зеркало из https://github.com/mozilla/kaldi.git
sandbox/online: merging changes from trunk
git-svn-id: https://svn.code.sf.net/p/kaldi/code/sandbox/online@4261 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
This commit is contained in:
Dan Povey
Коммит
0a6f544715
|
|
@ -17,6 +17,7 @@ bnf_num_gauss_sgmm=50000 # use fewer SGMM sub-states than the
|
|||
# non-bottleneck system (which has 80000).
|
||||
bnf_decode_acwt=0.066666
|
||||
|
||||
|
||||
# DNN hybrid system training parameters
|
||||
dnn_num_hidden_layers=4
|
||||
dnn_input_dim=4000
|
||||
|
|
@ -48,9 +49,16 @@ if [[ `hostname` == *.tacc.utexas.edu ]] ; then
|
|||
sgmm_train_extra_opts=( )
|
||||
sgmm_group_extra_opts=( --num_iters 25 )
|
||||
sgmm_denlats_extra_opts=( --num-threads 2 )
|
||||
sgmm_mmi_extra_opts=(--cmd "local/lonestar.py -pe smp 2")
|
||||
dnn_denlats_extra_opts=( --num-threads 2 )
|
||||
dnn_parallel_opts="-l gpu=1"
|
||||
|
||||
dnn_cpu_parallel_opts=(--minibatch-size 128 --max-change 10 --num-jobs-nnet 8 --num-threads 16 \
|
||||
--parallel-opts "-pe smp 16" )
|
||||
dnn_gpu_parallel_opts=(--minibatch-size 512 --max-change 40 --num-jobs-nnet 8 --num-threads 1)
|
||||
|
||||
dnn_gpu_mpe_parallel_opts=(--num-jobs-nnet 8 --num-threads 1)
|
||||
dnn_gpu_mpe_parallel_opts=(--num-jobs-nnet 8 --num-threads 1)
|
||||
dnn_parallel_opts="-l gpu=1"
|
||||
else
|
||||
decode_extra_opts=(--num-threads 6 --parallel-opts "-pe smp 6 -l mem_free=4G,ram_free=0.7G")
|
||||
sgmm_train_extra_opts=( --num-iters 25 )
|
||||
|
|
|
|||
|
|
@ -49,14 +49,15 @@ dnn_update_egs_opts=(--weight-threshold 0.7 --splice-width 4 --samples-per-iter
|
|||
|
||||
if [[ `hostname` == *.tacc.utexas.edu ]] ; then
|
||||
decode_extra_opts=( --num-threads 4 --parallel-opts "-pe smp 4" )
|
||||
sgmm_train_extra_opts=( )
|
||||
sgmm_train_extra_opts=( --num-iters 25 )
|
||||
sgmm_group_extra_opts=( )
|
||||
sgmm_denlats_extra_opts=( --num-threads 1 )
|
||||
dnn_denlats_extra_opts=( --num-threads 1 )
|
||||
|
||||
dnn_cpu_parallel_opts=(--minibatch-size 128 --num-jobs-nnet 8 --num-threads 16 \
|
||||
--parallel-opts "-pe smp 16" )
|
||||
dnn_gpu_parallel_opts=(--minibatch-size 512 --max-change 40 --num-jobs-nnet 4 --num-threads 1)
|
||||
dnn_gpu_parallel_opts=(--minibatch-size 512 --max-change 40 --num-jobs-nnet 4 --num-threads 1
|
||||
--parallel-opts "-pe smp 16" )
|
||||
|
||||
dnn_gpu_mpe_parallel_opts=(--num-jobs-nnet 4 --num-threads 1)
|
||||
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev/IARPA-babel102b-v0.5a_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=24
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev/IARPA-babel102b-v0.5a_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_102/conversational/eval/
|
||||
eval_data_list=/export/babel/data/splits/Assamese_Babel102/eval.list
|
||||
|
|
@ -46,19 +49,20 @@ shadow_data_dir=(
|
|||
/export/babel/data/102-assamese/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_102/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Assamese_Babel102/uem/102-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Assamese_Babel102/dev.list
|
||||
/export/babel/data/splits/Assamese_Babel102/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Assamese_Babel102/uem/102-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
||||
# Acoustic model parameters
|
||||
numLeavesTri1=1000
|
||||
numGaussTri1=10000
|
||||
|
|
@ -74,7 +78,6 @@ numGaussUBM=800
|
|||
numLeavesSGMM=10000
|
||||
numGaussSGMM=80000
|
||||
|
||||
|
||||
# Lexicon and Language Model parameters
|
||||
oovSymbol="<unk>"
|
||||
lexiconFlags="--romanized --oov <unk>"
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev/IARPA-babel102b-v0.5a_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=24
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev/IARPA-babel102b-v0.5a_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_102/conversational/eval/
|
||||
eval_data_list=/export/babel/data/splits/Assamese_Babel102/eval.list
|
||||
|
|
@ -46,19 +49,20 @@ shadow_data_dir=(
|
|||
/export/babel/data/102-assamese/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_102/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Assamese_Babel102/uem/102-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Assamese_Babel102/dev.list
|
||||
/export/babel/data/splits/Assamese_Babel102/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Assamese_Babel102/uem/102-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel102b-v0.5a_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
||||
unsup_data_dir=(/export/babel/data/102-assamese//release-current/conversational/training/
|
||||
/export/babel/data/102-assamese//release-current/conversational/untranscribed-training/
|
||||
)
|
||||
|
|
@ -83,7 +87,6 @@ numGaussUBM=750
|
|||
numLeavesSGMM=5000
|
||||
numGaussSGMM=18000
|
||||
|
||||
|
||||
# Lexicon and Language Model parameters
|
||||
oovSymbol="<unk>"
|
||||
lexiconFlags="--romanized --oov <unk>"
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev/IARPA-babel103b-v0.4b_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=12
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev/IARPA-babel103b-v0.4b_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_103/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Bengali_Babel103//eval.list
|
||||
|
|
@ -46,15 +49,17 @@ shadow_data_dir=(
|
|||
/export/babel/data/103-bengali/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_103/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Bengali_Babel103/uem/103-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Bengali_Babel103/dev.list
|
||||
/export/babel/data/splits/Bengali_Babel103/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Bengali_Babel103/uem/103-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
|
@ -73,7 +78,6 @@ numGaussUBM=800
|
|||
numLeavesSGMM=10000
|
||||
numGaussSGMM=80000
|
||||
|
||||
|
||||
# Lexicon and Language Model parameters
|
||||
oovSymbol="<unk>"
|
||||
lexiconFlags="--romanized --oov <unk>"
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev/IARPA-babel103b-v0.4b_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=12
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev/IARPA-babel103b-v0.4b_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_103/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Bengali_Babel103//eval.list
|
||||
|
|
@ -46,15 +49,17 @@ shadow_data_dir=(
|
|||
/export/babel/data/103-bengali/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_103/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Bengali_Babel103/uem/103-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Bengali_Babel103/dev.list
|
||||
/export/babel/data/splits/Bengali_Babel103/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Bengali_Babel103/uem/103-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel103b-v0.4b_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
|
@ -82,7 +87,6 @@ numGaussUBM=750
|
|||
numLeavesSGMM=5000
|
||||
numGaussSGMM=18000
|
||||
|
||||
|
||||
# Lexicon and Language Model parameters
|
||||
oovSymbol="<unk>"
|
||||
lexiconFlags="--romanized --oov <unk>"
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev/IARPA-babel201b-v0.2b_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[LimitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=20
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev/IARPA-babel201b-v0.2b_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[LimitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_201/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Haitian_Babel201//eval.list
|
||||
|
|
@ -46,15 +49,17 @@ shadow_data_dir=(
|
|||
/export/babel/data/201-haitian/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_201/conversational/eval
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Haitian_Babel201/uem/201-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Haitian_Babel201/dev.list
|
||||
/export/babel/data/splits/Haitian_Babel201/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Haitian_Babel201/uem/201-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev/IARPA-babel201b-v0.2b_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[LimitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=20
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev/IARPA-babel201b-v0.2b_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[LimitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_201/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Haitian_Babel201//eval.list
|
||||
|
|
@ -46,15 +49,17 @@ shadow_data_dir=(
|
|||
/export/babel/data/201-haitian/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_201/conversational/eval
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Haitian_Babel201/uem/201-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Haitian_Babel201/dev.list
|
||||
/export/babel/data/splits/Haitian_Babel201/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Haitian_Babel201/uem/201-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel201b-v0.2b_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev/IARPA-babel203b-v3.1a_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=18
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev/IARPA-babel203b-v3.1a_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_203/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Lao_Babel203//eval.list
|
||||
|
|
@ -46,27 +49,20 @@ shadow_data_dir=(
|
|||
/export/babel/data/203-lao/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_203/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Lao_Babel203/uem/203-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Lao_Babel203/dev.list
|
||||
/export/babel/data/splits/Lao_Babel203/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Lao_Babel203/uem/203-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
unsup_data_dir=(/export/babel/data/203-lao/release-current/conversational/training/
|
||||
/export/babel/data/203-lao/release-current/conversational/untranscribed-training/
|
||||
)
|
||||
unsup_data_list=(
|
||||
/export/babel/data/splits/Lao_Babel203/train.LimitedLP.untranscribed.list
|
||||
/export/babel/data/splits/Lao_Babel203/train.untranscribed.list
|
||||
)
|
||||
unsup_nj=64
|
||||
|
||||
# Acoustic model parameters
|
||||
numLeavesTri1=1000
|
||||
numGaussTri1=10000
|
||||
|
|
|
|||
|
|
@ -15,7 +15,8 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev/IARPA-babel203b-v3.1a_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=18
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev/IARPA-babel203b-v3.1a_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_203/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Lao_Babel203//eval.list
|
||||
|
|
@ -46,15 +49,17 @@ shadow_data_dir=(
|
|||
/export/babel/data/203-lao/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_203/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Lao_Babel203/uem/203-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Lao_Babel203/dev.list
|
||||
/export/babel/data/splits/Lao_Babel203/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Lao_Babel203/uem/203-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel203b-v3.1a_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
|
|
|||
|
|
@ -14,9 +14,12 @@ dev2h_stm_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev
|
|||
dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.scoring.ecf.xml
|
||||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/splits/Tamil_Babel204/IARPA-babel204b-v1.1b_conv-dev.radical.kwlist.xml
|
||||
#dev2h_more_kwlists=(
|
||||
# [limitedLP]=
|
||||
# )
|
||||
dev2h_more_kwlists=(
|
||||
[bbn1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist.xml
|
||||
[bbn2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist2.xml
|
||||
[ibm1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist3.xml
|
||||
[ibm2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=18
|
||||
|
||||
|
|
@ -33,19 +36,41 @@ dev10h_more_kwlists=(
|
|||
[bbn2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist2.xml
|
||||
[ibm1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist3.xml
|
||||
[ibm2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist4.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist5.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
unsup_data_dir=(/export/babel/data/204-tamil/release-current/conversational/training/
|
||||
/export/babel/data/204-tamil/release-current/conversational/untranscribed-training/
|
||||
)
|
||||
unsup_data_cmudb=/export/babel/data/splits/Tamil_Babel204/uem/db-unsup-jhuseg-v8-utt.dat
|
||||
unsup_data_list=(
|
||||
/export/babel/data/splits/Tamil_Babel204/train.LimitedLP.untranscribed.list
|
||||
/export/babel/data/splits/Tamil_Babel204/train.untranscribed.list
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/204-tamil/release-current/conversational/eval/
|
||||
eval_data_list=/export/babel/data/splits/Tamil_Babel204/eval.list
|
||||
eval_data_cmudb=/export/babel/data/splits/Tamil_Babel204/uem/db-shadow-jhuseg-v8-utt.dat
|
||||
eval_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.scoring.ecf.xml
|
||||
eval_kwlist_file=/export/babel/data/splits/Tamil_Babel204/IARPA-babel204b-v1.1b_conv-dev.radical.kwlist.xml
|
||||
eval_nj=64
|
||||
|
||||
#Shadow data files
|
||||
shadow_data_dir=(
|
||||
/export/babel/data/204-tamil/release-current/conversational/dev/
|
||||
/export/babel/data/204-tamil/release-current/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Tamil_Babel204/uem/204-shadow-v0-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Tamil_Babel204/dev.list
|
||||
/export/babel/data/splits/Tamil_Babel204/eval.list
|
||||
)
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.scoring.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/splits/Tamil_Babel204/IARPA-babel204b-v1.1b_conv-dev.radical.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[bbn1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist.xml
|
||||
[bbn2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist2.xml
|
||||
[ibm1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist3.xml
|
||||
[ibm2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist4.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist5.xml
|
||||
)
|
||||
unsup_nj=64
|
||||
shadow_nj=64
|
||||
|
||||
|
||||
|
||||
# Acoustic model parameters
|
||||
numLeavesTri1=1000
|
||||
|
|
|
|||
|
|
@ -14,9 +14,12 @@ dev2h_stm_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev
|
|||
dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.scoring.ecf.xml
|
||||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/splits/Tamil_Babel204/IARPA-babel204b-v1.1b_conv-dev.radical.kwlist.xml
|
||||
#dev2h_more_kwlists=(
|
||||
# [limitedLP]=
|
||||
# )
|
||||
dev2h_more_kwlists=(
|
||||
[bbn1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist.xml
|
||||
[bbn2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist2.xml
|
||||
[ibm1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist3.xml
|
||||
[ibm2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=18
|
||||
|
||||
|
|
@ -33,10 +36,41 @@ dev10h_more_kwlists=(
|
|||
[bbn2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist2.xml
|
||||
[ibm1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist3.xml
|
||||
[ibm2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist4.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist5.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/204-tamil/release-current/conversational/eval/
|
||||
eval_data_list=/export/babel/data/splits/Tamil_Babel204/eval.list
|
||||
eval_data_cmudb=/export/babel/data/splits/Tamil_Babel204/uem/db-shadow-jhuseg-v8-utt.dat
|
||||
eval_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.scoring.ecf.xml
|
||||
eval_kwlist_file=/export/babel/data/splits/Tamil_Babel204/IARPA-babel204b-v1.1b_conv-dev.radical.kwlist.xml
|
||||
eval_nj=64
|
||||
|
||||
#Shadow data files
|
||||
shadow_data_dir=(
|
||||
/export/babel/data/204-tamil/release-current/conversational/dev/
|
||||
/export/babel/data/204-tamil/release-current/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Tamil_Babel204/uem/204-shadow-v0-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Tamil_Babel204/dev.list
|
||||
/export/babel/data/splits/Tamil_Babel204/eval.list
|
||||
)
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev/IARPA-babel204b-v1.1b_conv-dev.scoring.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/splits/Tamil_Babel204/IARPA-babel204b-v1.1b_conv-dev.radical.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[bbn1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist.xml
|
||||
[bbn2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist2.xml
|
||||
[ibm1]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist3.xml
|
||||
[ibm2]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist4.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel204b-v1.1b_conv-dev.kwlist5.xml
|
||||
)
|
||||
shadow_nj=64
|
||||
|
||||
|
||||
unsup_data_dir=(/export/babel/data/204-tamil/release-current/conversational/training/
|
||||
/export/babel/data/204-tamil/release-current/conversational/untranscribed-training/
|
||||
)
|
||||
|
|
|
|||
|
|
@ -15,8 +15,9 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev/IARPA-babel206b-v0.1e_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
)
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=18
|
||||
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev/IARPA-babel206b-v0.1e_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_206/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Zulu_Babel206//eval.list
|
||||
|
|
@ -46,15 +49,17 @@ shadow_data_dir=(
|
|||
/export/babel/data/206-zulu/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_206/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Zulu_Babel206/uem/206-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Zulu_Babel206/dev.list
|
||||
/export/babel/data/splits/Zulu_Babel206/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Zulu_Babel206/uem/206-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
|
|
|||
|
|
@ -15,8 +15,9 @@ dev2h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev
|
|||
dev2h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev/IARPA-babel206b-v0.1e_conv-dev.mitllfa3.rttm
|
||||
dev2h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
dev2h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
)
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev2h_subset_ecf=true
|
||||
dev2h_nj=18
|
||||
|
||||
|
|
@ -29,10 +30,12 @@ dev10h_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-de
|
|||
dev10h_rttm_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev/IARPA-babel206b-v0.1e_conv-dev.mitllfa3.rttm
|
||||
dev10h_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
dev10h_more_kwlists=(
|
||||
[limitedLP]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
)
|
||||
dev10h_nj=32
|
||||
|
||||
|
||||
#Official EVAL period evaluation data files
|
||||
eval_data_dir=/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_206/conversational/eval
|
||||
eval_data_list=/export/babel/data/splits/Zulu_Babel206//eval.list
|
||||
|
|
@ -46,18 +49,21 @@ shadow_data_dir=(
|
|||
/export/babel/data/206-zulu/release-current/conversational/dev
|
||||
/export/babel/data/IARPA-BABEL_OP1_dev_eval/BABEL_OP1_206/conversational/eval/
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Zulu_Babel206/uem/206-shadow-v0-cleaned-utt.dat
|
||||
shadow_data_list=(
|
||||
/export/babel/data/splits/Zulu_Babel206/dev.list
|
||||
/export/babel/data/splits/Zulu_Babel206/eval.list
|
||||
)
|
||||
shadow_data_cmudb=/export/babel/data/splits/Zulu_Babel206/uem/206-shadow-v0-utt.dat
|
||||
shadow_ecf_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.ecf.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
shadow_kwlist_file=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
shadow_more_kwlists=(
|
||||
[FullLPdev]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist.xml
|
||||
[llp]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-dev.kwlist2.xml
|
||||
[eval]=/export/babel/data/scoring/IndusDB/IARPA-babel206b-v0.1e_conv-eval.kwlist4.xml
|
||||
|
||||
)
|
||||
shadow_nj=32
|
||||
|
||||
|
||||
unsup_data_dir=(/export/babel/data/206-zulu/release-current/conversational/training/
|
||||
/export/babel/data/206-zulu/release-current/conversational/untranscribed-training/
|
||||
)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,33 @@
|
|||
#!/bin/bash
|
||||
set -e
|
||||
set -o pipefail
|
||||
|
||||
. ./cmd.sh; . ./path.sh;
|
||||
|
||||
|
||||
#(
|
||||
#bash filter_data.sh --cmd "$decode_cmd" data/shadow.uem eval.uem exp/sgmm5_mmi_b0.1/decode_*shadow.uem_it*
|
||||
#bash filter_data.sh --cmd "$decode_cmd" data/shadow.uem eval.uem exp_bnf/sgmm7_mmi_b0.1/decode_*shadow.uem_it*
|
||||
#) &
|
||||
#bash filter_data.sh --cmd "$decode_cmd" data/shadow.uem eval.uem exp/tri6*_nnet*/decode_shadow.uem*
|
||||
#wait
|
||||
|
||||
(
|
||||
bash filter_data.sh --cmd "$decode_cmd" data/shadow.uem dev10h.uem exp_bnf/sgmm7_mmi_b0.1/decode_*shadow.uem_it*
|
||||
#bash filter_data.sh --cmd "$decode_cmd" data/shadow.uem dev10h.uem exp/sgmm5_mmi_b0.1/decode_*shadow.uem_it*
|
||||
) &
|
||||
bash filter_data.sh --cmd "$decode_cmd" data/shadow.uem dev10h.uem exp/tri6*_nnet*/decode_shadow.uem
|
||||
wait
|
||||
|
||||
wait
|
||||
exit
|
||||
|
||||
bash make_release.sh --dryrun false --dir exp/sgmm5_mmi_b0.1 --data data/shadow.uem --master dev10h.uem lang.conf ./release
|
||||
bash make_release.sh --dryrun false --dir exp/tri6b_nnet --data data/shadow.uem --master dev10h.uem lang.conf ./release
|
||||
bash make_release.sh --dryrun false --dir exp_bnf/sgmm7_mmi_b0.1 --data data/shadow.uem --master dev10h.uem lang.conf ./release
|
||||
|
||||
bash make_release.sh --dryrun false --dir exp/sgmm5_mmi_b0.1 --extrasys "NEWJHU" --data data/dev10h.uem --master dev10h.uem lang.conf ./release
|
||||
bash make_release.sh --dryrun false --dir exp/tri6b_nnet --extrasys "NEWJHU" --data data/dev10h.uem --master dev10h.uem lang.conf ./release
|
||||
bash make_release.sh --dryrun false --dir exp_bnf/sgmm7_mmi_b0.1 --extrasys "NEWJHU" --data data/dev10h.uem --master dev10h.uem lang.conf ./release
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,125 @@
|
|||
min_lmwt=5
|
||||
max_lmwt=25
|
||||
cer=0
|
||||
cmd=run.pl
|
||||
. ./utils/parse_options.sh
|
||||
|
||||
min_lmwt_start=$min_lmwt
|
||||
max_lmwt_start=$max_lmwt
|
||||
|
||||
datadir=$1; shift
|
||||
name=$1; shift
|
||||
. ./lang.conf
|
||||
|
||||
set -e
|
||||
set -o pipefail
|
||||
|
||||
[ ! -d $datadir/compounds/$name ] && echo "Component called $name does not exist" && exit 1
|
||||
ecf=$datadir/compounds/$name/ecf.xml
|
||||
cat $ecf | grep -P -o '(?<=audio_filename\=")[^"]*' > $datadir/compounds/$name/files.list
|
||||
filelist=$datadir/compounds/$name/files.list
|
||||
[ -f $datadir/compounds/$name/rttm ] && rttm=$datadir/compounds/$name/rttm
|
||||
[ -f $datadir/compounds/$name/stm ] && stm=$datadir/compounds/$name/stm
|
||||
|
||||
if [ -f $ecf ] ; then
|
||||
duration=`head -1 $ecf |\
|
||||
grep -o -E "duration=\"[0-9]*[ \.]*[0-9]*\"" |\
|
||||
perl -e 'while($m=<>) {$m=~s/.*\"([0-9.]+)\".*/\1/; print $m/2.0;}'`
|
||||
echo "INFO: Using duration $duration seconds (from ECF)."
|
||||
else
|
||||
echo "WARNING: Using default duration. ECF wasn't specified?"
|
||||
duration=9999
|
||||
fi
|
||||
|
||||
inputname=`basename $datadir`
|
||||
outputname=$name
|
||||
|
||||
while (( "$#" )); do
|
||||
resultdir=$1;shift
|
||||
echo "Processing data directory $resultdir"
|
||||
|
||||
[ ! -d $resultdir ] && echo "Decode dir $resultdir does not exist!" && exit 1;
|
||||
|
||||
targetdir=$resultdir/$outputname
|
||||
|
||||
min_existing=
|
||||
max_existing=
|
||||
for lmw in `seq $min_lmwt_start $max_lmwt_start`; do
|
||||
[ -d $resultdir/score_$lmw ] && [ -z $min_existing ] && min_existing=$lmw
|
||||
[ -d $resultdir/score_$lmw ] && [ ! -z $min_existing ] && max_existing=$lmw
|
||||
done
|
||||
[ -z $min_existing ] && echo "Data directories to be scored could not be found!" && exit 1
|
||||
[ -z $max_existing ] && echo "Data directories to be scored could not be found!" && exit 1
|
||||
min_lmwt=$min_existing
|
||||
max_lmwt=$max_existing
|
||||
echo "Found data directories for range LMWT=$min_lmwt:$max_lmwt"
|
||||
|
||||
$cmd LMWT=$min_lmwt:$max_lmwt $targetdir/scoring/filter.LMWT.log \
|
||||
set -e';' set -o pipefail';' \
|
||||
mkdir -p $targetdir/score_LMWT/';'\
|
||||
test -f $resultdir/score_LMWT/$inputname.ctm '&&' \
|
||||
utils/filter_scp.pl $filelist $resultdir/score_LMWT/$inputname.ctm '>' \
|
||||
$targetdir/score_LMWT/$outputname.ctm || exit 1
|
||||
|
||||
if [ ! -z $stm ] && [ -f $stm ] ; then
|
||||
echo "For scoring CTMs, this STM is used $stm"
|
||||
local/score_stm.sh --min-lmwt $min_lmwt --max-lmwt $max_lmwt --cer $cer --cmd "$cmd" $datadir/compounds/$name data/lang $targetdir
|
||||
else
|
||||
echo "Not running scoring, $datadir/compounds/$name/stm does not exist"
|
||||
fi
|
||||
|
||||
|
||||
kws_tasks="kws"
|
||||
|
||||
for kws in `cat $datadir/extra_kws_tasks`; do
|
||||
kws_tasks+=" ${kws}_kws"
|
||||
done
|
||||
|
||||
for kws in $kws_tasks ; do
|
||||
echo "Processing KWS task: $kws"
|
||||
mkdir -p $targetdir/$kws
|
||||
filter=$targetdir/$kws/utterances
|
||||
grep -F -f $filelist $datadir/segments | tee $targetdir/$kws/segments | \
|
||||
awk '{print $1, $2}' | tee $targetdir/$kws/utter_map |\
|
||||
awk '{print $1}' > $filter
|
||||
|
||||
kwlist=$datadir/$kws/kwlist.xml
|
||||
|
||||
echo -e "\tFiltering..."
|
||||
#$cmd LMWT=$min_lmwt:$max_lmwt $targetdir/$kws/kws_filter.LMWT.log \
|
||||
# set -e';' set -o pipefail';' \
|
||||
# mkdir -p $targetdir/${kws}_LMWT';'\
|
||||
# cat $resultdir/${kws}_LMWT/'result.*' \| grep -F -f $filter \> $targetdir/${kws}_LMWT/result || exit 1
|
||||
|
||||
$cmd LMWT=$min_lmwt:$max_lmwt $targetdir/$kws/kws_filter.LMWT.log \
|
||||
set -e';' set -o pipefail';' \
|
||||
mkdir -p $targetdir/${kws}_LMWT';'\
|
||||
cat $resultdir/${kws}_LMWT/'result.*' \| utils/filter_scp.pl -f 1 $filter \> $targetdir/${kws}_LMWT/result || exit -1
|
||||
|
||||
|
||||
echo -e "\tWrite normalized..."
|
||||
$cmd LMWT=$min_lmwt:$max_lmwt $targetdir/$kws/kws_write_normalized.LMWT.log \
|
||||
set -e';' set -o pipefail';' \
|
||||
cat $targetdir/${kws}_LMWT/result \| \
|
||||
utils/write_kwslist.pl --flen=0.01 --duration=$duration \
|
||||
--segments=$targetdir/$kws/segments --normalize=true --remove-dup=true\
|
||||
--map-utter=$targetdir/$kws/utter_map --digits=3 - $targetdir/${kws}_LMWT/kwslist.xml || exit 1
|
||||
|
||||
echo -e "\tWrite unnormalized..."
|
||||
$cmd LMWT=$min_lmwt:$max_lmwt $targetdir/$kws/kws_write_unnormalized.LMWT.log \
|
||||
set -e';' set -o pipefail';' \
|
||||
cat $targetdir/${kws}_LMWT/result \| \
|
||||
utils/write_kwslist.pl --flen=0.01 --duration=$duration \
|
||||
--segments=$targetdir/$kws/segments --normalize=false --remove-dup=true\
|
||||
--map-utter=$targetdir/$kws/utter_map - $targetdir/${kws}_LMWT/kwslist.unnormalized.xml || exit 1
|
||||
|
||||
if [ ! -z $rttm ] ; then
|
||||
echo -e "\tScoring..."
|
||||
$cmd LMWT=$min_lmwt:$max_lmwt $targetdir/$kws/kws_score.LMWT.log \
|
||||
set -e';' set -o pipefail';' \
|
||||
local/kws_score.sh --ecf $ecf --rttm $rttm --kwlist $kwlist $datadir $targetdir/${kws}_LMWT || exit 1
|
||||
else
|
||||
echo -e "\tNot scoring..."
