - feat_utter = gconf.cumat_type.new_from_host(feat_utter) +-- feat_utter = gconf.cumat_type.new_from_host(feat_utter) if frm_ext ~= nil and frm_ext > 0 then local step = frm_ext * 2 + 1 -- expand the feature @@ -26,8 +26,8 @@ function nerv.speech_utils.global_transf(feat_utter, global_transf, global_transf:init(input[1]:nrow()) global_transf:propagate(input, output) -- trim frames - expanded = gconf.mmat_type(output[1]:nrow() - frm_trim * 2, output[1]:ncol()) - output[1]:copy_toh(expanded, frm_trim, feat_utter:nrow() - frm_trim) + expanded = gconf.cumat_type(output[1]:nrow() - frm_trim * 2, output[1]:ncol()) + expanded:copy_fromd(output[1], frm_trim, feat_utter:nrow() - frm_trim) collectgarbage("collect") return expanded end -- cgit v1.2.3-70-g09d2