|
||||
fi
|
||||
done
|
||||
done
|
||||
|
|
@ -0,0 +1,229 @@
|
|||
if [ -z $1 ] ; then
|
||||
dir=`pwd`
|
||||
else
|
||||
dir=$1
|
||||
fi
|
||||
echo $dir
|
||||
|
||||
|
||||
convertsecs() {
|
||||
h=$(($1/3600))
|
||||
m=$((($1/60)%60))
|
||||
s=$(($1%60))
|
||||
printf "%02d:%02d:%02d\n" $h $m $s
|
||||
}
|
||||
|
||||
function process {
|
||||
count=1
|
||||
if [ ! -z $1 ]; then
|
||||
count=$1
|
||||
fi
|
||||
|
||||
replace=""
|
||||
for a in `seq 1 $count` ; do
|
||||
replace+="\t"
|
||||
done
|
||||
|
||||
(
|
||||
eval `grep "group=all"`
|
||||
echo -n "threads=$total_threads"
|
||||
echo -n " cpu_time=$total_cpu_time wall_time=$clock_time"
|
||||
echo -n " human_cpu_time="`convertsecs $total_cpu_time`
|
||||
echo -n " human_wall_time="`convertsecs $clock_time`
|
||||
echo ""
|
||||
) | sed 's/^/'$replace'/g'
|
||||
}
|
||||
|
||||
function legend {
|
||||
echo -ne '"'"$@"'" '
|
||||
}
|
||||
|
||||
legend Parameterization dev/train
|
||||
local/summarize_logs.pl $dir/exp/make_*/*train*/ | process
|
||||
|
||||
if [ -d $dir/data/local/extend ] ; then
|
||||
legend "Extending the lexicon"
|
||||
local/summarize_logs.pl $dir/data/local/extend/tmp/log | process
|
||||
fi
|
||||
|
||||
legend "Training upto stage tri5"
|
||||
local/summarize_logs.pl $dir/exp/mono*/log $dir/exp/tri{1..5}/log $dir/exp/tri{1..4}_ali*/log | process
|
||||
|
||||
legend "SGMM2 stage training"
|
||||
local/summarize_logs.pl $dir/exp/ubm5/log $dir/exp/sgmm5/log $dir/exp/tri5_ali/log | process
|
||||
|
||||
legend "SGMM2+bMMI stage training"
|
||||
local/summarize_logs.pl $dir/exp/sgmm5_*/log $dir/exp/ubm5/log $dir/exp/sgmm5_denlats/log/* | process
|
||||
|
||||
nnet=tri6_nnet
|
||||
[ ! -d $dir/exp/$nnet ] && nnet=tri6b_nnet
|
||||
|
||||
legend "DNN stage training GPU"
|
||||
local/summarize_logs.pl $dir/exp/$nnet/log | process
|
||||
|
||||
legend "BNF stage training"
|
||||
local/summarize_logs.pl $dir/exp_bnf/tri6_bnf/log | process
|
||||
|
||||
legend "BNF stage training GPU"
|
||||
local/summarize_logs.pl $dir/exp_bnf/tri{5,6}/log $dir/exp_bnf/sgmm7*/log \
|
||||
$dir/exp_bnf/sgmm7_denlats/log/* $dir/exp_bnf/ubm7 | process
|
||||
|
||||
legend "SEGMENTATION TRAINING: "
|
||||
local/summarize_logs.pl $dir/exp/tri4_train_seg_ali/log \
|
||||
$dir/exp/make_plp_pitch/train_seg/ \
|
||||
$dir/exp/tri4b_seg/log | process
|
||||
|
||||
semisup=exp_bnf_semisup2
|
||||
if [ -d $dir/param_bnf_semisup ] || [ -d $dir/param_bnf_semisup2 ] ; then
|
||||
[ ! -d $dir/$semisup ] && semisup=exp_bnf_semisup
|
||||
|
||||
decode=unsup.seg
|
||||
legend "BNF_SEMISUP training, segmentation "
|
||||
local/summarize_logs.pl $dir/exp/make_seg/$decode/log \
|
||||
$dir/exp/make_seg/$decode/make_plp/ \
|
||||
$dir/exp/tri4b_seg/decode_${decode}/log \
|
||||
$dir/exp/make_plp/$decode | process
|
||||
|
||||
legend "BNF_SEMISUP training, ecode unsup.seg TRI5 "
|
||||
local/summarize_logs.pl $dir/exp/tri5/decode_*${decode}*/log | process
|
||||
legend "BNF_SEMISUP training, ecode unsup.seg PLP "
|
||||
local/summarize_logs.pl $dir/exp/{sgmm5,sgmm5_mmi_b0.1}/decode_*${decode}*/log | process
|
||||
legend "BNF_SEMISUP training, ecode unsup.seg DNN "
|
||||
local/summarize_logs.pl $dir/exp/$nnet/decode_*${decode}*/log | process
|
||||
legend "BNF_SEMISUP training, data preparation for BNF_SEMISUP "
|
||||
local/summarize_logs.pl $dir/exp/combine2_post/unsup.seg/log \
|
||||
$dir/exp/combine2_post/unsup.seg/decode_unsup.seg/log\
|
||||
$dir/exp/tri6_nnet_ali/log | process
|
||||
|
||||
legend "BNF_SEMISUP training, TRAIN BNF_SEMISUP BNF GPU "
|
||||
local/summarize_logs.pl $dir/$semisup/tri6_bnf/log | process
|
||||
legend "BNF_SEMISUP training, TRAIN BNF_SEMISUP BNF "
|
||||
local/summarize_logs.pl $dir/$semisup/tri{5,6}/log $dir/exp_bnf/sgmm7*/log \
|
||||
$dir/exp_bnf/sgmm7_denlats/log/* $dir/exp_bnf/ubm7 | process
|
||||
fi
|
||||
|
||||
if [ -d $dir/exp/tri6_nnet_mpe ] ; then
|
||||
legend "DNN_MPE stage CPU training"
|
||||
local/summarize_logs.pl $dir/exp/tri6_nnet_ali/log/ \
|
||||
$dir/exp/tri6_nnet_denlats/log/* | process
|
||||
|
||||
legend "DNN_MPE stage GPU training"
|
||||
local/summarize_logs.pl $dir/exp/tri6_nnet_mpe/log/ | process
|
||||
fi
|
||||
|
||||
#~decode=dev10h.seg
|
||||
#~legend "DEV10H.SEG decoding"
|
||||
#~legend "Segmentation: "
|
||||
#~local/summarize_logs.pl $dir/exp/make_seg/$decode/log \
|
||||
#~ $dir/exp/make_seg/$decode/make_plp/ \
|
||||
#~ $dir/exp/tri4b_seg/decode_${decode}/log \
|
||||
#~ $dir/exp/make_plp/$decode | process
|
||||
#~legend "Decode $decode TRI5: "
|
||||
#~local/summarize_logs.pl $dir/exp/tri5/decode_*${decode}*/log | process
|
||||
#~legend "Decode $decode PLP: "
|
||||
#~local/summarize_logs.pl $dir/exp/{sgmm5,sgmm5_mmi_b0.1}/decode_*${decode}*/log | process
|
||||
#~legend "Decode $decode DNN: "
|
||||
#~local/summarize_logs.pl $dir/exp/$nnet/decode_*${decode}*/log | process
|
||||
#~legend "Decode $decode PLP: "
|
||||
#~local/summarize_logs.pl $dir/exp/{sgmm5,sgmm5_mmi_b0.1}/decode_*${decode}*/log | process
|
||||
|
||||
legend "G2P and confusion matrix: "
|
||||
local/summarize_logs.pl $dir/exp/conf_matrix/log $dir/exp/g2p/log | process
|
||||
if [ -d $dir/data/shadow2.uem ]; then
|
||||
decode=shadow2.uem
|
||||
else
|
||||
decode=shadow.uem
|
||||
fi
|
||||
|
||||
legend "Segmentation $decode: provided..."
|
||||
echo
|
||||
#--legend "Segmentation: "
|
||||
#--local/summarize_logs.pl $dir/exp/make_seg/$decode/log \
|
||||
#-- $dir/exp/make_seg/$decode/make_plp/ \
|
||||
#-- $dir/exp/tri4b_seg/decode_${decode}/log \
|
||||
#-- $dir/exp/make_plp/$decode | process
|
||||
legend "Parametrization: "
|
||||
local/summarize_logs.pl $dir/exp/make_plp/$decode | process
|
||||
legend "Decode $decode TRI5: "
|
||||
local/summarize_logs.pl $dir/exp/tri5/decode_*${decode}*/log | process
|
||||
legend "Decode $decode PLP: "
|
||||
local/summarize_logs.pl $dir/exp/{sgmm5,sgmm5_mmi_b0.1}/decode_*${decode}*/log | process
|
||||
legend "Decode $decode DNN: "
|
||||
local/summarize_logs.pl $dir/exp/$nnet/decode_*${decode}*/log | process
|
||||
legend "Decode $decode BNF: "
|
||||
local/summarize_logs.pl $dir/exp_bnf/{tri6,sgmm7,sgmm7_mmi_b0.1}/decode_*${decode}*/log | process
|
||||
if [ -d $dir/$semisup ] ; then
|
||||
legend "Decode $decode BNF_SEMISUP: "
|
||||
local/summarize_logs.pl $dir/$semisup/{tri6,sgmm7,sgmm7_mmi_b0.1}/decode_*${decode}*/log | process
|
||||
fi
|
||||
if [ -d $dir/exp/tri6_nnet_mpe ] ; then
|
||||
legend "Decode $decode DNN_MPE: "
|
||||
local/summarize_logs.pl $dir/exp/tri6_nnet_mpe/decode_${decode}_epoch*/log | process
|
||||
fi
|
||||
|
||||
legend "Indexing $decode PLP: "
|
||||
local/summarize_logs.pl $dir/exp/sgmm5_mmi_b0.1/decode_*${decode}*/kws_indices*/log | process
|
||||
legend "Indexing $decode DNN: "
|
||||
local/summarize_logs.pl $dir/exp/$nnet/decode_*${decode}*/kws_indices*/log | process
|
||||
legend "Indexing $decode BNF: "
|
||||
local/summarize_logs.pl $dir/exp_bnf/sgmm7_mmi_b0.1/decode_*${decode}*/kws_indices*/log | process
|
||||
if [ -d $dir/$semisup ] ; then
|
||||
legend "Indexing $decode BNF_SEMISUP: "
|
||||
local/summarize_logs.pl $dir/$semisup/sgmm7_mmi_b0.1/decode_*${decode}*/kws_indices*/log | process
|
||||
fi
|
||||
if [ -d $dir/exp/tri6_nnet_mpe ] ; then
|
||||
legend "Indexing $decode DNN_MPE: "
|
||||
local/summarize_logs.pl $dir/exp/tri6_nnet_mpe/decode_${decode}_epoch*/kws_indices*/log | process
|
||||
fi
|
||||
|
||||
legend "Search $decode PLP: "
|
||||
local/summarize_logs.pl $dir/exp/sgmm5_mmi_b0.1/decode_*${decode}*/evalKW_kws \
|
||||
$dir/exp/sgmm5_mmi_b0.1/decode_*${decode}*/evalKW_kws_*/log | process
|
||||
legend "Search $decode DNN: "
|
||||
local/summarize_logs.pl $dir/exp/$nnet/decode_*${decode}*/evalKW_kws \
|
||||
$dir/exp/$nnet/decode_*${decode}*/evalKW_kws_*/log | process
|
||||
legend "Search $decode BNF: "
|
||||
local/summarize_logs.pl $dir/exp_bnf/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_kws \
|
||||
$dir/exp_bnf/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_kws_*/log | process
|
||||
if [ -d $dir/$semisup ] ; then
|
||||
legend "Search $decode BNF_SEMISUP: "
|
||||
local/summarize_logs.pl $dir/$semisup/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_kws/ \
|
||||
$dir/$semisup/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_kws*/log | process
|
||||
fi
|
||||
if [ -d $dir/exp/tri6_nnet_mpe ] ; then
|
||||
legend "Search $decode DNN_MPE: "
|
||||
local/summarize_logs.pl $dir/exp/tri6_nnet_mpe/decode_${decode}_epoch*/evalKW_kws \
|
||||
$dir/exp/tri6_nnet_mpe/decode_${decode}_epoch*/evalKW_kws*/log | process
|
||||
fi
|
||||
|
||||
legend "Proxies generation: "
|
||||
local/summarize_logs.pl $dir/data/$decode/evalKW_oov_kws/g2p/log \
|
||||
$dir/data/$decode/evalKW_oov_kws/tmp/split/log | process
|
||||
legend "Search $decode PLP: "
|
||||
local/summarize_logs.pl $dir/exp/sgmm5_mmi_b0.1/decode_*${decode}*/evalKW_oov_kws \
|
||||
$dir/exp/sgmm5_mmi_b0.1/decode_*${decode}*/evalKW_oov_kws_*/log | process
|
||||
legend "Search $decode DNN: "
|
||||
local/summarize_logs.pl $dir/exp/$nnet/decode_*${decode}*/evalKW_oov_kws \
|
||||
$dir/exp/$nnet/decode_*${decode}*/evalKW_oov_kws_*/log | process
|
||||
legend "Search $decode BNF: "
|
||||
local/summarize_logs.pl $dir/exp_bnf/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_oov_kws \
|
||||
$dir/exp_bnf/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_oov_kws_*/log | process
|
||||
|
||||
if [ -d $dir/$semisup ] ; then
|
||||
legend "Search $decode BNF_SEMISUP: "
|
||||
local/summarize_logs.pl $dir/$semisup/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_oov_kws/ \
|
||||
$dir/$semisup/sgmm7_mmi_b0.1/decode_*${decode}*/evalKW_oov_kws*/log | process
|
||||
fi
|
||||
|
||||
|
||||
if [ -d $dir/exp/tri6_nnet_mpe ] ; then
|
||||
legend "Search $decode DNN_MPE: "
|
||||
local/summarize_logs.pl $dir/exp/tri6_nnet_mpe/decode_${decode}_epoch*/evalKW_oov_kws \
|
||||
$dir/exp/tri6_nnet_mpe/decode_${decode}_epoch*/evalKW_oov_kws*/log | process
|
||||
fi
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
|
@ -122,6 +122,6 @@ if [ $nlex -ne $nwlist ] ; then
|
|||
echo "WARNING: Lexicon : $nlex words"
|
||||
echo "WARNING:Diff example: "
|
||||
diff <(cut -f 1 $output_lex | sort -u ) \
|
||||
<(cut -f 1 $output/wordlist.orig.txt | sort -u )
|
||||
<(cut -f 1 $output/wordlist.orig.txt | sort -u ) || true
|
||||
fi
|
||||
exit 0
|
||||
|
|
|
|||
|
|
@ -89,7 +89,15 @@ if [ ! -z "$oov_prob_file" ]; then
|
|||
lmfile=$destdir/lm_tmp.gz
|
||||
fi
|
||||
|
||||
gunzip -c $lmfile | \
|
||||
if [[ $lmfile == *.bz2 ]] ; then
|
||||
decompress="bunzip2 -c $lmfile"
|
||||
elif [[ $lmfile == *.gz ]] ; then
|
||||
decompress="gunzip -c $lmfile"
|
||||
else
|
||||
decompress="cat $lmfile"
|
||||
fi
|
||||
|
||||
$decompress | \
|
||||
grep -v '<s> <s>' | grep -v '</s> <s>' | grep -v '</s> </s>' | \
|
||||
arpa2fst - | \
|
||||
fstprint | \
|
||||
|
|
@ -97,7 +105,7 @@ gunzip -c $lmfile | \
|
|||
utils/s2eps.pl | \
|
||||
fstcompile --isymbols=$langdir/words.txt \
|
||||
--osymbols=$langdir/words.txt --keep_isymbols=false --keep_osymbols=false | \
|
||||
fstrmepsilon > $destdir/G.fst || exit 1
|
||||
fstrmepsilon | fstarcsort --sort_type=olabel > $destdir/G.fst || exit 1
|
||||
fstisstochastic $destdir/G.fst || true;
|
||||
|
||||
if $cleanup; then
|
||||
|
|
|
|||
|
|
@ -73,13 +73,10 @@ mkdir -p $out_decode
|
|||
|
||||
if [ $stage -lt -1 ]; then
|
||||
mkdir -p $out_decode/log
|
||||
if [ ! -f $out_decode/.best_path.done ]; then
|
||||
$cmd JOB=1:$nj $out_decode/log/best_path.JOB.log \
|
||||
lattice-best-path --acoustic-scale=0.1 \
|
||||
"ark,s,cs:gunzip -c $decode_dir/lat.JOB.gz |" \
|
||||
ark:/dev/null "ark:| gzip -c > $out_decode/best_path_ali.JOB.gz" || exit 1
|
||||
touch $out_decode/.best_path.done
|
||||
fi
|
||||
$cmd JOB=1:$nj $out_decode/log/best_path.JOB.log \
|
||||
lattice-best-path --acoustic-scale=0.1 \
|
||||
"ark,s,cs:gunzip -c $decode_dir/lat.JOB.gz |" \
|
||||
ark:/dev/null "ark:| gzip -c > $out_decode/best_path_ali.JOB.gz" || exit 1
|
||||
fi
|
||||
|
||||
weights_sum=0.0
|
||||
|
|
|
|||
|
|
@ -14,36 +14,15 @@ fi
|
|||
check_variables_are_set
|
||||
|
||||
if [ ! -f ${dataset_dir}/kws/.done ] ; then
|
||||
if [ "$dataset_kind" == "shadow" ]; then
|
||||
# we expect that the ${dev2shadow} as well as ${eval2shadow} already exist
|
||||
if [ ! -f data/${dev2shadow}/kws/.done ]; then
|
||||
echo "Error: data/${dev2shadow}/kws/.done does not exist."
|
||||
echo "Create the directory data/${dev2shadow} first, by calling $0 --dir $dev2shadow --dataonly"
|
||||
exit 1
|
||||
fi
|
||||
if [ ! -f data/${eval2shadow}/kws/.done ]; then
|
||||
echo "Error: data/${eval2shadow}/kws/.done does not exist."
|
||||
echo "Create the directory data/${eval2shadow} first, by calling $0 --dir $eval2shadow --dataonly"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
local/kws_data_prep.sh --case_insensitive $case_insensitive \
|
||||
"${icu_opt[@]}" \
|
||||
data/lang ${dataset_dir} ${datadir}/kws || exit 1
|
||||
utils/fix_data_dir.sh ${dataset_dir}
|
||||
|
||||
touch ${dataset_dir}/kws/.done
|
||||
else # This will work for both supervised and unsupervised dataset kinds
|
||||
kws_flags=(--use-icu true)
|
||||
if [ "${dataset_kind}" == "supervised" ] ; then
|
||||
kws_flags+=(--rttm-file $my_rttm_file )
|
||||
fi
|
||||
if $my_subset_ecf ; then
|
||||
kws_flags+=(--subset-ecf $my_data_list)
|
||||
fi
|
||||
local/kws_setup.sh --case_insensitive $case_insensitive \
|
||||
"${kws_flags[@]}" "${icu_opt[@]}" \
|
||||
$my_ecf_file $my_kwlist_file data/lang ${dataset_dir} || exit 1
|
||||
kws_flags=( --use-icu true )
|
||||
if [ "${dataset_kind}" == "supervised" ] ; then
|
||||
kws_flags+=(--rttm-file $my_rttm_file )
|
||||
fi
|
||||
if $my_subset_ecf ; then
|
||||
kws_flags+=(--subset-ecf $my_data_list)
|
||||
fi
|
||||
local/kws_setup.sh --case_insensitive $case_insensitive \
|
||||
"${kws_flags[@]}" "${icu_opt[@]}" \
|
||||
$my_ecf_file $my_kwlist_file data/lang ${dataset_dir} || exit 1
|
||||
touch ${dataset_dir}/kws/.done
|
||||
fi
|
||||
|
|
|
|||
|
|
@ -22,13 +22,23 @@ function register_extraid {
|
|||
}
|
||||
|
||||
function setup_oov_search {
|
||||
local nbest=500
|
||||
#Basic lexicon
|
||||
#local phone_beam=-1
|
||||
#local phone_nbest=-1
|
||||
#local beam=5
|
||||
#local nbest=500
|
||||
|
||||
#Extended lexicon
|
||||
local nbest=-1
|
||||
local beam=-1
|
||||
local phone_nbest=300
|
||||
local phone_beam=5
|
||||
|
||||
local phone_cutoff=5
|
||||
|
||||
local g2p_nbest=10
|
||||
local g2p_mass=0.95
|
||||
local beam=5
|
||||
local phone_beam=4
|
||||
local phone_nbest=-1
|
||||
local phone_cutoff=5
|
||||
|
||||
|
||||
local data_dir=$1
|
||||
local source_dir=$2
|
||||
|
|
@ -37,10 +47,15 @@ function setup_oov_search {
|
|||
local kwsdatadir=$data_dir/${extraid}_kws
|
||||
|
||||
mkdir -p $kwsdatadir
|
||||
cp $source_dir/kwlist*.xml $kwsdatadir
|
||||
cp $source_dir/ecf.xml $kwsdatadir
|
||||
cp $source_dir/utter_* $kwsdatadir
|
||||
[ -f $source_dir/rttm ] && cp $source_dir/rttm $kwsdatadir
|
||||
|
||||
if [ "${dataset_kind}" == "supervised" ] ; then
|
||||
for file in $source_dir/rttm ; do
|
||||
cp -f $file $kwsdatadir
|
||||
done
|
||||
fi
|
||||
for file in $source_dir/utter_* $source_dir/kwlist*.xml $source_dir/ecf.xml ; do
|
||||
cp -f $file $kwsdatadir
|
||||
done
|
||||
|
||||
kwlist=$source_dir/kwlist_outvocab.xml
|
||||
#Get the KW list
|
||||
|
|
@ -84,55 +99,50 @@ function setup_oov_search {
|
|||
}
|
||||
|
||||
|
||||
if [ "$dataset_kind" == "shadow" ]; then
|
||||
true #we do not support multiple kw lists for shadow set system
|
||||
|
||||
else # This will work for both supervised and unsupervised dataset kinds
|
||||
kws_flags=( --use-icu true )
|
||||
if [ "${dataset_kind}" == "supervised" ] ; then
|
||||
#The presence of the file had been already verified, so just
|
||||
#add the correct switches
|
||||
kws_flags+=(--rttm-file $my_rttm_file )
|
||||
fi
|
||||
if $my_subset_ecf ; then
|
||||
kws_flags+=(--subset-ecf $my_data_list)
|
||||
fi
|
||||
|
||||
if [ ! -f $dataset_dir/.done.kws.oov ] ; then
|
||||
setup_oov_search $dataset_dir $dataset_dir/kws oov
|
||||
register_extraid $dataset_dir oov
|
||||
touch $dataset_dir/.done.kws.oov
|
||||
fi
|
||||
if [ ${#my_more_kwlists[@]} -ne 0 ] ; then
|
||||
|
||||
touch $dataset_dir/extra_kws_tasks
|
||||
|
||||
for extraid in "${!my_more_kwlists[@]}" ; do
|
||||
#The next line will help us in running only one. We don't really
|
||||
#know in which directory the KWS setup will reside in, so we will
|
||||
#place the .done file directly into the data directory
|
||||
[ -f $dataset_dir/.done.kws.$extraid ] && continue;
|
||||
kwlist=${my_more_kwlists[$extraid]}
|
||||
|
||||
local/kws_setup.sh --extraid $extraid --case_insensitive $case_insensitive \
|
||||
"${kws_flags[@]}" "${icu_opt[@]}" \
|
||||
$my_ecf_file $kwlist data/lang ${dataset_dir} || exit 1
|
||||
|
||||
#Register the dataset for default running...
|
||||
#We can do it without any problem here -- the kws_stt_tasks will not
|
||||
#run it, unless called with --run-extra-tasks true switch
|
||||
register_extraid $dataset_dir $extraid
|
||||
touch $dataset_dir/.done.kws.$extraid
|
||||
done
|
||||
for extraid in "${!my_more_kwlists[@]}" ; do
|
||||
#The next line will help us in running only one. We don't really
|
||||
#know in which directory the KWS setup will reside in, so we will
|
||||
#place the .done file directly into the data directory
|
||||
[ -f $dataset_dir/.done.kws.${extraid}_oov ] && continue;
|
||||
setup_oov_search $dataset_dir $dataset_dir/${extraid}_kws ${extraid}_oov
|
||||
register_extraid $dataset_dir ${extraid}_oov
|
||||
touch $dataset_dir/.done.kws.${extraid}_oov
|
||||
done
|
||||
fi
|
||||
kws_flags=( --use-icu true )
|
||||
if [ "${dataset_kind}" == "supervised" ] ; then
|
||||
#The presence of the file had been already verified, so just
|
||||
#add the correct switches
|
||||
kws_flags+=(--rttm-file $my_rttm_file )
|
||||
fi
|
||||
if $my_subset_ecf ; then
|
||||
kws_flags+=(--subset-ecf $my_data_list)
|
||||
fi
|
||||
|
||||
if [ ! -f $dataset_dir/.done.kws.oov ] ; then
|
||||
setup_oov_search $dataset_dir $dataset_dir/kws oov || exit 1
|
||||
register_extraid $dataset_dir oov
|
||||
touch $dataset_dir/.done.kws.oov
|
||||
fi
|
||||
if [ ${#my_more_kwlists[@]} -ne 0 ] ; then
|
||||
|
||||
touch $dataset_dir/extra_kws_tasks
|
||||
|
||||
for extraid in "${!my_more_kwlists[@]}" ; do
|
||||
#The next line will help us in running only one. We don't really
|
||||
#know in which directory the KWS setup will reside in, so we will
|
||||
#place the .done file directly into the data directory
|
||||
[ -f $dataset_dir/.done.kws.$extraid ] && continue;
|
||||
kwlist=${my_more_kwlists[$extraid]}
|
||||
|
||||
local/kws_setup.sh --extraid $extraid --case_insensitive $case_insensitive \
|
||||
"${kws_flags[@]}" "${icu_opt[@]}" \
|
||||
$my_ecf_file $kwlist data/lang ${dataset_dir} || exit 1
|
||||
|
||||
#Register the dataset for default running...
|
||||
#We can do it without any problem here -- the kws_stt_tasks will not
|
||||
#run it, unless called with --run-extra-tasks true switch
|
||||
register_extraid $dataset_dir $extraid
|
||||
touch $dataset_dir/.done.kws.$extraid
|
||||
done
|
||||
for extraid in "${!my_more_kwlists[@]}" ; do
|
||||
#The next line will help us in running only one. We don't really
|
||||
#know in which directory the KWS setup will reside in, so we will
|
||||
#place the .done file directly into the data directory
|
||||
[ -f $dataset_dir/.done.kws.${extraid}_oov ] && continue;
|
||||
setup_oov_search $dataset_dir $dataset_dir/${extraid}_kws ${extraid}_oov
|
||||
register_extraid $dataset_dir ${extraid}_oov
|
||||
touch $dataset_dir/.done.kws.${extraid}_oov
|
||||
done
|
||||
fi
|
||||
|
||||
|
|
|
|||
|
|
@ -1,4 +1,4 @@
|
|||
#!/usr/bin/perl
|
||||
#!/usr/bin/env perl
|
||||
|
||||
# Copyright 2012 Johns Hopkins University (Author: Guoguo Chen, Jan Trmal)
|
||||
# Apache 2.0.
|
||||
|
|
|
|||
|
|
@ -0,0 +1,101 @@
|
|||
#! /usr/bin/env python
|
||||
|
||||
import argparse, sys
|
||||
from argparse import ArgumentParser
|
||||
import re
|
||||
|
||||
def main():
|
||||
parser = ArgumentParser(description='Convert kaldi data directory to uem dat files',
|
||||
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
|
||||
parser.add_argument('--verbose', type=int, \
|
||||
dest='verbose', default=0, \
|
||||
help='Give higher verbose for more logging')
|
||||
parser.add_argument('--get-text', action='store_true', \
|
||||
help='Get text in dat file')
|
||||
parser.add_argument('--prefix', type=str, \
|
||||
help='Add db file name as db-<prefix>-{utt/spk}.dat')
|
||||
parser.add_argument('kaldi_dir', \
|
||||
help='Kaldi data directory')
|
||||
parser.add_argument('output_dir', \
|
||||
help='Directory to store uem dat files')
|
||||
parser.usage=':'.join(parser.format_usage().split(':')[1:]) \
|
||||
+ 'e.g. : %(prog)s --prefix 203-lao-v0 data/dev10h.seg CMU_db'
|
||||
options = parser.parse_args()
|
||||
|
||||
if options.get_text:
|
||||
try:
|
||||
text_file = open(options.kaldi_dir+'/text', 'r')
|
||||
except IOError as e:
|
||||
repr(e)
|
||||
sys.stderr.write("%s: No such file %s\n" % (sys.argv[0], options.kaldi_dir+'/text'))
|
||||
sys.exit(1)
|
||||
|
||||
try:
|
||||
segments_file = open(options.kaldi_dir+'/segments', 'r')
|
||||
except IOError as e:
|
||||
repr(e)
|
||||
sys.stderr.write("%s: No such file %s\n" % (sys.argv[0], options.kaldi_dir+'/segments'))
|
||||
sys.exit(1)
|
||||
|
||||
try:
|
||||
scp_file = open(options.kaldi_dir+'/wav.scp', 'r')
|
||||
except IOError as e:
|
||||
repr(e)
|
||||
sys.stderr.write("%s: No such file %s\n" % (sys.argv[0], options.kaldi_dir+'/wav.scp'))
|
||||
sys.exit(1)
|
||||
|
||||
reco2file_map = {}
|
||||
for line in scp_file.readlines():
|
||||
splits = line.strip().split()
|
||||
m = re.search(r".*/(?P<file_name>[0-9A-Za-z_]*\.(sph|wav)).*", line)
|
||||
if not m:
|
||||
sys.stderr.write("%s does not contain a valid speech file (.wav or .sph)\n" % line.strip())
|
||||
sys.exit(1)
|
||||
reco2file_map[splits[0]] = m.group('file_name')
|
||||
# End for
|
||||
|
||||
spk2utt_map = {}
|
||||
|
||||
if options.prefix == None:
|
||||
prefix = options.kaldi_dir.split('/')[-1].split('.')[0]
|
||||
else:
|
||||
prefix = options.prefix
|
||||
|
||||
try:
|
||||
utt_dat = open(options.output_dir+'/db-'+prefix+'-utt.dat', 'w')
|
||||
spk_dat = open(options.output_dir+'/db-'+prefix+'-spk.dat', 'w')
|
||||
except IOError as e:
|
||||
repr(e)
|
||||
sys.stderr.write("%s: Could not write dat files in %s\n" % (sys.argv[0], options.output_dir))
|
||||
sys.exit(1)
|
||||
|
||||
for line in segments_file.readlines():
|
||||
utt_id, file_id, start, end = line.strip().split()
|
||||
|
||||
if (options.get_text):
|
||||
splits = text_file.readline().split()
|
||||
while splits[0] < utt_id:
|
||||
splits = text_file.readline().split()
|
||||
text = ' '.join(splits[1:])
|
||||
else:
|
||||
text = ""
|
||||
|
||||
utt_dat.write("{UTTID %s} {UTT %s} {SPK %s} {FROM %s} {TO %s} {TEXT %s}\n" % (utt_id, utt_id, file_id, start, end, text))
|
||||
spk2utt_map.setdefault(file_id, [])
|
||||
spk2utt_map[file_id].append(utt_id)
|
||||
|
||||
for spk, utts in spk2utt_map.items():
|
||||
try:
|
||||
spk_dat.write("{SEGS %s} {ADC %s} {CONV %s.wav} {CHANNEL 1} {DUR }\n" % (' '.join(utts), reco2file_map[spk], spk))
|
||||
except KeyError as e:
|
||||
repr(e)
|
||||
sys.stderr.write("%s: Error in getting file for %s\n" % (sys.argv[0], spk))
|
||||
sys.exit(1)
|
||||
# End for
|
||||
|
||||
segments_file.close()
|
||||
utt_dat.close()
|
||||
spk_dat.close()
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
|
|
@ -85,6 +85,7 @@ fi
|
|||
mkdir -p $kwsdatadir
|
||||
|
||||
if [ -z $subset_ecf ] ; then
|
||||
test -f $kwsdatadir/ecf.xml && rm -f $kwsdatadir/ecf.xml
|
||||
cp "$ecf_file" $kwsdatadir/ecf.xml || exit 1
|
||||
else
|
||||
local/make_ecf_subset.sh $subset_ecf $ecf_file > $kwsdatadir/ecf.xml
|
||||
|
|
@ -107,10 +108,12 @@ if $kwlist_wordlist ; then
|
|||
echo '</kwlist>'
|
||||
) > $kwsdatadir/kwlist.xml || exit 1
|
||||
else
|
||||
test -f $kwsdatadir/kwlist.xml && rm -f $kwsdatadir/kwlist.xml
|
||||
cp "$kwlist_file" $kwsdatadir/kwlist.xml || exit 1
|
||||
fi
|
||||
|
||||
if [ ! -z $rttm_file ] ; then
|
||||
test -f $kwsdatadir/rttm && rm -f $kwsdatadir/rttm
|
||||
cp "$rttm_file" $kwsdatadir/rttm || exit 1
|
||||
fi
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,333 @@
|
|||
#!/usr/bin/env python
|
||||
from pylauncher import *
|
||||
import pylauncher
|
||||
import sys
|
||||
|
||||
import os
|
||||
import errno
|
||||
|
||||
def make_path(path):
|
||||
try:
|
||||
os.makedirs(path)
|
||||
except OSError as exception:
|
||||
if exception.errno != errno.EEXIST:
|
||||
raise
|
||||
elif not os.path.isdir(path):
|
||||
raise
|
||||
|
||||
def tail(n, filename):
|
||||
import subprocess
|
||||
p=subprocess.Popen(['tail','-n',str(n),filename], stdout=subprocess.PIPE)
|
||||
soutput,sinput=p.communicate()
|
||||
soutput=soutput.split("\n")
|
||||
return soutput
|
||||
|
||||
def KaldiLauncher(lo, **kwargs):
|
||||
import time;
|
||||
jobid = JobId()
|
||||
debug = kwargs.pop("debug","")
|
||||
qdir= os.path.join(lo.qdir, lo.taskname);
|
||||
cores = lo.nof_threads;
|
||||
|
||||
ce=SSHExecutor(workdir=qdir, debug=debug, force_workdir=True, catch_output=True)
|
||||
ce.outstring="out."
|
||||
ce.execstring=lo.taskname + "."
|
||||
|
||||
hostpool=HostPool(hostlist=HostListByName(), commandexecutor=ce )
|
||||
|
||||
completion=lambda x:FileCompletion( taskid=x, stamproot="done.", stampdir=qdir)
|
||||
|
||||
logfiles = list()
|
||||
commands = list()
|
||||
for q in xrange(lo.jobstart, lo.jobend+1):
|
||||
s = "bash " + lo.queue_scriptfile + " " + str(q)
|
||||
commands.append(s)
|
||||
|
||||
logfile = lo.logfile.replace("${PY_LAUNCHER_ID}", str(q))
|
||||
logfiles.append(logfile)
|
||||
|
||||
generator=ListCommandlineGenerator(list=commands, cores=cores)
|
||||
tasks = TaskGenerator(generator, completion=completion, debug=debug )
|
||||
|
||||
job = LauncherJob( hostpool=hostpool, taskgenerator=tasks, debug=debug,**kwargs)
|
||||
|
||||
job.run()
|
||||
#At this point all the .done files should exist and everything should be finalized.
|
||||
num_failed=0;
|
||||
time.sleep(1); #Lets wait for a while to give the shared fs time to sync
|
||||
error_pending=True
|
||||
for logfile in logfiles:
|
||||
import time
|
||||
sched_rate=[0, 0.5, 1, 2, 4, 8, 15, 32 ];
|
||||
for delay in sched_rate:
|
||||
time.sleep(delay);
|
||||
if os.path.isfile(logfile):
|
||||
break;
|
||||
if not os.path.isfile(logfile):
|
||||
sys.stderr.write("ERROR: " + "The following file is missing:\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + logfile + "\n")
|
||||
sys.stderr.write("ERROR: " + "That means something went wrong, but we don't know what. Try to figure out what and fix it\n");
|
||||
sys.exit(-1);
|
||||
|
||||
error_pending=True;
|
||||
for delay in sched_rate:
|
||||
time.sleep(delay);
|
||||
|
||||
lines=tail(10, logfile)
|
||||
with_status=filter(lambda x:re.search(r'with status (\d+)', x), lines)
|
||||
|
||||
if len(with_status) == 0:
|
||||
sys.stderr.write("The last line(s) of the log-file " + logfile + " does not seem"
|
||||
" to indicate return status as expected\n");
|
||||
elif len(with_status) > 1:
|
||||
sys.stderr.write("The last line(s) of the log-file " + logfile + " does seem"
|
||||
" to indicate multiple return statuses \n");
|
||||
else:
|
||||
status_re=re.search(r'with status (\d+)', with_status[0]);
|
||||
status=status_re.group(1);
|
||||
if status == '0':
|
||||
error_pending=False;
|
||||
break;
|
||||
sys.stderr.write("INFO: Waiting for status in files, sleeping %d seconds\n" % (delay,))
|
||||
if error_pending:
|
||||
num_failed+=1;
|
||||
|
||||
if num_failed != 0:
|
||||
sys.stderr.write(sys.argv[0] + ": " + str(num_failed) + "/" + str(len(logfiles)) + " failed \n");
|
||||
sys.stderr.write(sys.argv[0] + ": See " + lo.logfile.replace("${PY_LAUNCHER_ID}", "*" ) + " for details\n");
|
||||
sys.exit(-1);
|
||||
|
||||
#Remove service files. Be careful not to remove something that might be needed in problem diagnostics
|
||||
for i in xrange(len(commands)):
|
||||
out_file=os.path.join(qdir, ce.outstring+str(i))
|
||||
|
||||
#First, let's wait on files missing (it might be that those are missing
|
||||
#just because of slow shared filesystem synchronization
|
||||
if not os.path.isfile(out_file):
|
||||
import time
|
||||
sched_rate=[0.5, 1, 2, 4, 8 ];
|
||||
for delay in sched_rate:
|
||||
time.sleep(delay);
|
||||
if os.path.isfile(out_file):
|
||||
break;
|
||||
if not os.path.isfile(out_file):
|
||||
sys.stderr.write("ERROR: " + "The following file is missing:\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + out_file + "\n")
|
||||
sys.stderr.write("ERROR: " + "That means something went wrong, but we don't know what. Try to figure out what and fix it\n");
|
||||
sys.exit(-1);
|
||||
|
||||
if os.stat(out_file).st_size != 0:
|
||||
sys.stderr.write("ERROR: " + "The following file has non-zero size:\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + out_file + "\n")
|
||||
sys.stderr.write("ERROR: " + "That means something went wrong, but we don't know what. Try to figure out what and fix it\n");
|
||||
sys.exit(-1);
|
||||
else:
|
||||
exec_file=os.path.join(qdir, ce.execstring+str(i))
|
||||
done_file=os.path.join(qdir, "done."+str(i))
|
||||
if (not os.path.isfile(exec_file) ) or (not os.path.isfile(done_file)):
|
||||
sys.stderr.write("ERROR: " + "One of the following files is missing:\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + exec_file + "\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + done_file + "\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + out_file + "\n")
|
||||
sys.stderr.write("ERROR: " + "That means something went wrong, but we don't know what. Try to figure out what and fix it\n");
|
||||
sys.exit(-1);
|
||||
elif os.stat(done_file).st_size != 0:
|
||||
sys.stderr.write("ERROR: " + "The following file has non-zero size:\n")
|
||||
sys.stderr.write("ERROR: " + "\t" + done_file + "\n")
|
||||
sys.stderr.write("ERROR: " + "That means something went wrong, but we don't know what. Try to figure out what and fix it\n");
|
||||
sys.exit(-1);
|
||||
else:
|
||||
os.remove(exec_file)
|
||||
os.remove(done_file)
|
||||
os.remove(out_file)
|
||||
try:
|
||||
os.rmdir(qdir)
|
||||
except OSError:
|
||||
sys.stderr.write("ERROR: " + "Failed to remove the pylauncher task dir " + qdir + "\n");
|
||||
sys.stderr.write("ERROR: " + "Find out what is wrong and fix it\n")
|
||||
sys.exit(-1);
|
||||
|
||||
#print job.final_report()
|
||||
|
||||
class LauncherOpts:
|
||||
def __init__(self):
|
||||
self.sync=0
|
||||
self.nof_threads = 1
|
||||
self.qsub_opts = None
|
||||
|
||||
self.jobname=None
|
||||
self.jobstart=None
|
||||
self.jobend=None
|
||||
pass
|
||||
|
||||
def CmdLineParser(argv):
|
||||
import re;
|
||||
sync=0
|
||||
qsub_opts=''
|
||||
nof_threads=1
|
||||
|
||||
while len(argv) >= 2 and argv[0].startswith('-'):
|
||||
switch = argv.pop(0);
|
||||
|
||||
if switch == '-V':
|
||||
qsub_opts += switch + ' ';
|
||||
else:
|
||||
option = argv.pop(0)
|
||||
|
||||
if switch == "-sync" and (option in ['Y', 'y']):
|
||||
sync=1;
|
||||
qsub_opts += switch + ' ' + option + ' ';
|
||||
if switch == "-pe":
|
||||
option2 = argv.pop(0);
|
||||
qsub_opts += option2 + ' ';
|
||||
nof_threads = int(option2);
|
||||
|
||||
#Now we have to parse the JOB specifier
|
||||
jobname = ""
|
||||
jobstart = 0
|
||||
jobend = 0
|
||||
if (re.match( r"^[A-Za-z_]\w*=\d+:\d+$", argv[0])):
|
||||
m=re.match( r"^([A-Za-z_]\w*)=(\d+):(\d+)$", argv[0])
|
||||
jobname=m.group(1)
|
||||
jobstart=int(m.group(2))
|
||||
jobend=int(m.group(3))
|
||||
argv.pop(0)
|
||||
elif(re.match( r"^[A-Za-z_]\w*=\d+$", argv[0])):
|
||||
m=re.match( r"^([A-Za-z_]\w*)=(\d+)$", argv[0])
|
||||
jobname=m.group(1)
|
||||
jobstart=int(m.group(2))
|
||||
jobend=int(m.group(2))
|
||||
argv.pop(0)
|
||||
elif re.match("^.+=.*:.*$", argv[0]):
|
||||
print >> sys.stderr, "warning: suspicious JOB argument " + argv[0];
|
||||
|
||||
if jobstart > jobend:
|
||||
sys.stderr.write("lonestar.py: JOBSTART("+ str(jobstart) + ") must be lower than JOBEND(" + str(jobend) + ")\n")
|
||||
sys.exit(1)
|
||||
|
||||
logfile=argv.pop(0)
|
||||
|
||||
opts=LauncherOpts()
|
||||
opts.sync = sync
|
||||
opts.nof_threads=nof_threads;
|
||||
opts.qsub_opts=qsub_opts
|
||||
opts.varname=jobname
|
||||
opts.jobstart=jobstart
|
||||
opts.jobend=jobend
|
||||
opts.logfile=logfile
|
||||
|
||||
opts.cmd = escape_cmd(argv);
|
||||
|
||||
return (opts, argv)
|
||||
|
||||
def escape_cmd(argv):
|
||||
cmd =""
|
||||
for x in argv:
|
||||
#print x + " -> ",
|
||||
if re.search("^\S+$", x):
|
||||
#print " A -> ",
|
||||
cmd += x + " "
|
||||
elif '"' in x:
|
||||
cmd += "'''" + x + "''' "
|
||||
else:
|
||||
cmd += "\"" + x + "\" "
|
||||
#print cmd
|
||||
return cmd
|
||||
|
||||
def setup_paths_and_vars(opts):
|
||||
cwd = os.getcwd()
|
||||
|
||||
if opts.varname and (opts.varname not in opts.logfile ) and (opts.jobstart != opts.jobend):
|
||||
print >>sys.stderr, "lonestar.py: you are trying to run a parallel job" \
|
||||
"but you are putting the output into just one log file (" + opts.logfile + ")";
|
||||
sys.exit(1)
|
||||
|
||||
if not os.path.isabs(opts.logfile):
|
||||
opts.logfile = os.path.join(cwd, opts.logfile);
|
||||
logfile=opts.logfile
|
||||
|
||||
dir = os.path.dirname(logfile)
|
||||
base = os.path.basename(logfile)
|
||||
qdir = os.path.join(dir, "q");
|
||||
|
||||
if re.search("log/*q", qdir, flags=re.IGNORECASE):
|
||||
qdir = re.sub("log/*q", "/q", qdir, flags=re.IGNORECASE)
|
||||
|
||||
|
||||
queue_logfile= os.path.join(qdir, base)
|
||||
if opts.varname:
|
||||
queue_logfile = re.sub("\.?"+opts.varname, "", queue_logfile)
|
||||
|
||||
taskname=os.path.basename(queue_logfile)
|
||||
taskname = taskname.replace(".log", "");
|
||||
if taskname == "":
|
||||
print >> sys.stderr, "lonestar.py: you specified the log file name in such form " \
|
||||
"that leads to an empty task name ("+logfile + ")";
|
||||
sys.exit(1)
|
||||
|
||||
if not os.path.isabs(queue_logfile):
|
||||
queue_logfile= os.path.join(cwd, queue_logfile)
|
||||
|
||||
if opts.varname:
|
||||
opts.logfile = opts.logfile.replace(opts.varname, "${PY_LAUNCHER_ID}")
|
||||
opts.cmd = opts.cmd.replace(opts.varname, "${PY_LAUNCHER_ID}");
|
||||
|
||||
queue_scriptfile=queue_logfile;
|
||||
if re.search("\.[a-zA-Z]{1,5}$", queue_scriptfile):
|
||||
queue_scriptfile = re.sub("\.[a-zA-Z]{1,5}$", ".sh", queue_scriptfile);
|
||||
if not os.path.isabs(queue_scriptfile):
|
||||
queue_scriptfile= os.path.join(cwd, queue_scriptfile)
|
||||
|
||||
|
||||
make_path(qdir)
|
||||
make_path(dir)
|
||||
|
||||
opts.qdir = qdir
|
||||
opts.log_dir = dir
|
||||
opts.queue_scriptfile = queue_scriptfile
|
||||
opts.queue_logfile = queue_logfile
|
||||
opts.taskname = taskname
|
||||
|
||||
return opts
|
||||
|
||||
|
||||
|
||||
def create_scriptfile(scriptname, opts):
|
||||
import os
|
||||
logfile = opts.logfile
|
||||
cmd = opts.cmd
|
||||
nof_threads=opts.nof_threads;
|
||||
cwd = os.getcwd()
|
||||
#print scriptname
|
||||
f = open(scriptname, "wb")
|
||||
f.write("#!/bin/bash\n")
|
||||
f.write("export PY_LAUNCHER_ID=$1; shift;\n")
|
||||
f.write("cd " + cwd + "\n")
|
||||
f.write(". ./path.sh\n")
|
||||
f.write("( echo '#' Running on `hostname`\n")
|
||||
f.write(" echo '#' Started at `date`\n")
|
||||
f.write(" echo -n '# '; cat <<EOF\n")
|
||||
f.write(cmd + "\n")
|
||||
f.write("EOF\n")
|
||||
f.write(") > " +logfile + "\n")
|
||||
f.write("time1=`date +\"%s\"`\n")
|
||||
f.write("( " + cmd + ") 2>>" + logfile + " >>" + logfile + " \n")
|
||||
f.write("ret=$?\n")
|
||||
f.write("time2=`date +\"%s\"`\n")
|
||||
f.write("echo '#' Accounting time=$(($time2 - $time1)) threads=" + str(nof_threads) + " >> " + logfile + "\n")
|
||||
|
||||
f.write("echo '#' Finished at `date` with status $ret >>" + logfile + "\n")
|
||||
f.write("exit $ret \n")
|
||||
f.close()
|
||||
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
(opts, cmd) = CmdLineParser(sys.argv[1:]);
|
||||
setup_paths_and_vars(opts)
|
||||
create_scriptfile(opts.queue_scriptfile, opts);
|
||||
|
||||
#pylauncher.ClassicLauncher(["true && sleep 10s", "false || sleep 1s" ], debug="job+host+task+exec+ssh")
|
||||
KaldiLauncher(opts, debug="")
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,81 @@
|
|||
#! /usr/bin/bash
|
||||
|
||||
VARIABLES=`diff <(compgen -A variable) <(. ./lang.conf.orig; compgen -A variable) | grep '^>'| sed 's/^> *//g'`
|
||||
|
||||
. ./conf/common_vars.sh
|
||||
. ./lang.conf.orig
|
||||
|
||||
for variable in $VARIABLES ; do
|
||||
|
||||
eval VAL=\$${variable}
|
||||
if [[ $VAL =~ /export/babel/data/ ]] ; then
|
||||
eval $variable=${VAL/${BASH_REMATCH[0]}/"/work/02359/jtrmal/"/}
|
||||
#declare -x $variable
|
||||
declare -p $variable
|
||||
fi
|
||||
done
|
||||
|
||||
for kwlist in $( (compgen -A variable) | grep _data_list ) ; do
|
||||
declare -p $kwlist
|
||||
eval KEYS="\${!${kwlist}[@]}"
|
||||
#declare -p my_more_kwlist_keys
|
||||
for key in $KEYS # make sure you include the quotes there
|
||||
do
|
||||
#echo $key
|
||||
eval VAL="\${${kwlist}[$key]}"
|
||||
#echo $my_more_kwlist_val
|
||||
if [[ $VAL =~ /export/babel/data/ ]] ; then
|
||||
eval $kwlist["$key"]=${VAL/${BASH_REMATCH[0]}/"/work/02359/jtrmal/"/}
|
||||
fi
|
||||
done
|
||||
declare -p $kwlist
|
||||
done
|
||||
unset VAL
|
||||
unset KEYS
|
||||
|
||||
for kwlist in $( (compgen -A variable) | grep _data_dir ) ; do
|
||||
declare -p $kwlist
|
||||
eval KEYS="\${!${kwlist}[@]}"
|
||||
#declare -p my_more_kwlist_keys
|
||||
for key in $KEYS # make sure you include the quotes there
|
||||
do
|
||||
#echo $key
|
||||
eval VAL="\${${kwlist}[$key]}"
|
||||
#echo $my_more_kwlist_val
|
||||
if [[ $VAL =~ /export/babel/data/ ]] ; then
|
||||
eval $kwlist["$key"]=${VAL/${BASH_REMATCH[0]}/"/work/02359/jtrmal/"/}
|
||||
fi
|
||||
done
|
||||
declare -p $kwlist
|
||||
done
|
||||
unset VAL
|
||||
unset KEYS
|
||||
|
||||
for kwlist in $( (compgen -A variable) | grep _more_kwlists ) ; do
|
||||
declare -p $kwlist
|
||||
eval KEYS="\${!${kwlist}[@]}"
|
||||
#declare -p my_more_kwlist_keys
|
||||
for key in $KEYS # make sure you include the quotes there
|
||||
do
|
||||
#echo $key
|
||||
eval VAL="\${${kwlist}[$key]}"
|
||||
#echo $my_more_kwlist_val
|
||||
if [[ $VAL =~ /export/babel/data/ ]] ; then
|
||||
eval $kwlist["$key"]=${VAL/${BASH_REMATCH[0]}/"/work/02359/jtrmal/"/}
|
||||
fi
|
||||
done
|
||||
declare -p $kwlist
|
||||
done
|
||||
unset VAL
|
||||
unset KEYS
|
||||
|
||||
if [ "$babel_type" == "limited" ] ; then
|
||||
train_nj=32
|
||||
else
|
||||
train_nj=64
|
||||
fi
|
||||
dev10h_nj=60
|
||||
unsup_nj=120
|
||||
shadow_nj=60
|
||||
shadow2_nj=120
|
||||
eval_nj=120
|
||||
|
|
@ -96,6 +96,7 @@ use Getopt::Long;
|
|||
#
|
||||
########################################################################
|
||||
|
||||
print STDERR "$0 " . join(" ", @ARGV) . "\n";
|
||||
GetOptions("fragmentMarkers=s" => \$fragMarkers,
|
||||
"oov=s" => \$OOV_symbol,
|
||||
"vocab=s" => \$vocabFile,
|
||||
|
|
@ -165,7 +166,7 @@ if (-d $TranscriptionDir) {
|
|||
open (TRANSCRIPT, $inputspec) || die "Unable to open $filename";
|
||||
while ($line=<TRANSCRIPT>) {
|
||||
chomp $line;
|
||||
if ($line =~ m:^\[([0-9]+\.*[0-9]*)\]$:) {
|
||||
if ($line =~ m:^\s*\[([0-9]+\.*[0-9]*)\]\s*$:) {
|
||||
$thisTimeMark = $1;
|
||||
if ($thisTimeMark < $prevTimeMark) {
|
||||
print STDERR ("$0 ERROR: Found segment with negative duration in $filename\n");
|
||||
|
|
@ -245,6 +246,7 @@ if (-d $TranscriptionDir) {
|
|||
} else {
|
||||
# This is a just regular spoken word
|
||||
if ($vocabFile && (! $inVocab{$w}) && $fragMarkers) {
|
||||
print "Not in vocab: $w\n";
|
||||
# $w is a potential OOV token
|
||||
# Remove fragMarkers to see if $w becomes in-vocabulary
|
||||
while ($w =~ m:^(\S+[$fragMarkers]|[$fragMarkers]\S+)$:) {
|
||||
|
|
|
|||
|
|
@ -61,7 +61,7 @@ total_time=0
|
|||
t1=$(date +%s)
|
||||
|
||||
if [ $stage -le 0 ] ; then
|
||||
steps/decode_nolats.sh $decoder_extra_opts --write-words false --write-alignments true \
|
||||
steps/decode_nolats.sh ${decode_extra_opts+} --write-words false --write-alignments true \
|
||||
--cmd "$cmd" --nj $nj --beam $beam --max-active $max_active \
|
||||
$model_dir/phone_graph $datadir $model_dir/decode_${dirid} || exit 1
|
||||
fi
|
||||
|
|
@ -122,7 +122,7 @@ mkdir -p $output_dir
|
|||
mkdir -p $temp_dir/log
|
||||
|
||||
local/resegment/segmentation.py --verbose 2 $segmentation_opts \
|
||||
$temp_dir/pred $temp_dir/phone_map.txt 2> $temp_dir/log/resegment.log | \
|
||||
$temp_dir/pred $temp_dir/phone_map.txt 2>$temp_dir/log/resegment.log | \
|
||||
sort > $output_dir/segments || exit 1
|
||||
|
||||
if [ ! -s $output_dir/segments ] ; then
|
||||
|
|
|
|||
|
|
@ -31,8 +31,6 @@ skip_scoring=false
|
|||
extra_kws=false
|
||||
cmd=run.pl
|
||||
max_states=150000
|
||||
dev2shadow=
|
||||
eval2shadow=
|
||||
wip=0.5 #Word insertion penalty
|
||||
#End of options
|
||||
|
||||
|
|
@ -53,16 +51,6 @@ data_dir=$1;
|
|||
lang_dir=$2;
|
||||
decode_dir=$3;
|
||||
|
||||
type=normal
|
||||
if [ ! -z ${dev2shadow} ] && [ ! -z ${eval2shadow} ] ; then
|
||||
type=shadow
|
||||
elif [ -z ${dev2shadow} ] && [ -z ${eval2shadow} ] ; then
|
||||
type=normal
|
||||
else
|
||||
echo "Switches --dev2shadow and --eval2shadow must be used simultaneously" > /dev/stderr
|
||||
exit 1
|
||||
fi
|
||||
|
||||
##NB: The first ".done" files are used for backward compatibility only
|
||||
##NB: should be removed in a near future...
|
||||
if [ ! -f $decode_dir/.score.done ] && [ ! -f $decode_dir/.done.score ]; then
|
||||
|
|
@ -70,11 +58,7 @@ if [ ! -f $decode_dir/.score.done ] && [ ! -f $decode_dir/.done.score ]; then
|
|||
--min-lmwt ${min_lmwt} --max-lmwt ${max_lmwt} \
|
||||
$data_dir $lang_dir $decode_dir
|
||||
|
||||
if [[ "$type" == shadow* ]]; then
|
||||
local/split_ctms.sh --cmd "$cmd" --cer $cer \
|
||||
--min-lmwt ${min_lmwt} --max-lmwt ${max_lmwt}\
|
||||
$data_dir $decode_dir ${dev2shadow} ${eval2shadow}
|
||||
elif ! $skip_scoring ; then
|
||||
if ! $skip_scoring ; then
|
||||
local/score_stm.sh --cmd "$cmd" --cer $cer \
|
||||
--min-lmwt ${min_lmwt} --max-lmwt ${max_lmwt}\
|
||||
$data_dir $lang_dir $decode_dir
|
||||
|
|
@ -84,15 +68,9 @@ fi
|
|||
|
||||
if ! $skip_kws ; then
|
||||
if [ ! -f $decode_dir/.kws.done ] && [ ! -f $decode_dir/.done.kws ]; then
|
||||
if [[ "$type" == shadow* ]]; then
|
||||
local/shadow_set_kws_search.sh --cmd "$cmd" --max-states ${max_states} \
|
||||
--min-lmwt ${min_lmwt} --max-lmwt ${max_lmwt}\
|
||||
$data_dir $lang_dir $decode_dir ${dev2shadow} ${eval2shadow}
|
||||
else
|
||||
local/kws_search.sh --cmd "$cmd" --max-states ${max_states} \
|
||||
--min-lmwt ${min_lmwt} --max-lmwt ${max_lmwt} --skip-scoring $skip_scoring\
|
||||
--indices-dir $decode_dir/kws_indices $lang_dir $data_dir $decode_dir
|
||||
fi
|
||||
local/kws_search.sh --cmd "$cmd" --max-states ${max_states} \
|
||||
--min-lmwt ${min_lmwt} --max-lmwt ${max_lmwt} --skip-scoring $skip_scoring\
|
||||
--indices-dir $decode_dir/kws_indices $lang_dir $data_dir $decode_dir
|
||||
touch $decode_dir/.done.kws
|
||||
fi
|
||||
if $extra_kws && [ -f $data_dir/extra_kws_tasks ]; then
|
||||
|
|
|
|||
|
|
@ -0,0 +1,34 @@
|
|||
#!/bin/bash
|
||||
|
||||
. path.sh
|
||||
|
||||
format=pdf # pdf svg
|
||||
output=
|
||||
|
||||
. utils/parse_options.sh
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "usage: $0 [--format pdf|svg] [--output <path-to-output>] <utt-id> <lattice-ark> <word-list>"
|
||||
echo "e.g.: $0 utt-0001 \"test/lat.*.gz\" tri1/graph/words.txt"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
uttid=$1
|
||||
lat=$2
|
||||
words=$3
|
||||
|
||||
tmpdir=$(mktemp -d); trap "rm -r $tmpdir" EXIT # cleanup
|
||||
|
||||
gunzip -c $lat | lattice-to-fst ark:- ark,scp:$tmpdir/fst.ark,$tmpdir/fst.scp || exit 1
|
||||
! grep "^$uttid " $tmpdir/fst.scp && echo "ERROR : Missing utterance '$uttid' from gzipped lattice ark '$lat'" && exit 1
|
||||
fstcopy "scp:grep '^$uttid ' $tmpdir/fst.scp |" "scp:echo $uttid $tmpdir/$uttid.fst |" || exit 1
|
||||
fstdraw --portrait=true --osymbols=$words $tmpdir/$uttid.fst | dot -T${format} > $tmpdir/$uttid.${format}
|
||||
|
||||
if [ ! -z $output ]; then
|
||||
cp $tmpdir/$uttid.${format} $output
|
||||
fi
|
||||
|
||||
[ $format == "pdf" ] && evince $tmpdir/$uttid.pdf
|
||||
[ $format == "svg" ] && eog $tmpdir/$uttid.svg
|
||||
|
||||
exit 0
|
||||
|
|
@ -0,0 +1,121 @@
|
|||
#!/usr/bin/perl
|
||||
|
||||
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey). Apache 2.0.
|
||||
|
||||
#scalar(@ARGV) >= 1 && print STDERR "Usage: summarize_warnings.pl <log-dir>\n" && exit 1;
|
||||
|
||||
sub split_hundreds { # split list of filenames into groups of 100.
|
||||
my $names = shift @_;
|
||||
my @A = split(" ", $names);
|
||||
my @ans = ();
|
||||
while (@A > 0) {
|
||||
my $group = "";
|
||||
for ($x = 0; $x < 100 && @A>0; $x++) {
|
||||
$fname = pop @A;
|
||||
$group .= "$fname ";
|
||||
}
|
||||
push @ans, $group;
|
||||
}
|
||||
return @ans;
|
||||
}
|
||||
|
||||
sub parse_accounting_entry {
|
||||
$entry= shift @_;
|
||||
|
||||
@elems = split " ", $entry;
|
||||
|
||||
$time=undef;
|
||||
$threads=undef;
|
||||
foreach $elem (@elems) {
|
||||
if ( $elem=~ m/time=(\d+)/ ) {
|
||||
$elem =~ s/time=(\d+)/$1/;
|
||||
$time = $elem;
|
||||
} elsif ( $elem=~ m/threads=(\d+)/ ) {
|
||||
$elem =~ s/threads=(\d+)/$1/g;
|
||||
$threads = $elem;
|
||||
} else {
|
||||
die "Unknown entry \"$elem\" when parsing \"$entry\" \n";
|
||||
}
|
||||
}
|
||||
|
||||
if (defined($time) and defined($threads) ) {
|
||||
return ($time, $threads);
|
||||
} else {
|
||||
die "The accounting entry \"$entry\" did not contain all necessary attributes";
|
||||
}
|
||||
}
|
||||
|
||||
foreach $dir (@ARGV) {
|
||||
|
||||
#$dir = $ARGV[0];
|
||||
print $dir
|
||||
|
||||
! -d $dir && print STDERR "summarize_warnings.pl: no such directory $dir\n" ;
|
||||
|
||||
$dir =~ s:/$::; # Remove trailing slash.
|
||||
|
||||
|
||||
# Group the files into categories where all have the same base-name.
|
||||
foreach $f (glob ("$dir/*.log")) {
|
||||
$f_category = $f;
|
||||
# do next expression twice; s///g doesn't work as they overlap.
|
||||
$f_category =~ s:\.\d+\.(?!\d+):.*.:;
|
||||
#$f_category =~ s:\.\d+\.:.*.:;
|
||||
$fmap{$f_category} .= " $f";
|
||||
}
|
||||
}
|
||||
|
||||
foreach $c (sort (keys %fmap) ) {
|
||||
$n = 0;
|
||||
foreach $fgroup (split_hundreds($fmap{$c})) {
|
||||
$n += `grep -w WARNING $fgroup | wc -l`;
|
||||
}
|
||||
if ($n != 0) {
|
||||
print "$n warnings in $c\n"
|
||||
}
|
||||
}
|
||||
foreach $c (sort (keys %fmap)) {
|
||||
$n = 0;
|
||||
foreach $fgroup (split_hundreds($fmap{$c})) {
|
||||
$n += `grep -w ERROR $fgroup | wc -l`;
|
||||
}
|
||||
if ($n != 0) {
|
||||
print "$n errors in $c\n"
|
||||
}
|
||||
}
|
||||
|
||||
$supertotal_cpu_time=0.0;
|
||||
$supertotal_clock_time=0.0;
|
||||
$supertotal_threads=0.0;
|
||||
|
||||
foreach $c (sort (keys %fmap)) {
|
||||
$n = 0;
|
||||
|
||||
$total_cpu_time=0.0;
|
||||
$total_clock_time=0.0;
|
||||
$total_threads=0.0;
|
||||
foreach $fgroup (split_hundreds($fmap{$c})) {
|
||||
$lines=`grep -P "# Accounting:? " $fgroup |sed 's/.* Accounting:* *//g'`;
|
||||
|
||||
#print $lines ."\n";
|
||||
|
||||
@entries = split "\n", $lines;
|
||||
|
||||
foreach $line (@entries) {
|
||||
$time, $threads = parse_accounting_entry($line);
|
||||
|
||||
$total_cpu_time += $time * $threads;
|
||||
$total_threads += $threads;
|
||||
if ( $time > $total_clock_time ) {
|
||||
$total_clock_time = $time;
|
||||
}
|
||||
}
|
||||
}
|
||||
print "total_cpu_time=$total_cpu_time clock_time=$total_clock_time total_threads=$total_threads group=$c\n";
|
||||
|
||||
$supertotal_cpu_time += $total_cpu_time;
|
||||
$supertotal_clock_time += $total_clock_time;
|
||||
$supertotal_threads += $total_threads;
|
||||
}
|
||||
print "total_cpu_time=$supertotal_cpu_time clock_time=$supertotal_clock_time total_threads=$supertotal_threads group=all\n";
|
||||
|
||||
|
|
@ -1,57 +1,39 @@
|
|||
#!/bin/bash
|
||||
|
||||
lp=
|
||||
lr=
|
||||
ar=
|
||||
split=BaEval
|
||||
team=RADICAL
|
||||
corpusid=
|
||||
partition=
|
||||
scase=BaEval #BaDev|BaEval
|
||||
sysid=
|
||||
master=
|
||||
version=1
|
||||
relname=
|
||||
exp=c
|
||||
sysid=
|
||||
prim=c
|
||||
cer=0
|
||||
dryrun=true
|
||||
dir="exp/sgmm5_mmi_b0.1/"
|
||||
extrasys=""
|
||||
data=data/dev10h.seg
|
||||
master=dev10h
|
||||
final=false
|
||||
dev2shadow=dev10h.uem
|
||||
eval2shadow=eval.uem
|
||||
team=RADICAL
|
||||
|
||||
#end of configuration
|
||||
|
||||
|
||||
echo $0 " " "$@"
|
||||
|
||||
[ -f ./cmd.sh ] && . ./cmd.sh
|
||||
[ -f ./path.sh ] && . ./path.sh
|
||||
. ./utils/parse_options.sh
|
||||
|
||||
if [ $# -ne 2 ] ; then
|
||||
echo "Invalid number of parameters!"
|
||||
echo "Parameters " "$@"
|
||||
echo "$0 --ar <NTAR|TAR> --lr <BaseLR|BabelLR|OtherLR> --lp <FullLP|LimitedLP> --relname <NAME> [--version <version-nr> ] <config> <output>"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
|
||||
[ -z $lp ] && echo "Error -- you must specify --lp <FullLP|LimitedLP>" && exit 1
|
||||
if [ "$lp" != "FullLP" ] && [ "$lp" != "LimitedLP" ] ; then
|
||||
echo "Error -- you must specify --lp <FullLP|LimitedLP>" && exit 1
|
||||
fi
|
||||
|
||||
[ -z $lr ] && echo "Error -- you must specify --lr <BaseLR|BabelLR|OtherLR>" && exit 1
|
||||
if [ "$lr" != "BaseLR" ] && [ "$lr" != "BabelLR" ] && [ "$lr" != "OtherLR" ] ; then
|
||||
echo "Error -- you must specify --lr <BaseLR|BabelLR|OtherLR>" && exit 1
|
||||
fi
|
||||
[ -z $ar ] && echo "Error -- you must specify --ar <NTAR|TAR>" && exit 1
|
||||
if [ "$ar" != "NTAR" ] && [ "$ar" != "TAR" ] ; then
|
||||
echo "Error -- you must specify --ar <NTAR|TAR>" && exit 1
|
||||
fi
|
||||
[ -z $relname ] && echo "Error -- you must specify name" && exit 1
|
||||
|
||||
[ ! -f $1 ] && echo "Configuration $1 does not exist! " && exit 1
|
||||
. $1
|
||||
outputdir=$2
|
||||
|
||||
set -e
|
||||
set -o pipefail
|
||||
|
||||
function export_file {
|
||||
set -x
|
||||
#set -x
|
||||
source_file=$1
|
||||
target_file=$2
|
||||
if [ ! -f $source_file ] ; then
|
||||
|
|
@ -61,12 +43,17 @@ function export_file {
|
|||
if [ ! -f $target_file ] ; then
|
||||
if ! $dryrun ; then
|
||||
ln -s `readlink -f $source_file` $target_file || exit 1
|
||||
ls -al $target_file
|
||||
else
|
||||
echo "$source_file -> $target_file"
|
||||
fi
|
||||
|
||||
else
|
||||
echo "The file is already there, not doing anything. Either change the version (using --version), or delete that file manually)"
|
||||
exit 1
|
||||
fi
|
||||
fi
|
||||
#set +x
|
||||
return 0
|
||||
}
|
||||
|
||||
|
|
@ -97,38 +84,227 @@ function export_kws_file {
|
|||
return 0
|
||||
}
|
||||
|
||||
if [[ "$eval_kwlist_file" == *.kwlist.xml ]] ; then
|
||||
corpus=`basename $eval_kwlist_file .kwlist.xml`
|
||||
elif [[ "$eval_kwlist_file" == *.kwlist2.xml ]] ; then
|
||||
corpus=`basename $eval_kwlist_file .kwlist2.xml`
|
||||
function find_best_kws_result {
|
||||
local dir=$1
|
||||
local mask=$2
|
||||
local record=`(find $dir -name "sum.txt" -path "$mask" | xargs grep "^| *Occ") | cut -f 1,13,17 -d '|' | sed 's/|//g' | column -t | sort -r -n -k 3 | tail -n 1`
|
||||
echo $record >&2
|
||||
local file=`echo $record | awk -F ":" '{print $1}'`
|
||||
#echo $file >&2
|
||||
local path=`dirname $file`
|
||||
#echo $path >&2
|
||||
echo $path
|
||||
}
|
||||
|
||||
function find_best_stt_result {
|
||||
local dir=$1
|
||||
local mask=$2
|
||||
local record=`(find $dir -name "*.ctm.sys" -path "$mask" | xargs grep Avg) | sed 's/|//g' | column -t | sort -n -k 9 | head -n 1`
|
||||
|
||||
echo $record >&2
|
||||
local file=`echo $record | awk -F ":" '{print $1}'`
|
||||
#echo $file >&2
|
||||
local path=`dirname $file`
|
||||
#echo $path >&2
|
||||
echo $path
|
||||
}
|
||||
|
||||
function create_sysid {
|
||||
local best_one=$1
|
||||
local extrasys=$2
|
||||
local taskid=`basename $best_one`
|
||||
local system_path=`dirname $best_one`
|
||||
if [[ $system_path =~ .*sgmm5.* ]] ; then
|
||||
sysid=PLP
|
||||
elif [[ $system_path =~ .*nnet.* ]] ; then
|
||||
sysid=DNN
|
||||
elif [[ $system_path =~ .*sgmm7.* ]] ; then
|
||||
sysid=BNF
|
||||
else
|
||||
echo "Unknown system path ($system_path), cannot deduce the systemID" >&2
|
||||
exit 1
|
||||
fi
|
||||
if [ ! -z $extrasys ]; then
|
||||
sysid="${sysid}-${extrasys}"
|
||||
fi
|
||||
local kwsid=${taskid//kws_*/}
|
||||
kwsid=${kwsid//_/}
|
||||
if [ -z $kwsid ]; then
|
||||
echo ${sysid}
|
||||
else
|
||||
echo ${sysid}-$kwsid
|
||||
fi
|
||||
}
|
||||
|
||||
function get_ecf_name {
|
||||
local best_one=$1
|
||||
local taskid=`basename $best_one`
|
||||
local kwstask=${taskid//kws_*/kws}
|
||||
local kwlist=
|
||||
#echo $kwstask
|
||||
if [ -z $kwstask ] ; then
|
||||
#echo $data/kws/kwlist.xml
|
||||
kwlist= `readlink -f $data/kws/kwlist.xml`
|
||||
else
|
||||
#echo $data/$kwstask/kwlist.xml
|
||||
kwlist=`readlink -f $data/$kwstask/kwlist.xml`
|
||||
fi
|
||||
ecf=`head -n 1 $kwlist | grep -Po "(?<=ecf_filename=\")[^\"]*"`
|
||||
echo -e "\tFound ECF: $ecf" >&2
|
||||
echo $ecf
|
||||
return 0
|
||||
}
|
||||
|
||||
function compose_expid {
|
||||
local task=$1
|
||||
local best_one=$2
|
||||
local extraid=$3
|
||||
[ ! -z $extraid ] && extraid="-$extraid"
|
||||
local sysid=`create_sysid $best_one $extrasys`
|
||||
if [ "$task" == "KWS" ]; then
|
||||
ext="kwslist.xml"
|
||||
elif [ "$task" == "STT" ]; then
|
||||
ext="ctm"
|
||||
else
|
||||
echo "Incorrect task ID ($task) given to compose_expid function!" >&2
|
||||
exit 1
|
||||
fi
|
||||
echo "KWS14_${team}_${corpusid}_${partition}_${scase}_${task}_${prim}-${sysid}${extraid}_$version.$ext"
|
||||
return 0
|
||||
}
|
||||
|
||||
function figure_out_scase {
|
||||
local ecf=`basename $1`
|
||||
if [[ $ecf =~ IARPA-babel.*.ecf.xml ]] ; then
|
||||
local basnam=${ecf%%.ecf.xml}
|
||||
local scase=`echo $basnam | awk -F _ '{print $2}'`
|
||||
|
||||
if [ "$scase" = "conv-dev" ]; then
|
||||
echo "BaDev"
|
||||
elif [ "$scase" = "conv-eval" ]; then
|
||||
echo "BaEval"
|
||||
else
|
||||
echo "WARNING: The ECF file $ecf is probably not an official file" >&2
|
||||
echo "BaDev"
|
||||
return 1
|
||||
fi
|
||||
else
|
||||
echo "WARNING: The ECF file $ecf is probably not an official file" >&2
|
||||
echo "BaDev"
|
||||
return 1
|
||||
fi
|
||||
return 0
|
||||
}
|
||||
|
||||
function figure_out_partition {
|
||||
local ecf=`basename $1`
|
||||
if [[ $ecf =~ IARPA-babel.*.ecf.xml ]] ; then
|
||||
local basnam=${ecf%%.ecf.xml}
|
||||
local scase=`echo $basnam | awk -F _ '{print $2}'`
|
||||
|
||||
if [ "$scase" = "conv-dev" ]; then
|
||||
echo "conv-dev"
|
||||
elif [ "$scase" = "conv-eval" ]; then
|
||||
echo "conv-eval"
|
||||
else
|
||||
echo "WARNING: The ECF file $ecf is probably not an official file" >&2
|
||||
echo "conv-dev"
|
||||
return 1
|
||||
fi
|
||||
else
|
||||
echo "WARNING: The ECF file $ecf is probably not an official file" >&2
|
||||
echo "conv-dev"
|
||||
return 1
|
||||
fi
|
||||
return 0
|
||||
}
|
||||
|
||||
function figure_out_corpusid {
|
||||
local ecf=`basename $1`
|
||||
if [[ $ecf =~ IARPA-babel.*.ecf.xml ]] ; then
|
||||
local basnam=${ecf%%.ecf.xml}
|
||||
local corpusid=`echo $basnam | awk -F _ '{print $1}'`
|
||||
else
|
||||
echo "WARNING: The ECF file $ecf is probably not an official file" >&2
|
||||
local corpusid=${ecf%%.*}
|
||||
fi
|
||||
echo $corpusid
|
||||
}
|
||||
|
||||
#data=data/shadow.uem
|
||||
dirid=`basename $data`
|
||||
kws_tasks="kws "
|
||||
[ -f $data/extra_kws_tasks ] && kws_tasks+=`cat $data/extra_kws_tasks | awk '{print $1"_kws"}'`
|
||||
[ -d $data/compounds ] && compounds=`ls $data/compounds`
|
||||
|
||||
if [ -z "$compounds" ] ; then
|
||||
for kws in $kws_tasks ; do
|
||||
echo $kws
|
||||
best_one=`find_best_kws_result "$dir/decode_*${dirid}*/${kws}_*" "*"`
|
||||
sysid=`create_sysid $best_one $extrasys`
|
||||
ecf=`get_ecf_name $best_one`
|
||||
scase=`figure_out_scase $ecf` || break
|
||||
partition=`figure_out_partition $ecf` || break
|
||||
corpusid=`figure_out_corpusid $ecf`
|
||||
echo -e "\tEXPORT as:" `compose_expid KWS $best_one`
|
||||
done
|
||||
else
|
||||
echo "Unknown naming pattern of the kwlist file $eval_kwlist_file"
|
||||
exit 1
|
||||
[ -z $master ] && echo "You must choose the master compound (--master <compound>) for compound data set" && exit 1
|
||||
for kws in $kws_tasks ; do
|
||||
echo $kws
|
||||
best_one=`find_best_kws_result "$dir/decode_*${dirid}*/$master/${kws}_*" "*"`
|
||||
|
||||
for compound in $compounds ; do
|
||||
compound_best_one=`echo $best_one | sed ":$master/${kws}_:$compound/${kws}_:g"`
|
||||
echo -e "\tPREPARE EXPORT: $compound_best_one"
|
||||
sysid=`create_sysid $compound_best_one $extrasys`
|
||||
#ecf=`get_ecf_name $best_one`
|
||||
ecf=`readlink -f $data/compounds/$compound/ecf.xml`
|
||||
scase=`figure_out_scase $ecf`
|
||||
partition=`figure_out_partition $ecf`
|
||||
corpusid=`figure_out_corpusid $ecf`
|
||||
expid=`compose_expid KWS $compound_best_one`
|
||||
echo -e "\tEXPORT NORMALIZED as: $expid"
|
||||
expid_unnormalized=`compose_expid KWS $compound_best_one "unnorm"`
|
||||
echo -e "\tEXPORT UNNORMALIZED as: $expid_unnormalized"
|
||||
|
||||
export_kws_file $compound_best_one/kwslist.xml $compound_best_one/kwslist.fixed.xml $data/$kws/kwlist.xml $outputdir/$expid
|
||||
export_kws_file $compound_best_one/kwslist.unnormalized.xml $compound_best_one/kwslist.unnormalized.fixed.xml $data/$kws/kwlist.xml $outputdir/$expid_unnormalized
|
||||
done
|
||||
done
|
||||
fi
|
||||
#REMOVE the IARPA- prefix, if present
|
||||
#corpus=${corpora##IARPA-}
|
||||
|
||||
scores=`find -L $dir -name "sum.txt" -path "*${dev2shadow}_${eval2shadow}*" | xargs grep "| Occurrence" | cut -f 1,13 -d '|'| sed 's/:|//g' | column -t | sort -k 2 -n -r `
|
||||
[ -z "$scores" ] && echo "Nothing to export, exiting..." && exit 1
|
||||
##EXporting STT -- more straightforward, because there is only one task
|
||||
if [ -z "$compounds" ] ; then
|
||||
best_one=`find_best_stt_result "$dir/decode_*${dirid}*/score_*" "*"`
|
||||
echo -e "\tERROR: I don't know how to do this, yet"
|
||||
ecf=`get_ecf_name kws`
|
||||
sysid=`create_sysid $best_one $extrasys`
|
||||
scase=`figure_out_scase $ecf` || break
|
||||
partition=`figure_out_partition $ecf`
|
||||
corpusid=`figure_out_corpusid $ecf`
|
||||
expid=`compose_expid STT $best_one`
|
||||
echo -e "\tEXPORT NORMALIZED as: $expid"
|
||||
export_file $best_one/${dirid}.ctm $outputdir/$expid
|
||||
else
|
||||
[ -z $master ] && echo "You must choose the master compound (--master <compound>) for compound data set" && exit 1
|
||||
best_one=`find_best_stt_result "exp/sgmm5_mmi_b0.1/decode_*${dirid}*/$master/score_*" "*"`
|
||||
|
||||
echo "$scores" | head
|
||||
count=`echo "$scores" | wc -l`
|
||||
echo "Total result files: $count"
|
||||
best_score=`echo "$scores" | head -n 1 | cut -f 1 -d ' '`
|
||||
for compound in $compounds ; do
|
||||
compound_best_one=`echo $best_one | sed ":$master/${kws}_:$compound/${kws}_:g"`
|
||||
echo -e "\tPREPARE EXPORT: $compound_best_one"
|
||||
sysid=`create_sysid $compound_best_one $extrasys`
|
||||
#ecf=`get_ecf_name $best_one`
|
||||
ecf=`readlink -f $data/compounds/$compound/ecf.xml`
|
||||
scase=`figure_out_scase $ecf`
|
||||
partition=`figure_out_partition $ecf`
|
||||
corpusid=`figure_out_corpusid $ecf`
|
||||
expid=`compose_expid STT $compound_best_one`
|
||||
echo -e "\tEXPORT NORMALIZED as: $expid"
|
||||
|
||||
lmwt=`echo $best_score | sed 's:.*/kws_\([0-9][0-9]*\)/.*:\1:g'`
|
||||
echo "Best scoring file: $best_score"
|
||||
echo $lmwt
|
||||
base_dir=`echo $best_score | sed "s:\\(.*\\)/${dev2shadow}_${eval2shadow}/.*:\\1:g"`
|
||||
echo $base_dir
|
||||
|
||||
eval_dir=$base_dir/$eval2shadow/kws_$lmwt/
|
||||
eval_kwlist=$eval_dir/kwslist.xml
|
||||
eval_fixed_kwlist=$eval_dir/kwslist.fixed.xml
|
||||
eval_export_kwlist=$outputdir/KWS13_${team}_${corpus}_${split}_KWS_${lp}_${lr}_${ar}_${relname}_${version}.kwslist.xml
|
||||
|
||||
echo "export_kws_file $eval_kwlist $eval_fixed_kwlist $eval_kwlist_file $eval_export_kwlist"
|
||||
export_kws_file $eval_kwlist $eval_fixed_kwlist $eval_kwlist_file $eval_export_kwlist
|
||||
export_file $compound_best_one/${compound}.ctm $outputdir/$expid
|
||||
done
|
||||
fi
|
||||
|
||||
echo "Everything looks fine, good luck!"
|
||||
exit 0
|
||||
|
|
|
|||
|
|
@ -100,7 +100,7 @@ if [[ ! -f $lexicon || $lexicon -ot "$lexicon_file" ]]; then
|
|||
# Extend the original lexicon.
|
||||
# Will creates the files data/local/extend/{lexiconp.txt,oov2prob}.
|
||||
mv data/local/lexicon.txt data/local/lexicon_orig.txt
|
||||
local/extend_lexicon.sh --cmd "$train_cmd" \
|
||||
local/extend_lexicon.sh --cmd "$train_cmd" --cleanup false \
|
||||
--num-sent-gen $num_sent_gen --num-prons $num_prons \
|
||||
data/local/lexicon_orig.txt data/local/extend data/dev2h/text
|
||||
cp data/local/extend/lexiconp.txt data/local/
|
||||
|
|
@ -153,7 +153,7 @@ if [[ ! -f data/lang/G.fst || data/lang/G.fst -ot data/srilm/lm.gz ||\
|
|||
[ -f data/local/extend/original_oov_rates ] || exit 1;
|
||||
unk_fraction=`cat data/local/extend/original_oov_rates |\
|
||||
grep "token" | awk -v x=$unk_fraction_boost '{print $NF/100.0*x}'`
|
||||
extend_lexicon_param=(--unk-fraction $unk_fraction \
|
||||
extend_lexicon_param=(--cleanup false --unk-fraction $unk_fraction \
|
||||
--oov-prob-file data/local/extend/oov2prob)
|
||||
fi
|
||||
local/arpa2G.sh ${extend_lexicon_param[@]} \
|
||||
|
|
@ -177,6 +177,11 @@ if [ ! -f data/train/.plp.done ]; then
|
|||
touch data/train/.plp.done
|
||||
fi
|
||||
|
||||
echo -------------------------------------------------------------------------
|
||||
echo "Extended lexicon finished on" `date`. Now run the script run-1-main.sh
|
||||
echo -------------------------------------------------------------------------
|
||||
exit 0
|
||||
|
||||
mkdir -p exp
|
||||
|
||||
if [ ! -f data/train_sub3/.done ]; then
|
||||
|
|
@ -199,6 +204,11 @@ if [ ! -f data/train_sub3/.done ]; then
|
|||
touch data/train_sub3/.done
|
||||
fi
|
||||
|
||||
echo "------------------------------------------------------------------"
|
||||
echo "Now run the script run-1-main.sh"
|
||||
echo "------------------------------------------------------------------"
|
||||
exit 0
|
||||
|
||||
if [ ! -f exp/mono/.done ]; then
|
||||
echo ---------------------------------------------------------------------
|
||||
echo "Starting (small) monophone training in exp/mono on" `date`
|
||||
|
|
|
|||
|
|
@ -2,6 +2,7 @@
|
|||
|
||||
# This is not necessarily the top-level run.sh as it is in other directories. see README.txt first.
|
||||
tri5_only=false
|
||||
sgmm5_only=false
|
||||
|
||||
[ ! -f ./lang.conf ] && echo 'Language configuration does not exist! Use the configurations in conf/lang/* as a startup' && exit 1
|
||||
[ ! -f ./conf/common_vars.sh ] && echo 'the file conf/common_vars.sh does not exist!' && exit 1
|
||||
|
|
@ -287,6 +288,11 @@ if [ ! -f exp/sgmm5/.done ]; then
|
|||
touch exp/sgmm5/.done
|
||||
fi
|
||||
|
||||
if $sgmm5_only ; then
|
||||
echo "Exiting after stage SGMM5, as requested. "
|
||||
echo "Everything went fine. Done"
|
||||
exit 0;
|
||||
fi
|
||||
################################################################################
|
||||
# Ready to start discriminative SGMM training
|
||||
################################################################################
|
||||
|
|
|
|||
|
|
@ -5,6 +5,7 @@
|
|||
|
||||
# This parameter will be used when the training dies at a certain point.
|
||||
train_stage=-100
|
||||
dir=exp/tri6_nnet
|
||||
. ./utils/parse_options.sh
|
||||
|
||||
set -e
|
||||
|
|
@ -17,7 +18,7 @@ echo "Waiting till exp/tri5_ali/.done exists...."
|
|||
while [ ! -f exp/tri5_ali/.done ]; do sleep 30; done
|
||||
echo "...done waiting for exp/tri5_ali/.done"
|
||||
|
||||
if [ ! -f exp/tri6_nnet/.done ]; then
|
||||
if [ ! -f $dir/.done ]; then
|
||||
steps/nnet2/train_pnorm.sh \
|
||||
--stage $train_stage --mix-up $dnn_mixup \
|
||||
--initial-learning-rate $dnn_init_learning_rate \
|
||||
|
|
@ -27,7 +28,7 @@ if [ ! -f exp/tri6_nnet/.done ]; then
|
|||
--pnorm-output-dim $dnn_output_dim \
|
||||
--cmd "$train_cmd" \
|
||||
"${dnn_cpu_parallel_opts[@]}" \
|
||||
data/train data/lang exp/tri5_ali exp/tri6_nnet || exit 1
|
||||
data/train data/lang exp/tri5_ali $dir || exit 1
|
||||
|
||||
touch exp/tri6_nnet/.done
|
||||
touch $dir/.done
|
||||
fi
|
||||
|
|
|
|||
|
|
@ -1,7 +1,12 @@
|
|||
#!/bin/bash
|
||||
|
||||
. conf/common_vars.sh
|
||||
. ./lang.conf
|
||||
. conf/common_vars.sh
|
||||
|
||||
train_stage=-10
|
||||
dir=exp/tri6b_nnet
|
||||
|
||||
. ./utils/parse_options.sh
|
||||
|
||||
set -e
|
||||
set -o pipefail
|
||||
|
|
@ -12,12 +17,10 @@ dnn_pnorm_input_dim=3000
|
|||
dnn_pnorm_output_dim=300
|
||||
dnn_init_learning_rate=0.004
|
||||
dnn_final_learning_rate=0.001
|
||||
train_stage=-10
|
||||
temp_dir=`pwd`/nnet_gpu_egs
|
||||
ensemble_size=4
|
||||
initial_beta=0.1
|
||||
final_beta=5
|
||||
dir=exp/tri6b_nnet
|
||||
egs_dir=
|
||||
|
||||
# Wait till the main run.sh gets to the stage where's it's
|
||||
|
|
|
|||
|
|
@ -1,4 +1,6 @@
|
|||
#!/bin/bash
|
||||
dir=exp/tri6_nnet
|
||||
train_stage=-10
|
||||
|
||||
. conf/common_vars.sh
|
||||
. ./lang.conf
|
||||
|
|
@ -17,7 +19,7 @@ echo "Waiting till exp/tri5_ali/.done exists...."
|
|||
while [ ! -f exp/tri5_ali/.done ]; do sleep 30; done
|
||||
echo "...done waiting for exp/tri5_ali/.done"
|
||||
|
||||
if [ ! -f exp/tri6_nnet/.done ]; then
|
||||
if [ ! -f $dir/.done ]; then
|
||||
steps/nnet2/train_pnorm.sh \
|
||||
--stage $train_stage --mix-up $dnn_mixup \
|
||||
--initial-learning-rate $dnn_init_learning_rate \
|
||||
|
|
@ -27,8 +29,8 @@ if [ ! -f exp/tri6_nnet/.done ]; then
|
|||
--pnorm-output-dim $dnn_output_dim \
|
||||
--cmd "$train_cmd" \
|
||||
"${dnn_gpu_parallel_opts[@]}" \
|
||||
data/train data/lang exp/tri5_ali exp/tri6_nnet || exit 1
|
||||
data/train data/lang exp/tri5_ali $dir || exit 1
|
||||
|
||||
touch exp/tri6_nnet/.done
|
||||
touch $dir/.done
|
||||
fi
|
||||
|
||||
|
|
|
|||
|
|
@ -7,8 +7,6 @@ set -o pipefail
|
|||
|
||||
|
||||
dir=dev10h.pem
|
||||
dev2shadow=dev10h.uem
|
||||
eval2shadow=eval.uem
|
||||
kind=
|
||||
data_only=false
|
||||
fast_path=true
|
||||
|
|
@ -19,7 +17,6 @@ max_states=150000
|
|||
extra_kws=true
|
||||
vocab_kws=false
|
||||
wip=0.5
|
||||
shadow_set_extra_opts=( --wip $wip )
|
||||
|
||||
echo "run-4-test.sh $@"
|
||||
|
||||
|
|
@ -46,8 +43,6 @@ dataset_type=${dir%%.*}
|
|||
if [ -z ${kind} ] ; then
|
||||
if [ "$dataset_type" == "dev2h" ] || [ "$dataset_type" == "dev10h" ] ; then
|
||||
dataset_kind=supervised
|
||||
elif [ "$dataset_type" == "shadow" ] ; then
|
||||
dataset_kind=shadow
|
||||
else
|
||||
dataset_kind=unsupervised
|
||||
fi
|
||||
|
|
@ -139,55 +134,33 @@ function check_variables_are_set {
|
|||
fi
|
||||
}
|
||||
|
||||
if [ "$dataset_kind" == "shadow" ] ; then
|
||||
# we expect that the ${dev2shadow} as well as ${eval2shadow} already exist
|
||||
if [ ! -f data/${dev2shadow}/.done ]; then
|
||||
echo "Error: data/${dev2shadow}/.done does not exist."
|
||||
echo "Create the directory data/${dev2shadow} first"
|
||||
echo "e.g. by calling $0 --type $dev2shadow --dataonly"
|
||||
exit 1
|
||||
fi
|
||||
if [ ! -f data/${eval2shadow}/.done ]; then
|
||||
echo "Error: data/${eval2shadow}/.done does not exist."
|
||||
echo "Create the directory data/${eval2shadow} first."
|
||||
echo "e.g. by calling $0 --type $eval2shadow --dataonly"
|
||||
if [ ! -f data/raw_${dataset_type}_data/.done ]; then
|
||||
echo ---------------------------------------------------------------------
|
||||
echo "Subsetting the ${dataset_type} set"
|
||||
echo ---------------------------------------------------------------------
|
||||
|
||||
l1=${#my_data_dir[*]}
|
||||
l2=${#my_data_list[*]}
|
||||
if [ "$l1" -ne "$l2" ]; then
|
||||
echo "Error, the number of source files lists is not the same as the number of source dirs!"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
local/create_shadow_dataset.sh ${dataset_dir} \
|
||||
data/${dev2shadow} data/${eval2shadow}
|
||||
utils/fix_data_dir.sh ${datadir}
|
||||
nj_max=`cat $dataset_dir/wav.scp | wc -l`
|
||||
my_nj=64
|
||||
else
|
||||
if [ ! -f data/raw_${dataset_type}_data/.done ]; then
|
||||
echo ---------------------------------------------------------------------
|
||||
echo "Subsetting the ${dataset_type} set"
|
||||
echo ---------------------------------------------------------------------
|
||||
|
||||
l1=${#my_data_dir[*]}
|
||||
l2=${#my_data_list[*]}
|
||||
if [ "$l1" -ne "$l2" ]; then
|
||||
echo "Error, the number of source files lists is not the same as the number of source dirs!"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
resource_string=""
|
||||
if [ "$dataset_kind" == "unsupervised" ]; then
|
||||
resource_string+=" --ignore-missing-txt true"
|
||||
fi
|
||||
|
||||
for i in `seq 0 $(($l1 - 1))`; do
|
||||
resource_string+=" ${my_data_dir[$i]} "
|
||||
resource_string+=" ${my_data_list[$i]} "
|
||||
done
|
||||
local/make_corpus_subset.sh $resource_string ./data/raw_${dataset_type}_data
|
||||
touch data/raw_${dataset_type}_data/.done
|
||||
resource_string=""
|
||||
if [ "$dataset_kind" == "unsupervised" ]; then
|
||||
resource_string+=" --ignore-missing-txt true"
|
||||
fi
|
||||
my_data_dir=`readlink -f ./data/raw_${dataset_type}_data`
|
||||
[ -f $my_data_dir/filelist.list ] && my_data_list=$my_data_dir/filelist.list
|
||||
nj_max=`cat $my_data_list | wc -l` || nj_max=`ls $my_data_dir/audio | wc -l`
|
||||
|
||||
for i in `seq 0 $(($l1 - 1))`; do
|
||||
resource_string+=" ${my_data_dir[$i]} "
|
||||
resource_string+=" ${my_data_list[$i]} "
|
||||
done
|
||||
local/make_corpus_subset.sh $resource_string ./data/raw_${dataset_type}_data
|
||||
touch data/raw_${dataset_type}_data/.done
|
||||
fi
|
||||
my_data_dir=`readlink -f ./data/raw_${dataset_type}_data`
|
||||
[ -f $my_data_dir/filelist.list ] && my_data_list=$my_data_dir/filelist.list
|
||||
nj_max=`cat $my_data_list | wc -l` || nj_max=`ls $my_data_dir/audio | wc -l`
|
||||
if [ "$nj_max" -lt "$my_nj" ] ; then
|
||||
echo "Number of jobs ($my_nj) is too big!"
|
||||
echo "The maximum reasonable number of jobs is $nj_max"
|
||||
|
|
@ -234,10 +207,6 @@ if [ ! -f $dataset_dir/.done ] ; then
|
|||
echo "Valid dataset types are: seg, uem, pem";
|
||||
exit 1
|
||||
fi
|
||||
elif [ "$dataset_kind" == "shadow" ] ; then
|
||||
#We don't actually have to do anything here
|
||||
#The shadow dir is already set...
|
||||
true
|
||||
else
|
||||
echo "Unknown kind of the dataset: \"$dataset_kind\"!";
|
||||
echo "Valid dataset kinds are: supervised, unsupervised, shadow";
|
||||
|
|
@ -303,13 +272,13 @@ if ! $fast_path ; then
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_plp_extra_opts[@]}" \
|
||||
"${lmwt_plp_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang ${decode}
|
||||
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_plp_extra_opts[@]}" \
|
||||
"${lmwt_plp_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang ${decode}.si
|
||||
fi
|
||||
|
||||
|
|
@ -337,7 +306,7 @@ if [ -f exp/sgmm5/.done ]; then
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_plp_extra_opts[@]}" \
|
||||
"${lmwt_plp_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang exp/sgmm5/decode_fmllr_${dataset_id}
|
||||
fi
|
||||
fi
|
||||
|
|
@ -371,7 +340,7 @@ if [ -f exp/sgmm5/.done ]; then
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_plp_extra_opts[@]}" \
|
||||
"${lmwt_plp_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang $decode
|
||||
done
|
||||
fi
|
||||
|
|
@ -397,7 +366,7 @@ if [ -f exp/tri6_nnet/.done ]; then
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_dnn_extra_opts[@]}" \
|
||||
"${lmwt_dnn_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang $decode
|
||||
fi
|
||||
|
||||
|
|
@ -423,7 +392,7 @@ if [ -f exp/tri6a_nnet/.done ]; then
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_dnn_extra_opts[@]}" \
|
||||
"${lmwt_dnn_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang $decode
|
||||
fi
|
||||
|
||||
|
|
@ -447,6 +416,31 @@ if [ -f exp/tri6b_nnet/.done ]; then
|
|||
touch $decode/.done
|
||||
fi
|
||||
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${lmwt_dnn_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang $decode
|
||||
fi
|
||||
####################################################################
|
||||
##
|
||||
## DNN (ensemble) decoding
|
||||
##
|
||||
####################################################################
|
||||
if [ -f exp/tri6c_nnet/.done ]; then
|
||||
decode=exp/tri6c_nnet/decode_${dataset_id}
|
||||
if [ ! -f $decode/.done ]; then
|
||||
mkdir -p $decode
|
||||
steps/nnet2/decode.sh \
|
||||
--minimize $minimize --cmd "$decode_cmd" --nj $my_nj \
|
||||
--beam $dnn_beam --lat-beam $dnn_lat_beam \
|
||||
--skip-scoring true "${decode_extra_opts[@]}" \
|
||||
--transform-dir exp/tri5/decode_${dataset_id} \
|
||||
exp/tri5/graph ${dataset_dir} $decode | tee $decode/decode.log
|
||||
|
||||
touch $decode/.done
|
||||
fi
|
||||
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
|
|
@ -476,7 +470,7 @@ if [ -f exp/tri6_nnet_mpe/.done ]; then
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_dnn_extra_opts[@]}" \
|
||||
"${lmwt_dnn_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang $decode
|
||||
done
|
||||
fi
|
||||
|
|
@ -505,7 +499,7 @@ for dnn in tri6_nnet_semi_supervised tri6_nnet_semi_supervised2 \
|
|||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--skip-scoring $skip_scoring --extra-kws $extra_kws --wip $wip \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_dnn_extra_opts[@]}" \
|
||||
"${lmwt_dnn_extra_opts[@]}" \
|
||||
${dataset_dir} data/lang $decode
|
||||
fi
|
||||
done
|
||||
|
|
|
|||
|
|
@ -17,7 +17,7 @@ skip_scoring=false
|
|||
max_states=150000
|
||||
wip=0.5
|
||||
|
||||
echo "$0 $@"
|
||||
echo "run-5-test.sh $@"
|
||||
|
||||
. utils/parse_options.sh
|
||||
|
||||
|
|
@ -314,32 +314,34 @@ fi
|
|||
##
|
||||
####################################################################
|
||||
|
||||
for iter in 1 2 3 4; do
|
||||
decode=exp/sgmm5/decode_fmllr_${dirid}
|
||||
if [ ! -f $decode/.done ]; then
|
||||
for iter in 1 2 3 4; do
|
||||
# Decode SGMM+MMI (via rescoring).
|
||||
decode=exp/sgmm5_mmi_b0.1/decode_fmllr_${dirid}_it$iter
|
||||
if [ ! -f $decode/.done ]; then
|
||||
decode=exp/sgmm5_mmi_b0.1/decode_fmllr_${dirid}_it$iter
|
||||
if [ ! -f $decode/.done ]; then
|
||||
|
||||
mkdir -p $decode
|
||||
steps/decode_sgmm2_rescore.sh --skip-scoring true \
|
||||
--cmd "$decode_cmd" --iter $iter --transform-dir exp/tri5/decode_${dirid} \
|
||||
data/lang ${datadir} exp/sgmm5/decode_fmllr_${dirid} $decode | tee ${decode}/decode.log
|
||||
mkdir -p $decode
|
||||
steps/decode_sgmm2_rescore.sh --skip-scoring true \
|
||||
--cmd "$decode_cmd" --iter $iter --transform-dir exp/tri5/decode_${dirid} \
|
||||
data/lang ${datadir} exp/sgmm5/decode_fmllr_${dirid} $decode | tee ${decode}/decode.log
|
||||
|
||||
touch $decode/.done
|
||||
fi
|
||||
done
|
||||
touch $decode/.done
|
||||
fi
|
||||
done
|
||||
|
||||
#We are done -- all lattices has been generated. We have to
|
||||
#a)Run MBR decoding
|
||||
#b)Run KW search
|
||||
for iter in 1 2 3 4; do
|
||||
for iter in 1 2 3 4; do
|
||||
# Decode SGMM+MMI (via rescoring).
|
||||
decode=exp/sgmm5_mmi_b0.1/decode_fmllr_${dirid}_it$iter
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states --skip-scoring $skip_scoring\
|
||||
decode=exp/sgmm5_mmi_b0.1/decode_fmllr_${dirid}_it$iter
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states --skip-scoring $skip_scoring\
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt --wip $wip \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_plp_extra_opts[@]}" \
|
||||
${datadir} data/lang $decode
|
||||
done
|
||||
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_plp_extra_opts[@]}" \
|
||||
${datadir} data/lang $decode
|
||||
done
|
||||
fi
|
||||
|
||||
####################################################################
|
||||
##
|
||||
|
|
|
|||
|
|
@ -16,6 +16,7 @@ set -u #Fail on an undefined variable
|
|||
skip_kws=true
|
||||
skip_stt=false
|
||||
semisupervised=true
|
||||
unsup_string="_semisup"
|
||||
bnf_train_stage=-100
|
||||
bnf_weight_threshold=0.35
|
||||
ali_dir=exp/tri6_nnet_ali
|
||||
|
|
@ -31,7 +32,6 @@ fi
|
|||
|
||||
|
||||
if $semisupervised ; then
|
||||
unsup_string="_semi_supervised"
|
||||
egs_string="--egs-dir exp_bnf${unsup_string}/tri6_bnf/egs"
|
||||
else
|
||||
unsup_string="" #" ": supervised training, _semi_supervised: unsupervised BNF training
|
||||
|
|
@ -43,6 +43,22 @@ datadir=data/${dirid}
|
|||
exp_dir=exp_bnf${unsup_string}
|
||||
data_bnf_dir=data_bnf${unsup_string}
|
||||
param_bnf_dir=param_bnf${unsup_string}
|
||||
|
||||
if [ -z $ali_dir ] ; then
|
||||
# If alignment directory is not done, use exp/tri6_nnet_ali as alignment
|
||||
# directory
|
||||
ali_dir=exp/tri6_nnet_ali
|
||||
fi
|
||||
|
||||
if [ ! -f $ali_dir/.done ]; then
|
||||
echo "$0: Aligning supervised training data in exp/tri6_nnet_ali"
|
||||
[ ! -f exp/tri6_nnet/final.mdl ] && echo "exp/tri6_nnet/final.mdl not found!\nRun run-6-nnet.sh first!" && exit 1
|
||||
steps/nnet2/align.sh --cmd "$train_cmd" \
|
||||
--use-gpu no --transform-dir exp/tri5_ali --nj $train_nj \
|
||||
data/train data/lang exp/tri6_nnet $ali_dir || exit 1
|
||||
touch $ali_dir/.done
|
||||
fi
|
||||
|
||||
###############################################################################
|
||||
#
|
||||
# Semi-supervised BNF training
|
||||
|
|
@ -50,22 +66,22 @@ param_bnf_dir=param_bnf${unsup_string}
|
|||
###############################################################################
|
||||
[ ! -d $datadir ] && echo "Error: $datadir is not available!" && exit 1;
|
||||
mkdir -p $exp_dir/tri6_bnf
|
||||
if $semisupervised ; then
|
||||
echo "$0: Generate examples using unsupervised data in $exp_dir/tri6_nnet"
|
||||
if [ ! -f $exp_dir/tri6_bnf/egs/.done ]; then
|
||||
local/nnet2/get_egs_semi_supervised.sh \
|
||||
"${egs_cpu_opts[@]}" --io-opts "$egs_io_opts" \
|
||||
--transform-dir-sup exp/tri5_ali \
|
||||
--transform-dir-unsup exp/tri5/decode_${dirid} \
|
||||
--weight-threshold $bnf_weight_threshold \
|
||||
data/train $datadir data/lang \
|
||||
$ali_dir $decode_dir $exp_dir/tri6_bnf || exit 1;
|
||||
touch $exp_dir/tri6_bnf/egs/.done
|
||||
fi
|
||||
|
||||
fi
|
||||
|
||||
if [ ! -f $exp_dir/tri6_bnf/.done ]; then
|
||||
if $semisupervised ; then
|
||||
echo "$0: Generate examples using unsupervised data in $exp_dir/tri6_nnet"
|
||||
if [ ! -f $exp_dir/tri6_bnf/egs/.done ]; then
|
||||
local/nnet2/get_egs_semi_supervised.sh \
|
||||
"${dnn_update_egs_opts[@]}" \
|
||||
--transform-dir-sup exp/tri5_ali \
|
||||
--transform-dir-unsup exp/tri5/decode_${dirid} \
|
||||
--weight-threshold $bnf_weight_threshold \
|
||||
data/train $datadir data/lang \
|
||||
$ali_dir $decode_dir $exp_dir/tri6_bnf || exit 1;
|
||||
touch $exp_dir/tri6_bnf/egs/.done
|
||||
fi
|
||||
|
||||
fi
|
||||
|
||||
echo "$0: Train Bottleneck network"
|
||||
steps/nnet2/train_tanh_bottleneck.sh \
|
||||
--stage $bnf_train_stage --num-jobs-nnet $bnf_num_jobs \
|
||||
|
|
@ -86,7 +102,7 @@ fi
|
|||
if [ ! -f $data_bnf_dir/train_bnf/.done ]; then
|
||||
mkdir -p $data_bnf_dir
|
||||
# put the archives in ${param_bnf_dir}/.
|
||||
steps/nnet2/dump_bottleneck_features.sh --nj $train_nj --cmd "$train_cmd" \
|
||||
steps/nnet/make_bn_feats.sh --nj $train_nj --cmd "$train_cmd" \
|
||||
--transform-dir exp/tri5 data/train $data_bnf_dir/train_bnf \
|
||||
$exp_dir/tri6_bnf $param_bnf_dir $exp_dir/dump_bnf
|
||||
touch $data_bnf_dir/train_bnf/.done
|
||||
|
|
@ -95,7 +111,7 @@ fi
|
|||
if [ ! $data_bnf_dir/train/.done -nt $data_bnf_dir/train_bnf/.done ]; then
|
||||
steps/nnet/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
--nj $train_nj --transform-dir exp/tri5_ali $data_bnf_dir/train_sat data/train \
|
||||
exp/tri5_ali $exp_dir/make_fmllr_feats/log $param_bnf_dir/
|
||||
exp/tri5_ali $exp_dir/make_fmllr_feats/log $param_bnf_dir
|
||||
|
||||
steps/append_feats.sh --cmd "$train_cmd" --nj 4 \
|
||||
$data_bnf_dir/train_bnf $data_bnf_dir/train_sat $data_bnf_dir/train \
|
||||
|
|
|
|||
|
|
@ -25,7 +25,7 @@ if [ $babel_type == "full" ] && $semisupervised; then
|
|||
fi
|
||||
|
||||
if $semisupervised ; then
|
||||
unsup_string="_semi_supervised"
|
||||
unsup_string="_semisup"
|
||||
else
|
||||
unsup_string="" #" ": supervised training, _semi_supervised: unsupervised BNF training
|
||||
fi
|
||||
|
|
@ -45,8 +45,9 @@ if [ ! $exp_dir/sgmm7/.done -nt $exp_dir/ubm7/.done ]; then
|
|||
echo ---------------------------------------------------------------------
|
||||
echo "Starting $exp_dir/sgmm7 on" `date`
|
||||
echo ---------------------------------------------------------------------
|
||||
steps/train_sgmm2_group.sh \
|
||||
--cmd "$train_cmd" "${sgmm_group_extra_opts[@]}"\
|
||||
#steps/train_sgmm2_group.sh \
|
||||
steps/train_sgmm2.sh \
|
||||
--cmd "$train_cmd" "${sgmm_train_extra_opts[@]}"\
|
||||
$numLeavesSGMM $bnf_num_gauss_sgmm $data_bnf_dir/train data/lang \
|
||||
$exp_dir/tri6 $exp_dir/ubm7/final.ubm $exp_dir/sgmm7
|
||||
touch $exp_dir/sgmm7/.done
|
||||
|
|
|
|||
|
|
@ -38,6 +38,8 @@ if [ $# -ne 2 ]; then
|
|||
exit 1
|
||||
fi
|
||||
|
||||
set -u
|
||||
|
||||
unsup_datadir=$1
|
||||
unsup_postdir=$2
|
||||
unsup_dirid=`basename $unsup_datadir`
|
||||
|
|
@ -57,12 +59,12 @@ fi
|
|||
if [ ! -f $ali_dir/.done ]; then
|
||||
echo "$0: Aligning supervised training data in exp/tri6_nnet_ali"
|
||||
[ ! -f exp/tri6_nnet/final.mdl ] && echo "exp/tri6_nnet/final.mdl not found!\nRun run-6-nnet.sh first!" && exit 1
|
||||
steps/nnet2/align.sh --cmd "$decode_cmd" \
|
||||
steps/nnet2/align.sh --cmd "$train_cmd" \
|
||||
--use-gpu no --transform-dir exp/tri5_ali --nj $train_nj \
|
||||
data/train data/lang exp/tri6_nnet $ali_dir || exit 1
|
||||
touch $ali_dir/.done
|
||||
fi
|
||||
|
||||
exit 0
|
||||
echo "$0: Using supervised data alignments from $ali_dir"
|
||||
|
||||
###############################################################################
|
||||
|
|
@ -85,11 +87,10 @@ done
|
|||
mkdir -p exp/tri6_nnet_semi_supervised
|
||||
|
||||
if [ ! -f exp/tri6_nnet_semi_supervised/.egs.done ] ; then
|
||||
local/nnet2/get_egs_semi_supervised.sh $spk_vecs_opt \
|
||||
"${egs_gpu_opts[@]}" --io-opts "$egs_io_opts" \
|
||||
local/nnet2/get_egs_semi_supervised.sh --cmd "$train_cmd" \
|
||||
"${dnn_update_egs_opts[@]}" \
|
||||
--transform-dir-sup exp/tri5_ali \
|
||||
--transform-dir-unsup exp/tri5/decode_${dirid} \
|
||||
--weight-threshold $weight_threshold \
|
||||
--transform-dir-unsup exp/tri5/decode_${unsup_dirid} \
|
||||
data/train $unsup_datadir data/lang \
|
||||
$ali_dir $unsup_postdir exp/tri6_nnet_semi_supervised || exit 1;
|
||||
|
||||
|
|
|
|||
|
|
@ -46,13 +46,13 @@ fi
|
|||
if [ ! -f data_bnf/train_bnf/.done ]; then
|
||||
mkdir -p data_bnf
|
||||
# put the archives in plp/.
|
||||
steps/nnet2/dump_bottleneck_features.sh --nj $train_nj --cmd "$train_cmd" \
|
||||
steps/nnet/make_bn_feats.sh --nj $train_nj --cmd "$train_cmd" \
|
||||
--transform-dir exp/tri5 data/train data_bnf/train_bnf exp_bnf/tri6_bnf param_bnf exp_bnf/dump_bnf
|
||||
touch data_bnf/train_bnf/.done
|
||||
fi
|
||||
|
||||
if [ ! data_bnf/train/.done -nt data_bnf/train_bnf/.done ]; then
|
||||
steps/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
steps/nnet/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
--nj $train_nj --transform-dir exp/tri5_ali data_bnf/train_sat data/train \
|
||||
exp/tri5_ali exp_bnf/make_fmllr_feats/log param_bnf/
|
||||
|
||||
|
|
|
|||
|
|
@ -26,8 +26,8 @@ if [ ! exp_bnf/sgmm7/.done -nt exp_bnf/ubm7/.done ]; then
|
|||
echo ---------------------------------------------------------------------
|
||||
echo "Starting exp_bnf/sgmm7 on" `date`
|
||||
echo ---------------------------------------------------------------------
|
||||
steps/train_sgmm2_group.sh \
|
||||
--cmd "$train_cmd" "${sgmm_group_extra_opts[@]}"\
|
||||
steps/train_sgmm2.sh \
|
||||
--cmd "$train_cmd" \
|
||||
$numLeavesSGMM $bnf_num_gauss_sgmm data_bnf/train data/lang \
|
||||
exp_bnf/tri6 exp_bnf/ubm7/final.ubm exp_bnf/sgmm7
|
||||
touch exp_bnf/sgmm7/.done
|
||||
|
|
|
|||
|
|
@ -27,7 +27,7 @@ if [ ! exp_bnf/tri6_ali_50/.done -nt exp_bnf/tri6/.done ]; then
|
|||
echo "Aligning fMLLR system with 50 jobs"
|
||||
echo ---------------------------------------------------------------------
|
||||
steps/align_fmllr.sh \
|
||||
--boost-silence $boost_sil --nj 50 --cmd "$train_cmd" \
|
||||
--boost-silence $boost_sil --nj $train_nj --cmd "$train_cmd" \
|
||||
data_bnf/train_app data/lang exp_bnf/tri6 exp_bnf/tri6_ali_50
|
||||
touch exp_bnf/tri6_ali_50/.done
|
||||
fi
|
||||
|
|
|
|||
|
|
@ -33,12 +33,12 @@ if [ $babel_type == "full" ] && $semisupervised; then
|
|||
fi
|
||||
|
||||
if $semisupervised ; then
|
||||
unsup_string="_semi_supervised"
|
||||
unsup_string="_semisup"
|
||||
else
|
||||
unsup_string="" #" ": supervised training, _semi_supervised: unsupervised BNF training
|
||||
fi
|
||||
|
||||
if ! echo {dev10h,dev2h,eval,unsup}{,.uem,.seg} | grep -w "$type" >/dev/null; then
|
||||
if ! echo {dev10h,dev2h,eval,unsup,shadow}{,.uem,.seg} | grep -w "$type" >/dev/null; then
|
||||
# note: echo dev10.uem | grep -w dev10h will produce a match, but this
|
||||
# doesn't matter because dev10h is also a valid value.
|
||||
echo "Invalid variable type=${type}, valid values are " {dev10h,dev2h,eval,unsup}{,.uem,.seg}
|
||||
|
|
@ -61,7 +61,7 @@ my_nj=`cat exp/tri5/decode_${dirid}/num_jobs` || exit 1;
|
|||
if [ ! $data_bnf_dir/${dirid}_bnf/.done -nt exp/tri5/decode_${dirid}/.done ] || \
|
||||
[ ! $data_bnf_dir/${dirid}_bnf/.done -nt $exp_dir/tri6_bnf/.done ]; then
|
||||
# put the archives in $param_bnf_dir/.
|
||||
steps/nnet2/dump_bottleneck_features.sh --nj $my_nj --cmd "$train_cmd" \
|
||||
steps/nnet/make_bn_feats.sh --nj $my_nj --cmd "$train_cmd" \
|
||||
--transform-dir exp/tri5/decode_${dirid} data/${dirid} $data_bnf_dir/${dirid}_bnf $exp_dir/tri6_bnf $param_bnf_dir $exp_dir/dump_bnf
|
||||
touch $data_bnf_dir/${dirid}_bnf/.done
|
||||
fi
|
||||
|
|
@ -77,10 +77,14 @@ if [ ! $data_bnf_dir/${dirid}/.done -nt $data_bnf_dir/${dirid}_bnf/.done ]; then
|
|||
steps/compute_cmvn_stats.sh --fake $data_bnf_dir/${dirid} $exp_dir/make_fmllr_feats $param_bnf_dir
|
||||
rm -r $data_bnf_dir/${dirid}_sat
|
||||
if ! $skip_kws ; then
|
||||
cp -r data/${dirid}/kws* $data_bnf_dir/${dirid}/
|
||||
cp -r data/${dirid}/*kws* $data_bnf_dir/${dirid}/ || true
|
||||
fi
|
||||
touch $data_bnf_dir/${dirid}/.done
|
||||
fi
|
||||
if ! $skip_kws ; then
|
||||
cp -r data/${dirid}/*kws* $data_bnf_dir/${dirid}/ || true
|
||||
fi
|
||||
|
||||
|
||||
if $data_only ; then
|
||||
echo "Exiting, as data-only was requested... "
|
||||
|
|
@ -179,6 +183,8 @@ for iter in 1 2 3 4; do
|
|||
${datadir} data/lang $decode
|
||||
done
|
||||
|
||||
exit 0
|
||||
|
||||
if [ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ] || \
|
||||
[ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt exp_bnf/tri7_nnet/.done ]; then
|
||||
|
||||
|
|
|
|||
|
|
@ -22,7 +22,7 @@ if [ $# -ne 0 ]; then
|
|||
exit 1
|
||||
fi
|
||||
|
||||
if ! echo {dev10h,dev2h,eval,unsup}{,.uem,.seg} | grep -w "$type" >/dev/null; then
|
||||
if ! echo {shadow,dev10h,dev2h,eval,unsup}{,.uem,.seg,.pem} | grep -w "$type" >/dev/null; then
|
||||
# note: echo dev10.uem | grep -w dev10h will produce a match, but this
|
||||
# doesn't matter because dev10h is also a valid value.
|
||||
echo "Invalid variable type=${type}, valid values are " {dev10h,dev2h,eval,unsup}{,.uem,.seg}
|
||||
|
|
@ -38,17 +38,19 @@ datadir=data_bnf/${dirid}
|
|||
# Set my_nj; typically 64.
|
||||
my_nj=`cat exp/tri5/decode_${dirid}/num_jobs` || exit 1;
|
||||
|
||||
|
||||
test -d param_bnf || mkdir -p param_bnf
|
||||
mkdir -p param_bnf
|
||||
if [ ! data_bnf/${dirid}_bnf/.done -nt exp/tri5/decode_${dirid}/.done ] || \
|
||||
[ ! data_bnf/${dirid}_bnf/.done -nt exp_bnf/tri6_bnf/.done ]; then
|
||||
# put the archives in param_bnf/.
|
||||
steps/nnet2/dump_bottleneck_features.sh --nj $my_nj --cmd "$train_cmd" \
|
||||
|
||||
local/nnet/make_bn_feats.sh --nj $my_nj --cmd "$train_cmd" \
|
||||
--transform-dir exp/tri5/decode_${dirid} data/${dirid} data_bnf/${dirid}_bnf exp_bnf/tri6_bnf param_bnf exp_bnf/dump_bnf
|
||||
touch data_bnf/${dirid}_bnf/.done
|
||||
fi
|
||||
|
||||
if [ ! data_bnf/${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ]; then
|
||||
steps/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
steps/nnet/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
--nj 16 --transform-dir exp/tri5/decode_${dirid} data_bnf/${dirid}_sat data/${dirid} \
|
||||
exp/tri5_ali exp_bnf/make_fmllr_feats/log param_bnf/
|
||||
|
||||
|
|
@ -62,6 +64,9 @@ if [ ! data_bnf/${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ]; then
|
|||
fi
|
||||
touch data_bnf/${dirid}/.done
|
||||
fi
|
||||
if ! $skip_kws ; then
|
||||
cp -r data/${dirid}/*kws* data_bnf/${dirid}/ || true
|
||||
fi
|
||||
|
||||
|
||||
if $data_only ; then
|
||||
|
|
@ -161,6 +166,9 @@ for iter in 1 2 3 4; do
|
|||
${datadir} data/lang $decode
|
||||
done
|
||||
|
||||
echo "$0: Everything looking good...."
|
||||
exit 0
|
||||
|
||||
if [ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ] || \
|
||||
[ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt exp_bnf/tri7_nnet/.done ]; then
|
||||
|
||||
|
|
|
|||
|
|
@ -40,13 +40,13 @@ my_nj=`cat exp/tri5/decode_${dirid}/num_jobs` || exit 1;
|
|||
if [ ! data_bnf/${dirid}_bnf/.done -nt exp/tri5/decode_${dirid}/.done ] || \
|
||||
[ ! data_bnf/${dirid}_bnf/.done -nt exp_bnf/tri6_bnf/.done ]; then
|
||||
# put the archives in plp/.
|
||||
steps/nnet2/dump_bottleneck_features.sh --nj $my_nj --cmd "$train_cmd" \
|
||||
steps/nnet/make_bn_feats.sh --nj $my_nj --cmd "$train_cmd" \
|
||||
--transform-dir exp/tri5/decode_${dirid} data/${dirid} data_bnf/${dirid}_bnf exp_bnf/tri6_bnf param_bnf exp_bnf/dump_bnf
|
||||
touch data_bnf/${dirid}_bnf/.done
|
||||
fi
|
||||
|
||||
if [ ! data_bnf/${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ]; then
|
||||
steps/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
steps/nnet/make_fmllr_feats.sh --cmd "$train_cmd -tc 10" \
|
||||
--nj $train_nj --transform-dir exp/tri5/decode_${dirid} data_bnf/${dirid}_sat data/${dirid} \
|
||||
exp/tri5_ali exp_bnf/make_fmllr_feats/log param_bnf
|
||||
|
||||
|
|
@ -59,6 +59,7 @@ if [ ! data_bnf/${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ]; then
|
|||
touch data_bnf/${dirid}/.done
|
||||
fi
|
||||
|
||||
decode=exp_bnf/tri7_nnet/decode_${dirid}
|
||||
if [ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ] || \
|
||||
[ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt exp_bnf/tri7_nnet/.done ]; then
|
||||
|
||||
|
|
@ -70,7 +71,6 @@ if [ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ]
|
|||
utils/mkgraph.sh \
|
||||
data/lang exp_bnf/tri6 exp_bnf/tri6/graph |tee exp_bnf/tri6/mkgraph.log
|
||||
|
||||
decode=exp_bnf/tri7_nnet/decode_${dirid}
|
||||
if [ ! -f $decode/.done ]; then
|
||||
mkdir -p $decode
|
||||
steps/nnet2/decode.sh \
|
||||
|
|
@ -84,11 +84,11 @@ if [ ! exp_bnf/tri7_nnet/decode_${dirid}/.done -nt data_bnf/${dirid}_bnf/.done ]
|
|||
touch $decode/.done
|
||||
fi
|
||||
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt --wip $wip \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_bnf_extra_opts[@]}" \
|
||||
${datadir} data/lang $decode
|
||||
fi
|
||||
local/run_kws_stt_task.sh --cer $cer --max-states $max_states \
|
||||
--cmd "$decode_cmd" --skip-kws $skip_kws --skip-stt $skip_stt --wip $wip \
|
||||
"${shadow_set_extra_opts[@]}" "${lmwt_bnf_extra_opts[@]}" \
|
||||
${datadir} data/lang $decode
|
||||
|
||||
echo "$0: Everything looking good...."
|
||||
exit 0
|
||||
|
|
|
|||
|
|
@ -0,0 +1,110 @@
|
|||
#!/bin/bash
|
||||
|
||||
export NJ=`(. ./lang.conf > /dev/null; echo $train_nj )`
|
||||
export TYPE=`(. ./lang.conf > /dev/null; echo $babel_type )`
|
||||
|
||||
echo $NJ
|
||||
echo $TYPE
|
||||
|
||||
if [ "$TYPE" == "limited" ]; then
|
||||
T_SHORT="6:0:0"
|
||||
T_MEDIUM="12:0:0"
|
||||
T_LONG="24:0:0"
|
||||
T_EXTREME="48:0:0"
|
||||
BNF_NJ=$((16 * 4))
|
||||
DNN_NJ=$((16 * 4))
|
||||
elif [ "$TYPE" == "full" ]; then
|
||||
T_SHORT="6:0:0"
|
||||
T_MEDIUM="24:0:0"
|
||||
T_LONG="48:0:0"
|
||||
T_EXTREME="48:0:0"
|
||||
BNF_NJ=$((16 * 8))
|
||||
DNN_NJ=$((16 * 8))
|
||||
else
|
||||
echo "Unknown BABEL type! Exiting..."
|
||||
exit 1
|
||||
fi
|
||||
|
||||
|
||||
export SBATCH_JOBID
|
||||
|
||||
|
||||
function sbatch {
|
||||
#echo "sbatch " "${@}"
|
||||
output_name=""
|
||||
for param in "${@}"; do
|
||||
if [[ $param =~ ^\./.*sh ]]; then
|
||||
output_name=`basename $param`
|
||||
fi
|
||||
done
|
||||
if [ ! -z $output_name ]; then
|
||||
output_name="-o ${output_name}.%j"
|
||||
fi
|
||||
#echo "OUTPUT: $output_name"
|
||||
echo /usr/bin/sbatch --mail-type ALL --mail-user 'jtrmal@gmail.com' $output_name "${@}"
|
||||
jobid=$(/usr/bin/sbatch --mail-type ALL --mail-user 'jtrmal@gmail.com' $output_name "${@}" | tee /dev/stderr | grep "Submitted batch job" | awk '{print $4}' )
|
||||
SBATCH_JOBID=$jobid
|
||||
}
|
||||
|
||||
sbatch -p normal -n $NJ -t $T_SHORT ./run-1-main.sh --tri5-only true
|
||||
TRI5_ID=$SBATCH_JOBID
|
||||
|
||||
sbatch -p normal -n $NJ -t $T_LONG --dependency=afterok:$TRI5_ID ./run-1-main.sh
|
||||
PLP_ID=$SBATCH_JOBID
|
||||
|
||||
sbatch -p normal -n $NJ -t $T_SHORT --dependency=afterok:$TRI5_ID ./run-2-segmentation.sh
|
||||
SEG_ID=$SBATCH_JOBID
|
||||
|
||||
if [ "$TYPE" == "limited" ]; then
|
||||
sbatch -p gpu -n $DNN_NJ -t $T_MEDIUM --dependency=afterok:$TRI5_ID ./run-2a-nnet-ensemble-gpu.sh --dir exp/tri6_nnet/
|
||||
else
|
||||
sbatch -p gpu -n $DNN_NJ -t $T_MEDIUM --dependency=afterok:$TRI5_ID ./run-2a-nnet-gpu.sh
|
||||
DNN_ID=$SBATCH_JOBID
|
||||
sbatch -p gpu -n $DNN_NJ -t $T_MEDIUM --dependency=afterok:$DNN_ID ./run-2a-nnet-mpe.sh
|
||||
fi
|
||||
DNN_ID=$SBATCH_JOBID
|
||||
|
||||
sbatch -p gpu -n $BNF_NJ -t 24:0:0 --dependency=afterok:$TRI5_ID ./run-8a-kaldi-bnf.sh
|
||||
BNF_ID=$SBATCH_JOBID
|
||||
|
||||
sbatch -p normal -n $NJ -t $T_LONG --dependency=afterok:$BNF_ID ./run-8b-kaldi-bnf-sgmm.sh
|
||||
BNF_SGMM_ID=$SBATCH_JOBID
|
||||
|
||||
|
||||
#Decode DNNs and PLP systems
|
||||
sbatch -p normal -n 128 -t $T_MEDIUM --dependency=afterok:$DNN_ID:$PLP_ID ./run-5-anydecode.sh --fast-path true --skip-kws true --type dev10h
|
||||
DECODE_DNN_PLP_ID=$SBATCH_JOBID
|
||||
sbatch -p normal -n 16 -t $T_MEDIUM --dependency=afterok:$DECODE_DNN_PLP_ID ./run-5-anydecode.sh --fast-path true
|
||||
|
||||
#Decode BNF systems
|
||||
sbatch -p normal -n 128 -t $T_LONG --dependency=afterok:$BNF_SGMM_ID:$DECODE_DNN_PLP_ID ./run-8d-test-kaldi-bnf-sgmm.sh --skip-kws true --type dev10h
|
||||
DECODE_BNF_SGMM_ID=$SBATCH_JOBID
|
||||
sbatch -p normal -n 16 -t $T_MEDIUM --dependency=afterok:$DECODE_BNF_SGMM_ID ./run-8d-test-kaldi-bnf-sgmm.sh
|
||||
|
||||
exit 0
|
||||
|
||||
#For the discriminative training, we have to actually decode the unsup.seg
|
||||
#The unsup.seg needs segmentation to be done, i.e. it depends on the individual systems and on the segmentation
|
||||
if [ "$TYPE" == "limited" ]; then
|
||||
#First, setup data
|
||||
sbatch -p normal -n $NJ -t $T_LONG --dependency=afterok:$SEG_ID ./run-4-anydecode.sh --fast-path true --skip-scoring true --skip-kws true --dir unsup.seg --data-only true
|
||||
UNSUP_DATA_PREPARED=$SBATCH_JOBID
|
||||
sbatch -p normal -n 256 -t $T_LONG --dependency=afterok:$UNSUP_DATA_PREPARED:$DNN_ID:$PLP_ID ./run-4-anydecode.sh --fast-path true --skip-scoring true --skip-kws true --dir unsup.seg
|
||||
SEMI_PARTA_ID=$SBATCH_JOBID
|
||||
sbatch -p normal -n 256 -t $T_LONG --dependency=afterok:$UNSUP_DATA_PREPARED:$BNF_SGMM_ID:$DECODE_DNN_PLP_ID ./run-8d-test-kaldi-bnf-sgmm.sh --skip-kws true --skip-kws true --type unsup.seg
|
||||
SEMI_PARTB_ID=$SBATCH_JOBID
|
||||
fi
|
||||
|
||||
#
|
||||
#
|
||||
#We do not run BNF on the top of DNN by default (low performance)
|
||||
#sbatch -p gpu -n $BNF_NJ -t 24:0:0 --dependency=afterok:$BNF_ID ./run-8c-kaldi-bnf-dnn.sh
|
||||
#BNF_DNN_ID=$SBATCH_JOBID
|
||||
#The decoding depends on the BNF-SGMM in that sense that it expects the data directories to be prepared.
|
||||
#It can create the directories on its own, but do not run those two scripts in parallel -- because of no locking
|
||||
#this will result in crash as the scripts will overwrite each others's files
|
||||
#sbatch -p normal -n 128 -t $T_LONG --dependency=afterok:$BNF_DNN_ID:$DECODE_DNN_PLP_ID:$DECODE_BNF_SGMM_ID ./run-8e-test-kaldi-bnf-dnn.sh --skip-kws true
|
||||
#DECODE_BNF_DNN_ID=$SBATCH_JOBID
|
||||
#sbatch -p normal -n 16 -t $T_MEDIUM --dependency=afterok:$DECODE_BNF_DNN_ID ./run-8e-test-kaldi-bnf-dnn.sh
|
||||
|
||||
|
||||
|
|
@ -1,11 +1,10 @@
|
|||
for x in exp/*/decode_dev; do grep WER $x/wer_* | utils/best_wer.sh; done
|
||||
|
||||
%WER 50.00 [ 19571 / 39141, 1893 ins, 4738 del, 12940 sub ] exp/tri1/decode_dev/wer_12
|
||||
%WER 49.52 [ 19384 / 39141, 1774 ins, 5035 del, 12575 sub ] exp/tri2/decode_dev/wer_13
|
||||
%WER 42.57 [ 16664 / 39141, 1908 ins, 4080 del, 10676 sub ] exp/tri3a/decode_dev/wer_12
|
||||
%WER 35.67 [ 13963 / 39141, 1810 ins, 3347 del, 8806 sub ] exp/tri4a/decode_dev/wer_13
|
||||
%WER 32.09 [ 12560 / 39141, 1680 ins, 3131 del, 7749 sub ] exp/tri5a/decode_dev/wer_14
|
||||
|
||||
%WER 49.72 [ 19461 / 39141, 1999 ins, 4578 del, 12884 sub ] exp/tri1/decode_dev/wer_12
|
||||
%WER 49.00 [ 19181 / 39141, 1812 ins, 4848 del, 12521 sub ] exp/tri2/decode_dev/wer_13
|
||||
%WER 41.86 [ 16384 / 39141, 1735 ins, 4152 del, 10497 sub ] exp/tri3a/decode_dev/wer_13
|
||||
%WER 34.73 [ 13593 / 39141, 1719 ins, 3365 del, 8509 sub ] exp/tri4a/decode_dev/wer_14
|
||||
%WER 31.07 [ 12163 / 39141, 1869 ins, 2705 del, 7589 sub ] exp/tri5a/decode_dev/wer_13
|
||||
%WER 31.13 [ 12184 / 39141, 1939 ins, 2584 del, 7661 sub ] exp/tri5a_0.1/decode_dev/wer_12
|
||||
|
||||
%WER 23.66 [ 9259 / 39141, 1495 ins, 2432 del, 5332 sub ] exp/nnet6c4_gpu/decode_dev/wer_11
|
||||
|
||||
|
|
|
|||
|
|
@ -5,17 +5,30 @@
|
|||
|
||||
stage=0
|
||||
|
||||
calldata=
|
||||
while test $# -gt 0
|
||||
do
|
||||
case "$1" in
|
||||
--calldata) calldata=1
|
||||
;;
|
||||
*) break;
|
||||
;;
|
||||
esac
|
||||
shift
|
||||
done
|
||||
|
||||
. utils/parse_options.sh
|
||||
|
||||
if [ $# -eq 0 ]; then
|
||||
echo "$0 <fisher-dir-1> [<fisher-dir-2> ...]"
|
||||
echo "$0 [--calldata] <fisher-dir-1> [<fisher-dir-2> ...]"
|
||||
echo " e.g.: $0 /export/corpora3/LDC/LDC2004T19 /export/corpora3/LDC/LDC2005T19\\"
|
||||
echo " /export/corpora3/LDC/LDC2004S13 /export/corpora3/LDC/LDC2005S13"
|
||||
echo " (We also support a single directory that has the contents of all of them)"
|
||||
echo " If specified, --calldata will be used to map Kaldi speaker ID to real"
|
||||
echo " speaker PIN released with the Fisher corpus."
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
|
||||
# Check that the arguments are all absolute pathnames.
|
||||
|
||||
for dir in $*; do
|
||||
|
|
@ -178,5 +191,17 @@ if [ $stage -le 4 ]; then
|
|||
fi
|
||||
fi
|
||||
|
||||
if [ ! -z "$calldata" ]; then # fix speaker IDs
|
||||
cat $links/fe_03_p{1,2}_tran/doc/*calldata.tbl > $tmpdir/combined-calldata.tbl
|
||||
local/fisher_fix_speakerid.pl $tmpdir/combined-calldata.tbl data/train_all
|
||||
utils/utt2spk_to_spk2utt.pl data/train_all/utt2spk.new > data/train_all/spk2utt.new
|
||||
# patch files
|
||||
for f in spk2utt utt2spk text segments spk2gender; do
|
||||
cp data/train_all/$f data/train_all/$f.old || exit 1;
|
||||
cp data/train_all/$f.new data/train_all/$f || exit 1;
|
||||
done
|
||||
rm $tmpdir/combined-calldata.tbl
|
||||
fi
|
||||
|
||||
echo "Data preparation succeeded"
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,114 @@
|
|||
#!/usr/bin/perl -w
|
||||
|
||||
# Author: Peng Qi (pengqi@cs.stanford.edu)
|
||||
# This script maps Switchboard speaker IDs to the true physical speakers
|
||||
# and fixes the utterances IDs accordingly. Expected to be run one level of
|
||||
# directory above.
|
||||
|
||||
sub trim {
|
||||
(my $s = $_[0]) =~ s/^\s+|\s+$//g;
|
||||
return $s;
|
||||
}
|
||||
|
||||
if ($#ARGV != 1) {
|
||||
print "Usage: swbd1_fix_speakerid.pl <fisher-calldata-tbl-file> <data-dir>\n";
|
||||
print "E.g.: swbd1_fix_speakerid.pl data/local/train/combined-calldata.tbl data/train_all\n";
|
||||
}
|
||||
|
||||
$tab_file = $ARGV[0];
|
||||
$dir = $ARGV[1];
|
||||
|
||||
%conv_to_spk = ();
|
||||
|
||||
open(my $conv_tab, '<', $tab_file) or die "Could not open '$tab_file' $!\n";
|
||||
|
||||
while (my $line = <$conv_tab>) {
|
||||
chomp $line;
|
||||
|
||||
my @fields = split "," , $line;
|
||||
#$fields[0] = trim($fields[0]);
|
||||
$fields[5] = trim($fields[5]);
|
||||
$fields[10] = trim($fields[10]);
|
||||
$conv_to_spk{'fe_03_' . $fields[0] . '-A'} = $fields[5];
|
||||
$conv_to_spk{'fe_03_' . $fields[0] . '-B'} = $fields[10];
|
||||
}
|
||||
|
||||
close($conv_tab);
|
||||
|
||||
# fix utt2spk
|
||||
|
||||
%missingconv = ();
|
||||
|
||||
open(my $utt2spk, '<', $dir . '/utt2spk') or die "Could not open '$dir/utt2spk' $!\n";
|
||||
open(my $utt2spk_new, '>', $dir . '/utt2spk.new');
|
||||
|
||||
while (my $line = <$utt2spk>) {
|
||||
chomp $line;
|
||||
|
||||
my @fields = split " " , $line;
|
||||
my $convid = substr $fields[0], 0, 13;
|
||||
|
||||
if (exists $conv_to_spk{ $convid }) {
|
||||
my $spkid = $conv_to_spk{ $convid };
|
||||
$spkid = "fe_03_" . $spkid;
|
||||
my $newuttid = $spkid . '-' . (substr $fields[0], 6);
|
||||
|
||||
print $utt2spk_new "$newuttid $spkid\n";
|
||||
} else {
|
||||
my $convid = substr $convid, 6, 5;
|
||||
$missingconv{$convid} = 1;
|
||||
|
||||
print $utt2spk_new $fields[0]." ".$fields[1]."\n";
|
||||
}
|
||||
}
|
||||
|
||||
close($utt2spk);
|
||||
close($utt2spk_new);
|
||||
|
||||
foreach my $conv (keys %missingconv) {
|
||||
print "Warning: Conversation ID '$conv' not found in conv.tab, retaining old speaker IDs\n"
|
||||
}
|
||||
|
||||
# fix spk2gender
|
||||
|
||||
if (open(my $spk2gender, '<', $dir . '/spk2gender')) {
|
||||
open(my $spk2gender_new, '>', $dir . '/spk2gender.new');
|
||||
|
||||
while (my $line = <$spk2gender>) {
|
||||
chomp $line;
|
||||
|
||||
my @fields = split " ", $line;
|
||||
my $convid = $fields[0];
|
||||
|
||||
if (exists $conv_to_spk{ $convid }) {
|
||||
my $spkid = $conv_to_spk{ $convid };
|
||||
$spkid = "fe_03_" . $spkid;
|
||||
|
||||
print $spk2gender_new $spkid." ".$fields[1]."\n";
|
||||
} else {
|
||||
print $spk2gender_new $fields[0]." ".$fields[1]."\n";
|
||||
}
|
||||
}
|
||||
|
||||
close($spk2gender);
|
||||
close($spk2gender_new);
|
||||
}
|
||||
|
||||
# fix segments and text
|
||||
|
||||
foreach my $file ('segments','text') {
|
||||
open(my $oldfile, '<', "$dir/$file") or die "Could not open '$dir/$file' $!\n";
|
||||
open(my $newfile, '>', "$dir/$file.new");
|
||||
|
||||
while (my $line = <$oldfile>) {
|
||||
chomp $line;
|
||||
|
||||
my $convid = substr $line, 0, 13;
|
||||
if (exists $conv_to_spk{$convid}) {
|
||||
my $spkid = $conv_to_spk{$convid};
|
||||
print $newfile "fe_03_$spkid-" . (substr $line, 6) . "\n";
|
||||
} else {
|
||||
print $newfile "$line\n";
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -0,0 +1,36 @@
|
|||
#!/bin/bash
|
||||
|
||||
|
||||
# This script shows how you can do data-cleaning, and exclude data that has a
|
||||
# higher likelihood of being wrongly transcribed. see the RESULTS file; this
|
||||
# made essentially no difference in our case-- indicating, perhaps, that Fisher
|
||||
# transcripts are already clean enough.
|
||||
|
||||
|
||||
. cmd.sh
|
||||
. path.sh
|
||||
set -e
|
||||
|
||||
|
||||
steps/cleanup/find_bad_utts.sh --nj 200 --cmd "$train_cmd" data/train data/lang \
|
||||
exp/tri5a exp/tri5a_cleanup
|
||||
|
||||
# with threshold of 0.05 we keep 1.1 million out of 1.6 million utterances, and
|
||||
# around 8.7 million out of 18.1 million words
|
||||
# with threshold of 0.1 we keep 1.3 out of 1.6 million utterances, and around
|
||||
# 13.2 million out of 18.1 million words.
|
||||
thresh=0.1
|
||||
cat exp/tri5a_cleanup/all_info.txt | awk -v threshold=$thresh '{ errs=$2;ref=$3; if (errs <= threshold*ref) { print $1; } }' > uttlist
|
||||
utils/subset_data_dir.sh --utt-list uttlist data/train data/train.thresh$thresh
|
||||
|
||||
steps/align_fmllr.sh --nj 30 --cmd "$train_cmd" \
|
||||
data/train.thresh$thresh data/lang exp/tri4a exp/tri4a_ali_$thresh
|
||||
|
||||
steps/train_sat.sh --cmd "$train_cmd" \
|
||||
10000 300000 data/train data/lang exp/tri4a_ali_$thresh exp/tri5a_$thresh || exit 1;
|
||||
|
||||
|
||||
utils/mkgraph.sh data/lang_test exp/tri5a_$thresh exp/tri5a_$thresh/graph
|
||||
steps/decode_fmllr.sh --nj 25 --cmd "$decode_cmd" --config conf/decode.config \
|
||||
exp/tri5a_$thresh/graph data/dev exp/tri5a_$thresh/decode_dev
|
||||
|
||||
|
|
@ -10,6 +10,12 @@ set -e
|
|||
# the next command produces the data in local/train_all
|
||||
local/fisher_data_prep.sh /export/corpora3/LDC/LDC2004T19 /export/corpora3/LDC/LDC2005T19 \
|
||||
/export/corpora3/LDC/LDC2004S13 /export/corpora3/LDC/LDC2005S13
|
||||
# You could also try specifying the --calldata argument to this command as below.
|
||||
# If specified, the script will use actual speaker personal identification
|
||||
# numbers released with the dataset, i.e. real speaker IDs. Note: --calldata has
|
||||
# to be the first argument of this script.
|
||||
# local/fisher_data_prep.sh --calldata /export/corpora3/LDC/LDC2004T19 /export/corpora3/LDC/LDC2005T19 \
|
||||
# /export/corpora3/LDC/LDC2004S13 /export/corpora3/LDC/LDC2005S13
|
||||
|
||||
# at BUT:
|
||||
# local/fisher_data_prep.sh /mnt/matylda6/jhu09/qpovey/FISHER/LDC2005T19 /mnt/matylda2/data/FISHER/
|
||||
|
|
@ -156,12 +162,12 @@ steps/train_sat.sh --cmd "$train_cmd" \
|
|||
exp/tri5a/graph data/dev exp/tri5a/decode_dev
|
||||
)&
|
||||
|
||||
#
|
||||
# steps/cleanup/find_bad_utts.sh --nj 200 --cmd "$train_cmd" data/train data/lang \
|
||||
# exp/tri5a exp/tri5a_cleanup
|
||||
|
||||
|
||||
# The step below won't run by default; it demonstrates a data-cleaning method.
|
||||
# It doesn't seem to help in this setup; maybe the data was clean enough already.
|
||||
# local/run_data_cleaning.sh
|
||||
|
||||
# local/run_for_spkid.sh
|
||||
|
||||
# we don't have to results for the step below yet.
|
||||
# local/run_nnet2.sh
|
||||
|
|
|
|||
|
|
@ -168,6 +168,8 @@ exit 0
|
|||
# last time I created this.
|
||||
# Per-frame cross-entropy training
|
||||
%WER 1.66 [ 208 / 12533, 27 ins, 49 del, 132 sub ] exp/dnn4b_pretrain-dbn_dnn/decode/wer_3
|
||||
%WER 7.80 [ 978 / 12533, 83 ins, 151 del, 744 sub ] exp/dnn4b_pretrain-dbn_dnn/decode_ug/wer_6
|
||||
|
||||
# Sequence-based sMBR training
|
||||
%WER 1.64 [ 206 / 12533, 24 ins, 49 del, 133 sub ] exp/dnn4b_pretrain-dbn_dnn_smbr/decode_it1/wer_4
|
||||
%WER 1.62 [ 203 / 12533, 25 ins, 46 del, 132 sub ] exp/dnn4b_pretrain-dbn_dnn_smbr/decode_it2/wer_4
|
||||
|
|
|
|||
|
|
@ -7,8 +7,14 @@
|
|||
|
||||
train_cmd="queue.pl -l arch=*64"
|
||||
decode_cmd="queue.pl -l arch=*64"
|
||||
|
||||
# cuda_cmd is used for nnet1 scripts e.g. local/run_dnn.sh, but
|
||||
# in the nnet2 scripts e.g. local/run_nnet2.sh, this is not
|
||||
# used and we append options to train_cmd.
|
||||
cuda_cmd="queue.pl -l arch=*64 -l gpu=1"
|
||||
|
||||
#train_cmd="run.pl"
|
||||
# Do training locally. Note: for jobs on smallish subsets,
|
||||
# with run.pl we do training locally. Note: for jobs on smallish subsets,
|
||||
# it's way faster to run on a single machine with a handful of CPUs, as
|
||||
# you avoid the latency of starting GridEngine jobs.
|
||||
|
||||
|
|
|
|||
|
|
@ -2,5 +2,3 @@
|
|||
first_beam=16.0
|
||||
beam=20.0
|
||||
lattice_beam=10.0
|
||||
min_lmwt=2
|
||||
max_lmwt=10
|
||||
|
|
|
|||
|
|
@ -41,7 +41,7 @@ if [ ! -f $dir/final.mdl ]; then
|
|||
--minibatch-size "$minibatch_size" \
|
||||
--parallel-opts "$parallel_opts" \
|
||||
--num-jobs-nnet 4 \
|
||||
--num-epochs-extra 10 --add-layers-period 1 \
|
||||
--num-epochs 8 --num-epochs-extra 5 --add-layers-period 1 \
|
||||
--num-hidden-layers 2 \
|
||||
--mix-up 4000 \
|
||||
--initial-learning-rate 0.02 --final-learning-rate 0.004 \
|
||||
|
|
|
|||
|
|
@ -66,6 +66,8 @@ if [ $stage -le 2 ]; then
|
|||
# Decode (reuse HCLG graph)
|
||||
steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt 0.1 \
|
||||
$gmmdir/graph $data_fmllr/test $dir/decode || exit 1;
|
||||
steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config --acwt 0.1 \
|
||||
$gmmdir/graph_ug $data_fmllr/test $dir/decode_ug || exit 1;
|
||||
fi
|
||||
|
||||
|
||||
|
|
@ -92,6 +94,9 @@ if [ $stage -le 4 ]; then
|
|||
steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config \
|
||||
--nnet $dir/${ITER}.nnet --acwt $acwt \
|
||||
$gmmdir/graph $data_fmllr/test $dir/decode_it${ITER} || exit 1
|
||||
steps/nnet/decode.sh --nj 20 --cmd "$decode_cmd" --config conf/decode_dnn.config \
|
||||
--nnet $dir/${ITER}.nnet --acwt $acwt \
|
||||
$gmmdir/graph_ug $data_fmllr/test $dir/decode_ug_it${ITER} || exit 1
|
||||
done
|
||||
fi
|
||||
|
||||
|
|
@ -100,3 +105,16 @@ exit 0
|
|||
|
||||
# Getting results [see RESULTS file]
|
||||
# for x in exp/*/decode*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done
|
||||
|
||||
# Showing how model conversion to nnet2 works; note, we use the expanded variable
|
||||
# names here so be careful in case the script changes.
|
||||
# steps/nnet2/convert_nnet1_to_nnet2.sh exp/dnn4b_pretrain-dbn_dnn exp/dnn4b_nnet2
|
||||
# cp exp/tri3b/splice_opts exp/tri3b/cmvn_opts exp/tri3b/final.mat exp/dnn4b_nnet2/
|
||||
#
|
||||
# steps/nnet2/decode.sh --nj 10 --cmd "$decode_cmd" --transform-dir exp/tri3b/decode \
|
||||
# --config conf/decode.config exp/tri3b/graph data/test exp/dnn4b_nnet2/decode
|
||||
|
||||
# decoding results are essentially the same (any small difference is probably because
|
||||
# decode.config != decode_dnn.config).
|
||||
# %WER 1.58 [ 198 / 12533, 22 ins, 45 del, 131 sub ] exp/dnn4b_nnet2/decode/wer_3
|
||||
# %WER 1.59 [ 199 / 12533, 23 ins, 45 del, 131 sub ] exp/dnn4b_pretrain-dbn_dnn/decode/wer_3
|
||||
|
|
|
|||
|
|
@ -10,11 +10,18 @@
|
|||
## you unpacked this. We are just doing a "find" command to locate
|
||||
## the .sph files.
|
||||
|
||||
## The second input is optional, which should point to a directory containing
|
||||
## Switchboard transcriptions/documentations (specifically, the conv.tab file).
|
||||
## If specified, the script will try to use the actual speaker PINs provided
|
||||
## with the corpus instead of the conversation side ID (Kaldi default). We
|
||||
## will be using "find" to locate this file so we don't make any assumptions
|
||||
## on the directory structure. (Peng Qi, Aug 2014)
|
||||
|
||||
. path.sh
|
||||
|
||||
#check existing directories
|
||||
if [ $# != 1 ]; then
|
||||
echo "Usage: swbd1_data_prep_edin.sh /path/to/SWBD"
|
||||
if [ $# != 1 -a $# != 2 ]; then
|
||||
echo "Usage: swbd1_data_prep_edin.sh /path/to/SWBD [/path/to/SWBD_DOC]"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
|
|
@ -144,6 +151,17 @@ for f in spk2utt utt2spk wav.scp text segments reco2file_and_channel; do
|
|||
cp data/local/train/$f data/train/$f || exit 1;
|
||||
done
|
||||
|
||||
if [ $# == 2 ]; then # fix speaker IDs
|
||||
find $2 -name conv.tab > $dir/conv.tab
|
||||
local/swbd1_fix_speakerid.pl `cat $dir/conv.tab` data/train
|
||||
utils/utt2spk_to_spk2utt.pl data/train/utt2spk.new > data/train/spk2utt.new
|
||||
# patch files
|
||||
for f in spk2utt utt2spk text segments; do
|
||||
cp data/train/$f data/train/$f.old || exit 1;
|
||||
cp data/train/$f.new data/train/$f || exit 1;
|
||||
done
|
||||
rm $dir/conv.tab
|
||||
fi
|
||||
|
||||
echo Switchboard-1 data preparation succeeded.
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,89 @@
|
|||
#!/usr/bin/perl -w
|
||||
|
||||
# Author: Peng Qi (pengqi@cs.stanford.edu)
|
||||
# This script maps Switchboard speaker IDs to the true physical speakers
|
||||
# and fixes the utterances IDs accordingly. Expected to be run one level of
|
||||
# directory above.
|
||||
|
||||
sub trim {
|
||||
(my $s = $_[0]) =~ s/^\s+|\s+$//g;
|
||||
return $s;
|
||||
}
|
||||
|
||||
if ($#ARGV != 1) {
|
||||
print "Usage: swbd1_fix_speakerid.pl <swbd-conv-tab-file> <data-dir>\n";
|
||||
print "E.g.: swbd1_fix_speakerid.pl /datasets/SWBD1Transcripts/tables/conv.tab data/train\n";
|
||||
}
|
||||
|
||||
$tab_file = $ARGV[0];
|
||||
$dir = $ARGV[1];
|
||||
|
||||
%conv_to_spk = ();
|
||||
|
||||
open(my $conv_tab, '<', $tab_file) or die "Could not open '$tab_file' $!\n";
|
||||
|
||||
while (my $line = <$conv_tab>) {
|
||||
chomp $line;
|
||||
|
||||
my @fields = split "," , $line;
|
||||
#$fields[0] = trim($fields[0]);
|
||||
$fields[2] = trim($fields[2]);
|
||||
$fields[3] = trim($fields[3]);
|
||||
$conv_to_spk{'sw0' . $fields[0] . '-A'} = $fields[2];
|
||||
$conv_to_spk{'sw0' . $fields[0] . '-B'} = $fields[3];
|
||||
}
|
||||
|
||||
close($conv_tab);
|
||||
|
||||
# fix utt2spk
|
||||
|
||||
%missingconv = ();
|
||||
|
||||
open(my $utt2spk, '<', $dir . '/utt2spk') or die "Could not open '$dir/utt2spk' $!\n";
|
||||
open(my $utt2spk_new, '>', $dir . '/utt2spk.new');
|
||||
|
||||
while (my $line = <$utt2spk>) {
|
||||
chomp $line;
|
||||
|
||||
my @fields = split " " , $line;
|
||||
my $convid = substr $fields[0], 0, 9;
|
||||
|
||||
if (exists $conv_to_spk{ $convid }) {
|
||||
my $spkid = $conv_to_spk{ $convid };
|
||||
$spkid = "sw" . $spkid;
|
||||
my $newuttid = $spkid . '-' . (substr $fields[0], 2);
|
||||
|
||||
print $utt2spk_new "$newuttid $spkid\n";
|
||||
} else {
|
||||
my $convid = substr $convid, 3, 4;
|
||||
$missingconv{$convid} = 1;
|
||||
|
||||
print $utt2spk_new $fields[0]." ".$fields[1]."\n";
|
||||
}
|
||||
}
|
||||
|
||||
close($utt2spk);
|
||||
close($utt2spk_new);
|
||||
|
||||
foreach my $conv (keys %missingconv) {
|
||||
print "Warning: Conversation ID '$conv' not found in conv.tab, retaining old speaker IDs\n"
|
||||
}
|
||||
|
||||
# fix segments and text
|
||||
|
||||
foreach my $file ('segments','text') {
|
||||
open(my $oldfile, '<', "$dir/$file") or die "Could not open '$dir/$file' $!\n";
|
||||
open(my $newfile, '>', "$dir/$file.new");
|
||||
|
||||
while (my $line = <$oldfile>) {
|
||||
chomp $line;
|
||||
|
||||
my $convid = substr $line, 0, 9;
|
||||
if (exists $conv_to_spk{$convid}) {
|
||||
my $spkid = $conv_to_spk{$convid};
|
||||
print $newfile "sw$spkid-" . (substr $line, 2) . "\n";
|
||||
} else {
|
||||
print $newfile "$line\n";
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -15,6 +15,15 @@
|
|||
. path.sh
|
||||
set -e # exit on error
|
||||
|
||||
# Prepare Switchboard data. This command can also take a second optional argument
|
||||
# which specifies the directory to Switchboard documentations. Specifically, if
|
||||
# this argument is given, the script will look for the conv.tab file and correct
|
||||
# speaker IDs to the actual speaker personal identification numbers released in
|
||||
# the documentations. The documentations can be found here:
|
||||
# https://catalog.ldc.upenn.edu/docs/LDC97S62/
|
||||
# Note: if you are using this link, make sure you rename conv_tab.csv to conv.tab
|
||||
# after downloading.
|
||||
# Usage: local/swbd1_data_prep.sh /path/to/SWBD [/path/to/SWBD_docs]
|
||||
local/swbd1_data_prep.sh /export/corpora3/LDC/LDC97S62
|
||||
# local/swbd1_data_prep.sh /home/dpovey/data/LDC97S62
|
||||
# local/swbd1_data_prep.sh /data/corpora0/LDC97S62
|
||||
|
|
|
|||
|
|
@ -14,12 +14,13 @@ use_graphs=false
|
|||
# Begin configuration.
|
||||
scale_opts="--transition-scale=1.0 --self-loop-scale=0.1"
|
||||
acoustic_scale=0.1
|
||||
beam=20.0
|
||||
lattice_beam=10.0
|
||||
beam=15.0
|
||||
lattice_beam=8.0
|
||||
max_active=750
|
||||
transform_dir= # directory to find fMLLR transforms in.
|
||||
top_n_words=100 # Number of common words that we compile into each graph (most frequent
|
||||
# in $lang/text.
|
||||
stage=0
|
||||
stage=-1
|
||||
cleanup=true
|
||||
# End configuration options.
|
||||
|
||||
|
|
@ -64,14 +65,16 @@ cp $srcdir/{tree,final.mdl} $dir || exit 1;
|
|||
cp $srcdir/final.occs $dir;
|
||||
|
||||
|
||||
utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt <$data/text | \
|
||||
awk '{for(x=2;x<=NF;x++) print $x;}' | sort | uniq -c | \
|
||||
sort -rn > $dir/word_counts.int || exit 1;
|
||||
num_words=$(awk '{x+=$1} END{print x}' < $dir/word_counts.int) || exit 1;
|
||||
# print top-n words with their unigram probabilities.
|
||||
if [ $stage -le 0 ]; then
|
||||
utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt <$data/text | \
|
||||
awk '{for(x=2;x<=NF;x++) print $x;}' | sort | uniq -c | \
|
||||
sort -rn > $dir/word_counts.int || exit 1;
|
||||
num_words=$(awk '{x+=$1} END{print x}' < $dir/word_counts.int) || exit 1;
|
||||
# print top-n words with their unigram probabilities.
|
||||
|
||||
head -n $top_n_words $dir/word_counts.int | awk -v tot=$num_words '{print $1/tot, $2;}' >$dir/top_words.int
|
||||
utils/int2sym.pl -f 2 $lang/words.txt <$dir/top_words.int >$dir/top_words.txt
|
||||
head -n $top_n_words $dir/word_counts.int | awk -v tot=$num_words '{print $1/tot, $2;}' >$dir/top_words.int
|
||||
utils/int2sym.pl -f 2 $lang/words.txt <$dir/top_words.int >$dir/top_words.txt
|
||||
fi
|
||||
|
||||
if [ -f $srcdir/final.mat ]; then feat_type=lda; else feat_type=delta; fi
|
||||
echo "$0: feature type is $feat_type"
|
||||
|
|
@ -105,9 +108,9 @@ elif [ -f $srcdir/final.alimdl ]; then
|
|||
fi
|
||||
|
||||
|
||||
echo "$0: decoding $data using utterance-specific decoding graphs using model from $srcdir, output in $dir"
|
||||
if [ $stage -le 1 ]; then
|
||||
echo "$0: decoding $data using utterance-specific decoding graphs using model from $srcdir, output in $dir"
|
||||
|
||||
if [ $stage -le 0 ]; then
|
||||
rm $dir/edits.*.txt $dir/aligned_ref.*.txt 2>/dev/null
|
||||
|
||||
$cmd JOB=1:$nj $dir/log/decode.JOB.log \
|
||||
|
|
@ -116,7 +119,8 @@ if [ $stage -le 0 ]; then
|
|||
compile-train-graphs-fsts $scale_opts --read-disambig-syms=$lang/phones/disambig.int \
|
||||
$dir/tree $dir/final.mdl $lang/L_disambig.fst ark:- ark:- \| \
|
||||
gmm-latgen-faster --acoustic-scale=$acoustic_scale --beam=$beam \
|
||||
--lattice-beam=$lattice_beam --word-symbol-table=$lang/words.txt \
|
||||
--max-active=$max_active --lattice-beam=$lattice_beam \
|
||||
--word-symbol-table=$lang/words.txt \
|
||||
$dir/final.mdl ark:- "$feats" ark:- \| \
|
||||
lattice-oracle ark:- "ark:utils/sym2int.pl --map-oov $oov -f 2- $lang/words.txt $sdata/JOB/text|" \
|
||||
ark,t:- ark,t:$dir/edits.JOB.txt \| \
|
||||
|
|
@ -124,15 +128,16 @@ if [ $stage -le 0 ]; then
|
|||
fi
|
||||
|
||||
|
||||
if [ $stage -le 1 ]; then
|
||||
if [ $stage -le 2 ]; then
|
||||
if [ -f $dir/edits.1.txt ]; then
|
||||
for x in $(seq $nj); do cat $dir/edits.$x.txt; done > $dir/edits.txt
|
||||
for x in $(seq $nj); do cat $dir/aligned_ref.$x.txt; done > $dir/aligned_ref.txt
|
||||
# the awk commands below are to ensure that partially-written files don't confuse us.
|
||||
for x in $(seq $nj); do cat $dir/edits.$x.txt; done | awk '{if(NF==2){print;}}' > $dir/edits.txt
|
||||
for x in $(seq $nj); do cat $dir/aligned_ref.$x.txt; done | awk '{if(NF>=1){print;}}' > $dir/aligned_ref.txt
|
||||
else
|
||||
echo "$0: warning: no file $dir/edits.1.txt, using previously concatenated file if present."
|
||||
fi
|
||||
|
||||
# in case any utterances failed to align, get filtered copy of $data/text that's filtered.
|
||||
# in case any utterances failed to align, get filtered copy of $data/text
|
||||
utils/filter_scp.pl $dir/edits.txt < $data/text > $dir/text
|
||||
cat $dir/text | awk '{print $1, (NF-1);}' > $dir/length.txt
|
||||
|
||||
|
|
@ -162,4 +167,3 @@ if [ $stage -le 1 ]; then
|
|||
rm $dir/edits.*.txt $dir/aligned_ref.*.txt
|
||||
fi
|
||||
fi
|
||||
|
||||
|
|
|
|||
|
|
@ -1,12 +1,21 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright 2012-2013 Johns Hopkins University (Author: Daniel Povey)
|
||||
# Copyright 2012-2014 Johns Hopkins University (Author: Daniel Povey)
|
||||
# Vimal Manohar
|
||||
# Apache 2.0
|
||||
|
||||
##Changes
|
||||
# Vimal Manohar (Jan 2014):
|
||||
# Added options to boost silence probabilities in the model before
|
||||
# decoding. This can help in favoring the silence phones when
|
||||
# some silence regions are wrongly decoded as speech phones like glottal stops
|
||||
|
||||
# Begin configuration section.
|
||||
transform_dir=
|
||||
iter=
|
||||
model= # You can specify the model to use (e.g. if you want to use the .alimdl)
|
||||
boost_silence=1.0 # Boost silence pdfs in the model by this factor before decoding
|
||||
silence_phones_list= # List of silence phones that would be boosted before decoding
|
||||
stage=0
|
||||
nj=4
|
||||
cmd=run.pl
|
||||
|
|
@ -27,6 +36,8 @@ echo "$0 $@" # Print the command line for logging
|
|||
[ -f ./path.sh ] && . ./path.sh; # source the path.
|
||||
. parse_options.sh || exit 1;
|
||||
|
||||
[ -z $silence_phones_list ] && boost_silence=1.0
|
||||
|
||||
if [ $# != 3 ]; then
|
||||
echo "Usage: $0 [options] <graph-dir> <data-dir> <decode-dir>"
|
||||
echo "... where <decode-dir> is assumed to be a sub-directory of the directory"
|
||||
|
|
@ -106,10 +117,13 @@ if [ $stage -le 0 ]; then
|
|||
words="ark:/dev/null"
|
||||
fi
|
||||
|
||||
[ ! -z "$silence_phones_list" ] && \
|
||||
model="gmm-boost-silence --boost=$boost_silence $silence_phones_list $model - |"
|
||||
|
||||
$cmd $parallel_opts JOB=1:$nj $dir/log/decode.JOB.log \
|
||||
gmm-decode-faster$thread_string --max-active=$max_active --beam=$beam \
|
||||
--acoustic-scale=$acwt --allow-partial=true --word-symbol-table=$graphdir/words.txt \
|
||||
$model $graphdir/HCLG.fst "$feats" "$words" "$ali" || exit 1;
|
||||
"$model" $graphdir/HCLG.fst "$feats" "$words" "$ali" || exit 1;
|
||||
fi
|
||||
|
||||
exit 0;
|
||||
|
|
|
|||
|
|
@ -20,7 +20,7 @@ gselect=15 # Number of Gaussian-selection indices for SGMMs. [Note:
|
|||
first_pass_gselect=3 # Use a smaller number of Gaussian-selection indices in
|
||||
# the 1st pass of decoding (lattice generation).
|
||||
max_active=7000
|
||||
|
||||
max_mem=50000000
|
||||
#WARNING: This option is renamed lattice_beam (it was renamed to follow the naming
|
||||
# in the other scripts
|
||||
lattice_beam=6.0 # Beam we use in lattice generation.
|
||||
|
|
@ -131,7 +131,7 @@ if [ $stage -le 2 ]; then
|
|||
$cmd $parallel_opts JOB=1:$nj $dir/log/decode_pass1.JOB.log \
|
||||
sgmm2-latgen-faster$thread_string --max-active=$max_active --beam=$beam --lattice-beam=$lattice_beam \
|
||||
--acoustic-scale=$acwt --determinize-lattice=false --allow-partial=true \
|
||||
--word-symbol-table=$graphdir/words.txt "$gselect_opt_1stpass" $alignment_model \
|
||||
--word-symbol-table=$graphdir/words.txt --max-mem=$max_mem "$gselect_opt_1stpass" $alignment_model \
|
||||
$graphdir/HCLG.fst "$feats" "ark:|gzip -c > $dir/pre_lat.JOB.gz" || exit 1;
|
||||
fi
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,56 @@
|
|||
#!/bin/bash
|
||||
|
||||
# Copyright 2014 Johns Hopkins University (Author: Daniel Povey).
|
||||
# Apache 2.0.
|
||||
|
||||
# This script converts nnet1 into nnet2 models.
|
||||
# Note, it doesn't support all possible types of nnet1 models.
|
||||
|
||||
# Begin configuration section
|
||||
cleanup=true
|
||||
cmd=run.pl
|
||||
# End configuration section.
|
||||
|
||||
echo "$0 $@" # Print the command line for logging
|
||||
|
||||
[ -f ./path.sh ] && . ./path.sh; # source the path.
|
||||
. parse_options.sh || exit 1;
|
||||
|
||||
|
||||
if [ $# -ne 2 ]; then
|
||||
echo "Usage: $0 [options] <src-nnet1-dir> <dest-nnet2-dir>"
|
||||
echo "e.g.: $0 exp/dnn4b_pretrain-dbn_dnn_smbr exp/dnn4b_smbr_nnet2"
|
||||
exit 1;
|
||||
fi
|
||||
|
||||
src=$1
|
||||
dir=$2
|
||||
|
||||
mkdir -p $dir/log || exit 1;
|
||||
|
||||
for f in $src/final.mdl $src/final.nnet $src/final.feature_transform $src/ali_train_pdf.counts; do
|
||||
[ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
|
||||
done
|
||||
|
||||
|
||||
# We could do the following things all as one long piped command,
|
||||
# but it will be easier to debug if we make them separate.
|
||||
|
||||
$cmd $dir/log/convert_feature_transform.log \
|
||||
nnet1-to-raw-nnet $src/final.feature_transform $dir/0.raw || exit 1;
|
||||
|
||||
$cmd $dir/log/convert_model.log \
|
||||
nnet1-to-raw-nnet $src/final.nnet $dir/1.raw || exit 1;
|
||||
|
||||
$cmd $dir/log/append_model.log \
|
||||
raw-nnet-concat $dir/0.raw $dir/1.raw $dir/concat.raw || exit 1;
|
||||
|
||||
$cmd $dir/log/init_model.log \
|
||||
nnet-am-init $src/final.mdl $dir/concat.raw $dir/final_noprior.mdl || exit 1;
|
||||
|
||||
$cmd $dir/log/set_priors.log \
|
||||
nnet-adjust-priors $dir/final_noprior.mdl $src/ali_train_pdf.counts $dir/final.mdl || exit 1;
|
||||
|
||||
if $cleanup; then
|
||||
rm $dir/0.raw $dir/1.raw $dir/concat.raw $dir/final_noprior.mdl
|
||||
fi
|
||||
|
|
@ -65,6 +65,7 @@ for f in $graphdir/HCLG.fst $data/feats.scp $model $extra_files; do
|
|||
done
|
||||
|
||||
sdata=$data/split$nj;
|
||||
cmvn_opts=`cat $srcdir/cmvn_opts 2>/dev/null`
|
||||
thread_string=
|
||||
[ $num_threads -gt 1 ] && thread_string="-parallel --num-threads=$num_threads"
|
||||
|
||||
|
|
@ -79,7 +80,6 @@ if [ -z "$feat_type" ]; then
|
|||
echo "$0: feature type is $feat_type"
|
||||
fi
|
||||
|
||||
cmvn_opts=`cat $srcdir/cmvn_opts 2>/dev/null`
|
||||
splice_opts=`cat $srcdir/splice_opts 2>/dev/null`
|
||||
|
||||
case $feat_type in
|
||||
|
|
@ -90,16 +90,30 @@ case $feat_type in
|
|||
esac
|
||||
if [ ! -z "$transform_dir" ]; then
|
||||
echo "$0: using transforms from $transform_dir"
|
||||
if [ "$feat_type" == "lda" ]; then
|
||||
[ ! -f $transform_dir/trans.1 ] && echo "$0: no such file $transform_dir/trans.1" && exit 1;
|
||||
[ "$nj" -ne "`cat $transform_dir/num_jobs`" ] \
|
||||
&& echo "$0: #jobs mismatch with transform-dir." && exit 1;
|
||||
feats="$feats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk ark,s,cs:$transform_dir/trans.JOB ark:- ark:- |"
|
||||
[ ! -s $transform_dir/num_jobs ] && \
|
||||
echo "$0: expected $transform_dir/num_jobs to contain the number of jobs." && exit 1;
|
||||
nj_orig=$(cat $transform_dir/num_jobs)
|
||||
|
||||
if [ $feat_type == "raw" ]; then trans=raw_trans;
|
||||
else trans=trans; fi
|
||||
if [ $feat_type == "lda" ] && \
|
||||
! cmp $transform_dir/../final.mat $srcdir/final.mat && \
|
||||
! cmp $transform_dir/final.mat $srcdir/final.mat; then
|
||||
echo "$0: LDA transforms differ between $srcdir and $transform_dir"
|
||||
exit 1;
|
||||
fi
|
||||
if [ ! -f $transform_dir/$trans.1 ]; then
|
||||
echo "$0: expected $transform_dir/$trans.1 to exist (--transform-dir option)"
|
||||
exit 1;
|
||||
fi
|
||||
if [ $nj -ne $nj_orig ]; then
|
||||
# Copy the transforms into an archive with an index.
|
||||
for n in $(seq $nj_orig); do cat $transform_dir/$trans.$n; done | \
|
||||
copy-feats ark:- ark,scp:$dir/$trans.ark,$dir/$trans.scp || exit 1;
|
||||
feats="$feats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk scp:$dir/$trans.scp ark:- ark:- |"
|
||||
else
|
||||
[ ! -f $transform_dir/raw_trans.1 ] && echo "$0: no such file $transform_dir/raw_trans.1" && exit 1;
|
||||
[ "$nj" -ne "`cat $transform_dir/num_jobs`" ] \
|
||||
&& echo "$0: #jobs mismatch with transform-dir." && exit 1;
|
||||
feats="$feats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk ark,s,cs:$transform_dir/raw_trans.JOB ark:- ark:- |"
|
||||
# number of jobs matches with alignment dir.
|
||||
feats="$feats transform-feats --utt2spk=ark:$sdata/JOB/utt2spk ark:$transform_dir/$trans.JOB ark:- ark:- |"
|
||||
fi
|
||||
elif grep 'transform-feats --utt2spk' $srcdir/log/train.1.log >&/dev/null; then
|
||||
echo "$0: **WARNING**: you seem to be using a neural net system trained with transforms,"
|
||||
|
|
|
|||
|
|
@ -36,6 +36,7 @@ perspk=false
|
|||
first_opt=""
|
||||
speakers=false
|
||||
spk_list_specified=false
|
||||
utt_list_specified=false
|
||||
|
||||
if [ "$1" == "--per-spk" ]; then
|
||||
perspk=true;
|
||||
|
|
@ -55,6 +56,9 @@ elif [ "$1" == "--last" ]; then
|
|||
elif [ "$1" == "--spk-list" ]; then
|
||||
spk_list_specified=true
|
||||
shift;
|
||||
elif [ "$1" == "--utt-list" ]; then
|
||||
utt_list_specified=true
|
||||
shift;
|
||||
fi
|
||||
|
||||
|
||||
|
|
@ -64,6 +68,7 @@ if [ $# != 3 ]; then
|
|||
echo "Usage: "
|
||||
echo " subset_data_dir.sh [--speakers|--shortest|--first|--last|--per-spk] <srcdir> <num-utt> <destdir>"
|
||||
echo " subset_data_dir.sh [--spk-list <speaker-list-file>] <srcdir> <destdir>"
|
||||
echo " subset_data_dir.sh [--utt-list <utterance-list-file>] <srcdir> <destdir>"
|
||||
echo "By default, randomly selects <num-utt> utterances from the data directory."
|
||||
echo "With --speakers, randomly selects enough speakers that we have <num-utt> utterances"
|
||||
echo "With --per-spk, selects <num-utt> utterances per speaker, if available."
|
||||
|
|
@ -78,6 +83,10 @@ if $spk_list_specified; then
|
|||
spk_list=$1
|
||||
srcdir=$2
|
||||
destdir=$3
|
||||
elif $utt_list_specified; then
|
||||
utt_list=$1
|
||||
srcdir=$2
|
||||
destdir=$3
|
||||
else
|
||||
srcdir=$1
|
||||
numutt=$2
|
||||
|
|
@ -127,7 +136,13 @@ if $spk_list_specified; then
|
|||
utils/filter_scp.pl "$spk_list" $srcdir/spk2utt > $destdir/spk2utt || exit 1;
|
||||
utils/spk2utt_to_utt2spk.pl < $destdir/spk2utt > $destdir/utt2spk || exit 1;
|
||||
do_filtering; # bash function.
|
||||
exit 0;
|
||||
exit 0;
|
||||
elif $utt_list_specified; then
|
||||
mkdir -p $destdir
|
||||
utils/filter_scp.pl "$utt_list" $srcdir/utt2spk > $destdir/utt2spk || exit 1;
|
||||
utils/utt2spk_to_spk2utt.pl < $destdir/utt2spk > $destdir/spk2utt || exit 1;
|
||||
do_filtering; # bash function.
|
||||
exit 0;
|
||||
elif $speakers; then
|
||||
mkdir -p $destdir
|
||||
utils/shuffle_list.pl < $srcdir/spk2utt | awk -v numutt=$numutt '{ if (tot < numutt){ print; } tot += (NF-1); }' | \
|
||||
|
|
|
|||
|
|
@ -2,23 +2,7 @@
|
|||
|
||||
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey). Apache 2.0.
|
||||
|
||||
@ARGV != 1 && print STDERR "Usage: summarize_warnings.pl <log-dir>\n" && exit 1;
|
||||
|
||||
$dir = $ARGV[0];
|
||||
|
||||
! -d $dir && print STDERR "summarize_warnings.pl: no such directory $dir\n" && exit 1;
|
||||
|
||||
$dir =~ s:/$::; # Remove trailing slash.
|
||||
|
||||
|
||||
# Group the files into categories where all have the same base-name.
|
||||
foreach $f (glob ("$dir/*.log")) {
|
||||
$f_category = $f;
|
||||
# do next expression twice; s///g doesn't work as they overlap.
|
||||
$f_category =~ s:\.\d+\.:.*.:;
|
||||
$f_category =~ s:\.\d+\.:.*.:;
|
||||
$fmap{$f_category} .= " $f";
|
||||
}
|
||||
#scalar(@ARGV) >= 1 && print STDERR "Usage: summarize_warnings.pl <log-dir>\n" && exit 1;
|
||||
|
||||
sub split_hundreds { # split list of filenames into groups of 100.
|
||||
my $names = shift @_;
|
||||
|
|
@ -35,7 +19,53 @@ sub split_hundreds { # split list of filenames into groups of 100.
|
|||
return @ans;
|
||||
}
|
||||
|
||||
foreach $c (keys %fmap) {
|
||||
sub parse_accounting_entry {
|
||||
$entry= shift @_;
|
||||
|
||||
@elems = split " ", $entry;
|
||||
|
||||
$time=undef;
|
||||
$threads=undef;
|
||||
foreach $elem (@elems) {
|
||||
if ( $elem=~ m/time=(\d+)/ ) {
|
||||
$elem =~ s/time=(\d+)/$1/;
|
||||
$time = $elem;
|
||||
} elsif ( $elem=~ m/threads=(\d+)/ ) {
|
||||
$elem =~ s/threads=(\d+)/$1/g;
|
||||
$threads = $elem;
|
||||
} else {
|
||||
die "Unknown entry \"$elem\" when parsing \"$entry\" \n";
|
||||
}
|
||||
}
|
||||
|
||||
if (defined($time) and defined($threads) ) {
|
||||
return ($time, $threads);
|
||||
} else {
|
||||
die "The accounting entry \"$entry\" did not contain all necessary attributes";
|
||||
}
|
||||
}
|
||||
|
||||
foreach $dir (@ARGV) {
|
||||
|
||||
#$dir = $ARGV[0];
|
||||
print $dir
|
||||
|
||||
! -d $dir && print STDERR "summarize_warnings.pl: no such directory $dir\n" ;
|
||||
|
||||
$dir =~ s:/$::; # Remove trailing slash.
|
||||
|
||||
|
||||
# Group the files into categories where all have the same base-name.
|
||||
foreach $f (glob ("$dir/*.log")) {
|
||||
$f_category = $f;
|
||||
# do next expression twice; s///g doesn't work as they overlap.
|
||||
$f_category =~ s:\.\d+\.(?!\d+):.*.:;
|
||||
#$f_category =~ s:\.\d+\.:.*.:;
|
||||
$fmap{$f_category} .= " $f";
|
||||
}
|
||||
}
|
||||
|
||||
foreach $c (sort (keys %fmap) ) {
|
||||
$n = 0;
|
||||
foreach $fgroup (split_hundreds($fmap{$c})) {
|
||||
$n += `grep -w WARNING $fgroup | wc -l`;
|
||||
|
|
@ -44,7 +74,7 @@ foreach $c (keys %fmap) {
|
|||
print "$n warnings in $c\n"
|
||||
}
|
||||
}
|
||||
foreach $c (keys %fmap) {
|
||||
foreach $c (sort (keys %fmap)) {
|
||||
$n = 0;
|
||||
foreach $fgroup (split_hundreds($fmap{$c})) {
|
||||
$n += `grep -w ERROR $fgroup | wc -l`;
|
||||
|
|
@ -53,3 +83,39 @@ foreach $c (keys %fmap) {
|
|||
print "$n errors in $c\n"
|
||||
}
|
||||
}
|
||||
|
||||
$supertotal_cpu_time=0.0;
|
||||
$supertotal_clock_time=0.0;
|
||||
$supertotal_threads=0.0;
|
||||
|
||||
foreach $c (sort (keys %fmap)) {
|
||||
$n = 0;
|
||||
|
||||
$total_cpu_time=0.0;
|
||||
$total_clock_time=0.0;
|
||||
$total_threads=0.0;
|
||||
foreach $fgroup (split_hundreds($fmap{$c})) {
|
||||
$lines=`grep -a "# Accounting: " $fgroup |sed 's/.* Accounting: *//g'`;
|
||||
|
||||
#print $lines ."\n";
|
||||
|
||||
@entries = split "\n", $lines;
|
||||
|
||||
foreach $line (@entries) {
|
||||
$time, $threads = parse_accounting_entry($line);
|
||||
|
||||
$total_cpu_time += $time * $threads;
|
||||
$total_threads += $threads;
|
||||
if ( $time > $total_clock_time ) {
|
||||
$total_clock_time += $time;
|
||||
}
|
||||
}
|
||||
}
|
||||
print "total_cpu_time=$total_cpu_time clock_time=$total_clock_time total_threads=$total_threads group=$c\n";
|
||||
|
||||
$supertotal_cpu_time += $total_cpu_time;
|
||||
$supertotal_clock_time += $total_clock_time;
|
||||
$supertotal_threads += $total_threads;
|
||||
}
|
||||
print "total_cpu_time=$supertotal_cpu_time clock_time=$supertotal_clock_time total_threads=$supertotal_threads group=all\n";
|
||||
|
||||
|
|
|
|||
|
|
@ -128,8 +128,8 @@ void Randomize(const CuMatrixBase<Real> &src,
|
|||
|
||||
template<typename Real>
|
||||
void Splice(const CuMatrix<Real> &src, const CuArray<int32> &frame_offsets,
|
||||
CuMatrix<Real> *tgt) {
|
||||
|
||||
CuMatrixBase<Real> *tgt) {
|
||||
|
||||
KALDI_ASSERT(src.NumCols()*frame_offsets.Dim() == tgt->NumCols());
|
||||
KALDI_ASSERT(src.NumRows() == tgt->NumRows());
|
||||
|
||||
|
|
@ -167,7 +167,8 @@ void Splice(const CuMatrix<Real> &src, const CuArray<int32> &frame_offsets,
|
|||
|
||||
|
||||
template<typename Real>
|
||||
void Copy(const CuMatrix<Real> &src, const CuArray<int32> ©_from_indices, CuMatrix<Real> *tgt) {
|
||||
void Copy(const CuMatrix<Real> &src, const CuArray<int32> ©_from_indices,
|
||||
CuMatrixBase<Real> *tgt) {
|
||||
|
||||
KALDI_ASSERT(copy_from_indices.Dim() == tgt->NumCols());
|
||||
KALDI_ASSERT(src.NumRows() == tgt->NumRows());
|
||||
|
|
@ -207,13 +208,17 @@ template
|
|||
void RegularizeL1(CuMatrixBase<double> *weight, CuMatrixBase<double> *grad, double l1, double lr);
|
||||
|
||||
template
|
||||
void Splice(const CuMatrix<float> &src, const CuArray<int32> &frame_offsets, CuMatrix<float> *tgt);
|
||||
void Splice(const CuMatrix<float> &src, const CuArray<int32> &frame_offsets,
|
||||
CuMatrixBase<float> *tgt);
|
||||
template
|
||||
void Splice(const CuMatrix<double> &src, const CuArray<int32> &frame_offsets, CuMatrix<double> *tgt);
|
||||
void Splice(const CuMatrix<double> &src, const CuArray<int32> &frame_offsets,
|
||||
CuMatrixBase<double> *tgt);
|
||||
template
|
||||
void Copy(const CuMatrix<float> &src, const CuArray<int32> ©_from_indices, CuMatrix<float> *tgt);
|
||||
void Copy(const CuMatrix<float> &src, const CuArray<int32> ©_from_indices,
|
||||
CuMatrixBase<float> *tgt);
|
||||
template
|
||||
void Copy(const CuMatrix<double> &src, const CuArray<int32> ©_from_indices, CuMatrix<double> *tgt);
|
||||
void Copy(const CuMatrix<double> &src, const CuArray<int32> ©_from_indices,
|
||||
CuMatrixBase<double> *tgt);
|
||||
|
||||
template
|
||||
void Randomize(const CuMatrixBase<float> &src,
|
||||
|
|
|
|||
|
|
@ -61,7 +61,7 @@ void Randomize(const CuMatrixBase<Real> &src,
|
|||
template<typename Real>
|
||||
void Splice(const CuMatrix<Real> &src,
|
||||
const CuArray<int32> &frame_offsets,
|
||||
CuMatrix<Real> *tgt);
|
||||
CuMatrixBase<Real> *tgt);
|
||||
|
||||
/// Copies elements from src into tgt as given by copy_from_indices.
|
||||
/// The matrices src and tgt must have the same dimensions and
|
||||
|
|
@ -71,7 +71,7 @@ void Splice(const CuMatrix<Real> &src,
|
|||
template<typename Real>
|
||||
void Copy(const CuMatrix<Real> &src,
|
||||
const CuArray<int32> ©_from_indices,
|
||||
CuMatrix<Real> *tgt);
|
||||
CuMatrixBase<Real> *tgt);
|
||||
|
||||
|
||||
} // namespace cu
|
||||
|
|
|
|||
|
|
@ -78,10 +78,10 @@ class CuMatrixBase {
|
|||
CuMatrixBase<Real> *grad, Real l1, Real lr);
|
||||
friend void cu::Splice<Real>(const CuMatrix<Real> &src,
|
||||
const CuArray<int32> &frame_offsets,
|
||||
CuMatrix<Real> *tgt);
|
||||
CuMatrixBase<Real> *tgt);
|
||||
friend void cu::Copy<Real>(const CuMatrix<Real> &src,
|
||||
const CuArray<int32> ©_from_indices,
|
||||
CuMatrix<Real> *tgt);
|
||||
CuMatrixBase<Real> *tgt);
|
||||
friend void cu::Randomize<Real>(const CuMatrixBase<Real> &src,
|
||||
const CuArray<int32> ©_from_idx,
|
||||
CuMatrixBase<Real> *tgt);
|
||||
|
|
@ -290,9 +290,9 @@ class CuMatrixBase {
|
|||
void InvertElements();
|
||||
/// B = alpha * A
|
||||
void AddMat(Real alpha, const CuMatrixBase<Real> &A, MatrixTransposeType transA = kNoTrans);
|
||||
/// B = alpha * row + beta * B
|
||||
/// (for each column c of *this), c = alpha * col + beta * c
|
||||
void AddVecToCols(Real alpha, const CuVectorBase<Real> &col, Real beta = 1.0);
|
||||
/// B = alpha * row + beta * B
|
||||
/// (for each row r of *this), r = alpha * row + beta * r
|
||||
void AddVecToRows(Real alpha, const CuVectorBase<Real> &row, Real beta = 1.0);
|
||||
/// C = alpha * A(^T)*B(^T) + beta * C
|
||||
void AddMatMat(Real alpha, const CuMatrixBase<Real> &A, MatrixTransposeType transA,
|
||||
|
|
|
|||
|
|
@ -59,7 +59,7 @@ class CuVectorBase {
|
|||
const CuVectorBase<OtherReal> &v2);
|
||||
friend void cu::Splice<Real>(const CuMatrix<Real> &src,
|
||||
const CuArray<int32> &frame_offsets,
|
||||
CuMatrix<Real> *tgt);
|
||||
CuMatrixBase<Real> *tgt);
|
||||
friend class CuRand<Real>;
|
||||
|
||||
/// Dimensions
|
||||
|
|
|
|||
|
|
@ -171,7 +171,7 @@ class AffineTransform : public UpdatableComponent {
|
|||
}
|
||||
|
||||
/// Accessors to the component parameters
|
||||
const CuVector<BaseFloat>& GetBias() {
|
||||
const CuVector<BaseFloat>& GetBias() const {
|
||||
return bias_;
|
||||
}
|
||||
|
||||
|
|
@ -180,7 +180,7 @@ class AffineTransform : public UpdatableComponent {
|
|||
bias_.CopyFromVec(bias);
|
||||
}
|
||||
|
||||
const CuMatrix<BaseFloat>& GetLinearity() {
|
||||
const CuMatrix<BaseFloat>& GetLinearity() const {
|
||||
return linearity_;
|
||||
}
|
||||
|
||||
|
|
@ -190,11 +190,11 @@ class AffineTransform : public UpdatableComponent {
|
|||
linearity_.CopyFromMat(linearity);
|
||||
}
|
||||
|
||||
const CuVector<BaseFloat>& GetBiasCorr() {
|
||||
const CuVector<BaseFloat>& GetBiasCorr() const {
|
||||
return bias_corr_;
|
||||
}
|
||||
|
||||
const CuMatrix<BaseFloat>& GetLinearityCorr() {
|
||||
const CuMatrix<BaseFloat>& GetLinearityCorr() const {
|
||||
return linearity_corr_;
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -119,7 +119,7 @@ std::string MomentStatistics(const CuMatrix<Real> &mat) {
|
|||
* in N, out k*N, FrameOffset o_1,o_2,...,o_k
|
||||
* FrameOffset example 11frames: -5 -4 -3 -2 -1 0 1 2 3 4 5
|
||||
*/
|
||||
class Splice : public Component {
|
||||
class Splice: public Component {
|
||||
public:
|
||||
Splice(int32 dim_in, int32 dim_out)
|
||||
: Component(dim_in, dim_out)
|
||||
|
|
|
|||
|
|
@ -506,7 +506,7 @@ void UnitTestAffinePreconInputComponent() {
|
|||
|
||||
void UnitTestBlockAffineComponent() {
|
||||
BaseFloat learning_rate = 0.01,
|
||||
param_stddev = 0.1, bias_stddev = 1.0;
|
||||
param_stddev = 0.1, bias_stddev = 0.1;
|
||||
int32 num_blocks = 1 + rand() % 3,
|
||||
input_dim = num_blocks * (2 + rand() % 4),
|
||||
output_dim = num_blocks * (2 + rand() % 4);
|
||||
|
|
@ -655,6 +655,28 @@ void UnitTestFixedAffineComponent() {
|
|||
}
|
||||
}
|
||||
|
||||
void UnitTestFixedScaleComponent() {
|
||||
int32 m = 1 + rand() % 20;
|
||||
{
|
||||
CuVector<BaseFloat> vec(m);
|
||||
vec.SetRandn();
|
||||
FixedScaleComponent component;
|
||||
component.Init(vec);
|
||||
UnitTestGenericComponentInternal(component);
|
||||
}
|
||||
}
|
||||
|
||||
void UnitTestFixedBiasComponent() {
|
||||
int32 m = 1 + rand() % 20;
|
||||
{
|
||||
CuVector<BaseFloat> vec(m);
|
||||
vec.SetRandn();
|
||||
FixedBiasComponent component;
|
||||
component.Init(vec);
|
||||
UnitTestGenericComponentInternal(component);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
void UnitTestParsing() {
|
||||
|
|
@ -825,6 +847,8 @@ int main() {
|
|||
UnitTestDctComponent();
|
||||
UnitTestFixedLinearComponent();
|
||||
UnitTestFixedAffineComponent();
|
||||
UnitTestFixedScaleComponent();
|
||||
UnitTestFixedBiasComponent();
|
||||
UnitTestAffineComponentPreconditioned();
|
||||
UnitTestAffineComponentPreconditionedOnline();
|
||||
UnitTestAffineComponentModified();
|
||||
|
|
|
|||
|
|
@ -98,6 +98,10 @@ Component* Component::NewComponentOfType(const std::string &component_type) {
|
|||
ans = new FixedLinearComponent();
|
||||
} else if (component_type == "FixedAffineComponent") {
|
||||
ans = new FixedAffineComponent();
|
||||
} else if (component_type == "FixedScaleComponent") {
|
||||
ans = new FixedScaleComponent();
|
||||
} else if (component_type == "FixedBiasComponent") {
|
||||
ans = new FixedBiasComponent();
|
||||
} else if (component_type == "SpliceComponent") {
|
||||
ans = new SpliceComponent();
|
||||
} else if (component_type == "SpliceMaxComponent") {
|
||||
|
|
@ -290,6 +294,15 @@ Component *PermuteComponent::Copy() const {
|
|||
ans->reorder_ = reorder_;
|
||||
return ans;
|
||||
}
|
||||
void PermuteComponent::Init(const std::vector<int32> &reorder) {
|
||||
reorder_ = reorder;
|
||||
KALDI_ASSERT(!reorder.empty());
|
||||
std::vector<int32> indexes(reorder);
|
||||
std::sort(indexes.begin(), indexes.end());
|
||||
for (int32 i = 0; i < static_cast<int32>(indexes.size()); i++)
|
||||
KALDI_ASSERT(i == indexes[i] && "Not a permutation");
|
||||
}
|
||||
|
||||
|
||||
std::string Component::Info() const {
|
||||
std::stringstream stream;
|
||||
|
|
@ -1071,6 +1084,19 @@ AffineComponent::AffineComponent(const AffineComponent &component):
|
|||
bias_params_(component.bias_params_),
|
||||
is_gradient_(component.is_gradient_) { }
|
||||
|
||||
AffineComponent::AffineComponent(const CuMatrix<BaseFloat> &linear_params,
|
||||
const CuVector<BaseFloat> &bias_params,
|
||||
BaseFloat learning_rate):
|
||||
UpdatableComponent(learning_rate),
|
||||
linear_params_(linear_params),
|
||||
bias_params_(bias_params) {
|
||||
KALDI_ASSERT(linear_params.NumRows() == bias_params.Dim()&&
|
||||
bias_params.Dim() != 0);
|
||||
is_gradient_ = false;
|
||||
}
|
||||
|
||||
|
||||
|
||||
void AffineComponent::SetZero(bool treat_as_gradient) {
|
||||
if (treat_as_gradient) {
|
||||
SetLearningRate(1.0);
|
||||
|
|
@ -1193,11 +1219,16 @@ void AffineComponent::InitFromString(std::string args) {
|
|||
void AffineComponent::Propagate(const CuMatrixBase<BaseFloat> &in,
|
||||
int32, // num_chunks
|
||||
CuMatrix<BaseFloat> *out) const {
|
||||
KALDI_LOG << "First element of input is " << in(0, 0);
|
||||
KALDI_LOG << "Input sum is " << in.Sum();
|
||||
// No need for asserts as they'll happen within the matrix operations.
|
||||
out->Resize(in.NumRows(), linear_params_.NumRows());
|
||||
out->CopyRowsFromVec(bias_params_); // copies bias_params_ to each row
|
||||
// of *out.
|
||||
KALDI_LOG << "First element of output is " << (*out)(0, 0);
|
||||
KALDI_LOG << "Linearity sum is " << linear_params_.Sum();
|
||||
out->AddMatMat(1.0, in, kNoTrans, linear_params_, kTrans, 1.0);
|
||||
KALDI_LOG << "First element of output is " << (*out)(0, 0);
|
||||
}
|
||||
|
||||
void AffineComponent::UpdateSimple(const CuMatrixBase<BaseFloat> &in_value,
|
||||
|
|
@ -3435,86 +3466,50 @@ void SpliceComponent::Propagate(const CuMatrixBase<BaseFloat> &in,
|
|||
<< "Probably a code error.";
|
||||
out->Resize(num_chunks * output_chunk_size, output_dim);
|
||||
|
||||
if (0) { // rand() % 2 == 0) { // Occasionally do the older code,
|
||||
// this will flag any inconsistency in the tests.
|
||||
|
||||
for (int32 chunk = 0; chunk < num_chunks; chunk++) {
|
||||
CuSubMatrix<BaseFloat> input_chunk(in,
|
||||
chunk * input_chunk_size, input_chunk_size,
|
||||
0, input_dim),
|
||||
output_chunk(*out,
|
||||
chunk * output_chunk_size, output_chunk_size,
|
||||
0, output_dim);
|
||||
// 'indexes' is, for each index from 0 to (left_context_+right_context_+1)-1,
|
||||
// then for each row of "out", the corresponding row of "in" that we copy from.
|
||||
int32 num_splice = left_context_ + right_context_ + 1,
|
||||
const_dim = const_component_dim_;
|
||||
std::vector<std::vector<int32> > indexes(num_splice);
|
||||
// const_component_dim_ != 0, "const_indexes" will be used to determine which
|
||||
// row of "in" we copy the last part of each row of "out" from (this part is
|
||||
// not subject to splicing, it's assumed constant for each frame of "input".
|
||||
std::vector<int32> const_indexes(const_dim == 0 ? 0 : out->NumRows());
|
||||
|
||||
for (int32 c = 0; c < left_context_ + right_context_ + 1; c++) {
|
||||
CuSubMatrix<BaseFloat> input_part(input_chunk,
|
||||
c, output_chunk_size,
|
||||
0, input_dim - const_component_dim_),
|
||||
output_part(output_chunk,
|
||||
0, output_chunk_size,
|
||||
(input_dim - const_component_dim_) * c,
|
||||
input_dim - const_component_dim_);
|
||||
output_part.CopyFromMat(input_part);
|
||||
}
|
||||
//Append the constant component at the end of the output vector
|
||||
if (const_component_dim_ != 0) {
|
||||
CuSubMatrix<BaseFloat> input_part(input_chunk,
|
||||
0, output_chunk_size,
|
||||
InputDim() - const_component_dim_,
|
||||
const_component_dim_),
|
||||
output_part(output_chunk,
|
||||
0, output_chunk_size,
|
||||
OutputDim() - const_component_dim_,
|
||||
const_component_dim_);
|
||||
output_part.CopyFromMat(input_part);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
// 'indexes' is, for each index from 0 to (left_context_+right_context_+1)-1,
|
||||
// then for each row of "out", the corresponding row of "in" that we copy from.
|
||||
int32 num_splice = left_context_ + right_context_ + 1,
|
||||
const_dim = const_component_dim_;
|
||||
std::vector<std::vector<int32> > indexes(num_splice);
|
||||
// const_component_dim_ != 0, "const_indexes" will be used to determine which
|
||||
// row of "in" we copy the last part of each row of "out" from (this part is
|
||||
// not subject to splicing, it's assumed constant for each frame of "input".
|
||||
std::vector<int32> const_indexes(const_dim == 0 ? 0 : out->NumRows());
|
||||
for (int32 c = 0; c < num_splice; c++)
|
||||
indexes[c].resize(out->NumRows());
|
||||
|
||||
for (int32 c = 0; c < num_splice; c++)
|
||||
indexes[c].resize(out->NumRows());
|
||||
|
||||
for (int32 chunk = 0; chunk < num_chunks; chunk++) {
|
||||
for (int32 c = 0; c < num_splice; c++) {
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++) {
|
||||
indexes[c][chunk * output_chunk_size + offset] =
|
||||
chunk * input_chunk_size + c + offset;
|
||||
}
|
||||
}
|
||||
if (const_dim != 0) {
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++)
|
||||
const_indexes[chunk * output_chunk_size + offset] =
|
||||
chunk * input_chunk_size + offset; // there is
|
||||
// an arbitrariness here; since we assume the const_component
|
||||
// is constant within a chunk, it doesn't matter from where we copy.
|
||||
}
|
||||
}
|
||||
for (int32 chunk = 0; chunk < num_chunks; chunk++) {
|
||||
for (int32 c = 0; c < num_splice; c++) {
|
||||
int32 dim = input_dim - const_dim; // dimension we
|
||||
// are splicing
|
||||
CuSubMatrix<BaseFloat> in_part(in, 0, in.NumRows(),
|
||||
0, dim),
|
||||
out_part(*out, 0, out->NumRows(),
|
||||
c * dim, dim);
|
||||
out_part.CopyRows(in_part, indexes[c]);
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++) {
|
||||
indexes[c][chunk * output_chunk_size + offset] =
|
||||
chunk * input_chunk_size + c + offset;
|
||||
}
|
||||
}
|
||||
if (const_dim != 0) {
|
||||
CuSubMatrix<BaseFloat> in_part(in, 0, in.NumRows(),
|
||||
in.NumCols() - const_dim, const_dim),
|
||||
out_part(*out, 0, out->NumRows(),
|
||||
out->NumCols() - const_dim, const_dim);
|
||||
out_part.CopyRows(in_part, const_indexes);
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++)
|
||||
const_indexes[chunk * output_chunk_size + offset] =
|
||||
chunk * input_chunk_size + offset; // there is
|
||||
// an arbitrariness here; since we assume the const_component
|
||||
// is constant within a chunk, it doesn't matter from where we copy.
|
||||
}
|
||||
}
|
||||
}
|
||||
for (int32 c = 0; c < num_splice; c++) {
|
||||
int32 dim = input_dim - const_dim; // dimension we
|
||||
// are splicing
|
||||
CuSubMatrix<BaseFloat> in_part(in, 0, in.NumRows(),
|
||||
0, dim),
|
||||
out_part(*out, 0, out->NumRows(),
|
||||
c * dim, dim);
|
||||
out_part.CopyRows(in_part, indexes[c]);
|
||||
}
|
||||
if (const_dim != 0) {
|
||||
CuSubMatrix<BaseFloat> in_part(in, 0, in.NumRows(),
|
||||
in.NumCols() - const_dim, const_dim),
|
||||
out_part(*out, 0, out->NumRows(),
|
||||
out->NumCols() - const_dim, const_dim);
|
||||
out_part.CopyRows(in_part, const_indexes);
|
||||
}
|
||||
}
|
||||
|
||||
void SpliceComponent::Backprop(const CuMatrixBase<BaseFloat> &, // in_value
|
||||
|
|
@ -3537,111 +3532,69 @@ void SpliceComponent::Backprop(const CuMatrixBase<BaseFloat> &, // in_value
|
|||
|
||||
KALDI_ASSERT( OutputDim() == output_dim );
|
||||
|
||||
if (0) { // old code
|
||||
in_deriv->Resize(num_chunks * input_chunk_size, input_dim); // Will zero it.
|
||||
for (int32 chunk = 0; chunk < num_chunks; chunk++) {
|
||||
CuSubMatrix<BaseFloat> in_deriv_chunk(*in_deriv,
|
||||
chunk * input_chunk_size, input_chunk_size,
|
||||
0, input_dim),
|
||||
out_deriv_chunk(out_deriv,
|
||||
chunk * output_chunk_size, output_chunk_size,
|
||||
0, output_dim);
|
||||
in_deriv->Resize(num_chunks * input_chunk_size, input_dim, kUndefined);
|
||||
|
||||
|
||||
for (int32 c = 0; c < left_context_ + right_context_ + 1; c++) {
|
||||
CuSubMatrix<BaseFloat> in_deriv_part(in_deriv_chunk,
|
||||
c, output_chunk_size,
|
||||
0, input_dim - const_component_dim_),
|
||||
out_deriv_part(out_deriv_chunk,
|
||||
0, output_chunk_size,
|
||||
c * (input_dim - const_component_dim_),
|
||||
input_dim - const_component_dim_);
|
||||
in_deriv_part.AddMat(1.0, out_deriv_part);
|
||||
}
|
||||
|
||||
if (const_component_dim_ > 0) {
|
||||
CuSubMatrix<BaseFloat> out_deriv_const_part(out_deriv_chunk,
|
||||
0, output_chunk_size,
|
||||
output_dim - const_component_dim_,
|
||||
const_component_dim_);
|
||||
// Because we assume the "constant part" of the input is the same for all
|
||||
// input rows, it's not clear how to propagate the derivative back. We
|
||||
// propagate the same value to all copies of it, but you should only take
|
||||
// one of them, not sum them up. In practice this is only used at the
|
||||
// start of the network and the derivative probably won't ever be used.
|
||||
for (int32 c = 0; c < in_deriv_chunk.NumRows(); c++) {
|
||||
CuSubMatrix<BaseFloat> in_deriv_part(in_deriv_chunk, c, 1,
|
||||
input_dim - const_component_dim_,
|
||||
const_component_dim_);
|
||||
in_deriv_part.Row(0).AddRowSumMat(1.0, out_deriv_const_part);
|
||||
}
|
||||
}
|
||||
}
|
||||
} else {
|
||||
in_deriv->Resize(num_chunks * input_chunk_size, input_dim, kUndefined);
|
||||
|
||||
int32 num_splice = left_context_ + right_context_ + 1,
|
||||
const_dim = const_component_dim_;
|
||||
// 'indexes' is, for each index from 0 to num_splice - 1,
|
||||
// then for each row of "in_deriv", the corresponding row of "out_deriv" that
|
||||
// we add, or -1 if.
|
||||
int32 num_splice = left_context_ + right_context_ + 1,
|
||||
const_dim = const_component_dim_;
|
||||
// 'indexes' is, for each index from 0 to num_splice - 1,
|
||||
// then for each row of "in_deriv", the corresponding row of "out_deriv" that
|
||||
// we add, or -1 if.
|
||||
|
||||
std::vector<std::vector<int32> > indexes(num_splice);
|
||||
// const_dim != 0, "const_indexes" will be used to determine which
|
||||
// row of "in" we copy the last part of each row of "out" from (this part is
|
||||
// not subject to splicing, it's assumed constant for each frame of "input".
|
||||
std::vector<int32> const_indexes(const_dim == 0 ? 0 : in_deriv->NumRows(),
|
||||
-1);
|
||||
std::vector<std::vector<int32> > indexes(num_splice);
|
||||
// const_dim != 0, "const_indexes" will be used to determine which
|
||||
// row of "in" we copy the last part of each row of "out" from (this part is
|
||||
// not subject to splicing, it's assumed constant for each frame of "input".
|
||||
std::vector<int32> const_indexes(const_dim == 0 ? 0 : in_deriv->NumRows(),
|
||||
-1);
|
||||
|
||||
for (int32 c = 0; c < indexes.size(); c++)
|
||||
indexes[c].resize(in_deriv->NumRows(), -1); // set to -1 by default,
|
||||
// this gets interpreted by the CopyRows() code as a signal to zero the output...
|
||||
for (int32 c = 0; c < indexes.size(); c++)
|
||||
indexes[c].resize(in_deriv->NumRows(), -1); // set to -1 by default,
|
||||
// this gets interpreted by the CopyRows() code as a signal to zero the output...
|
||||
|
||||
int32 dim = input_dim - const_dim; // dimension we are splicing
|
||||
int32 dim = input_dim - const_dim; // dimension we are splicing
|
||||
|
||||
for (int32 chunk = 0; chunk < num_chunks; chunk++) {
|
||||
for (int32 c = 0; c < num_splice; c++)
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++)
|
||||
indexes[c][chunk * input_chunk_size + c + offset] =
|
||||
chunk * output_chunk_size + offset;
|
||||
for (int32 chunk = 0; chunk < num_chunks; chunk++) {
|
||||
for (int32 c = 0; c < num_splice; c++)
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++)
|
||||
indexes[c][chunk * input_chunk_size + c + offset] =
|
||||
chunk * output_chunk_size + offset;
|
||||
|
||||
// Note: when changing over to the CUDA code, we also changed
|
||||
// how the derivatives are propagated through the splicing layer
|
||||
// for the const-component-dim. The code was never being used,
|
||||
// so it doesn't matter. The way we now do it probably makes more
|
||||
// sense (to get the derivative, you'd have to sum over time, not
|
||||
// pick an arbitrary time)
|
||||
if (const_dim != 0)
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++)
|
||||
const_indexes[chunk * input_chunk_size + offset] =
|
||||
chunk * output_chunk_size + offset;
|
||||
}
|
||||
// Note: when changing over to the CUDA code, we also changed
|
||||
// how the derivatives are propagated through the splicing layer
|
||||
// for the const-component-dim. The code was never being used,
|
||||
// so it doesn't matter. The way we now do it probably makes more
|
||||
// sense (to get the derivative, you'd have to sum over time, not
|
||||
// pick an arbitrary time)
|
||||
if (const_dim != 0)
|
||||
for (int32 offset = 0; offset < output_chunk_size; offset++)
|
||||
const_indexes[chunk * input_chunk_size + offset] =
|
||||
chunk * output_chunk_size + offset;
|
||||
}
|
||||
|
||||
CuMatrix<BaseFloat> temp_mat(in_deriv->NumRows(), dim, kUndefined);
|
||||
CuMatrix<BaseFloat> temp_mat(in_deriv->NumRows(), dim, kUndefined);
|
||||
|
||||
for (int32 c = 0; c < num_splice; c++) {
|
||||
int32 dim = input_dim - const_dim; // dimension we
|
||||
// are splicing
|
||||
CuSubMatrix<BaseFloat> out_deriv_part(out_deriv, 0, out_deriv.NumRows(),
|
||||
c * dim, dim),
|
||||
in_deriv_part(*in_deriv, 0, in_deriv->NumRows(),
|
||||
0, dim);
|
||||
if (c == 0)
|
||||
in_deriv_part.CopyRows(out_deriv_part, indexes[c]);
|
||||
else {
|
||||
temp_mat.CopyRows(out_deriv_part, indexes[c]);
|
||||
in_deriv_part.AddMat(1.0, temp_mat);
|
||||
}
|
||||
}
|
||||
if (const_dim != 0) {
|
||||
CuSubMatrix<BaseFloat> out_deriv_part(out_deriv, 0, out_deriv.NumRows(),
|
||||
out_deriv.NumCols() - const_dim,
|
||||
const_dim),
|
||||
in_deriv_part(*in_deriv, 0, in_deriv->NumRows(),
|
||||
in_deriv->NumCols() - const_dim, const_dim);
|
||||
in_deriv_part.CopyRows(out_deriv_part, const_indexes);
|
||||
for (int32 c = 0; c < num_splice; c++) {
|
||||
int32 dim = input_dim - const_dim; // dimension we
|
||||
// are splicing
|
||||
CuSubMatrix<BaseFloat> out_deriv_part(out_deriv, 0, out_deriv.NumRows(),
|
||||
c * dim, dim),
|
||||
in_deriv_part(*in_deriv, 0, in_deriv->NumRows(),
|
||||
0, dim);
|
||||
if (c == 0)
|
||||
in_deriv_part.CopyRows(out_deriv_part, indexes[c]);
|
||||
else {
|
||||
temp_mat.CopyRows(out_deriv_part, indexes[c]);
|
||||
in_deriv_part.AddMat(1.0, temp_mat);
|
||||
}
|
||||
}
|
||||
if (const_dim != 0) {
|
||||
CuSubMatrix<BaseFloat> out_deriv_part(out_deriv, 0, out_deriv.NumRows(),
|
||||
out_deriv.NumCols() - const_dim,
|
||||
const_dim),
|
||||
in_deriv_part(*in_deriv, 0, in_deriv->NumRows(),
|
||||
in_deriv->NumCols() - const_dim, const_dim);
|
||||
in_deriv_part.CopyRows(out_deriv_part, const_indexes);
|
||||
}
|
||||
}
|
||||
|
||||
Component *SpliceComponent::Copy() const {
|
||||
|
|
@ -4159,6 +4112,142 @@ void FixedAffineComponent::Read(std::istream &is, bool binary) {
|
|||
}
|
||||
|
||||
|
||||
void FixedScaleComponent::Init(const CuVectorBase<BaseFloat> &scales) {
|
||||
KALDI_ASSERT(scales.Dim() != 0);
|
||||
scales_ = scales;
|
||||
}
|
||||
|
||||
void FixedScaleComponent::InitFromString(std::string args) {
|
||||
std::string orig_args = args;
|
||||
std::string filename;
|
||||
bool ok = ParseFromString("scales", &args, &filename);
|
||||
|
||||
if (!ok || !args.empty())
|
||||
KALDI_ERR << "Invalid initializer for layer of type "
|
||||
<< Type() << ": \"" << orig_args << "\"";
|
||||
|
||||
CuVector<BaseFloat> vec;
|
||||
ReadKaldiObject(filename, &vec);
|
||||
Init(vec);
|
||||
}
|
||||
|
||||
|
||||
std::string FixedScaleComponent::Info() const {
|
||||
std::stringstream stream;
|
||||
BaseFloat scales_size = static_cast<BaseFloat>(scales_.Dim()),
|
||||
scales_mean = scales_.Sum() / scales_size,
|
||||
scales_stddev = std::sqrt(VecVec(scales_, scales_) / scales_size)
|
||||
- (scales_mean * scales_mean);
|
||||
stream << Component::Info() << ", scales-mean=" << scales_mean
|
||||
<< ", scales-stddev=" << scales_stddev;
|
||||
return stream.str();
|
||||
}
|
||||
|
||||
void FixedScaleComponent::Propagate(const CuMatrixBase<BaseFloat> &in,
|
||||
int32 num_chunks,
|
||||
CuMatrix<BaseFloat> *out) const {
|
||||
*out = in;
|
||||
out->MulColsVec(scales_);
|
||||
}
|
||||
|
||||
void FixedScaleComponent::Backprop(const CuMatrixBase<BaseFloat> &, // in_value
|
||||
const CuMatrixBase<BaseFloat> &, // out_value
|
||||
const CuMatrixBase<BaseFloat> &out_deriv,
|
||||
int32, // num_chunks
|
||||
Component *, // to_update
|
||||
CuMatrix<BaseFloat> *in_deriv) const {
|
||||
*in_deriv = out_deriv;
|
||||
in_deriv->MulColsVec(scales_);
|
||||
}
|
||||
|
||||
Component* FixedScaleComponent::Copy() const {
|
||||
FixedScaleComponent *ans = new FixedScaleComponent();
|
||||
ans->scales_ = scales_;
|
||||
return ans;
|
||||
}
|
||||
|
||||
|
||||
void FixedScaleComponent::Write(std::ostream &os, bool binary) const {
|
||||
WriteToken(os, binary, "<FixedScaleComponent>");
|
||||
WriteToken(os, binary, "<Scales>");
|
||||
scales_.Write(os, binary);
|
||||
WriteToken(os, binary, "</FixedScaleComponent>");
|
||||
}
|
||||
|
||||
void FixedScaleComponent::Read(std::istream &is, bool binary) {
|
||||
ExpectOneOrTwoTokens(is, binary, "<FixedScaleComponent>", "<Scales>");
|
||||
scales_.Read(is, binary);
|
||||
ExpectToken(is, binary, "</FixedScaleComponent>");
|
||||
}
|
||||
|
||||
void FixedBiasComponent::Init(const CuVectorBase<BaseFloat> &bias) {
|
||||
KALDI_ASSERT(bias.Dim() != 0);
|
||||
bias_ = bias;
|
||||
}
|
||||
|
||||
void FixedBiasComponent::InitFromString(std::string args) {
|
||||
std::string orig_args = args;
|
||||
std::string filename;
|
||||
bool ok = ParseFromString("bias", &args, &filename);
|
||||
|
||||
if (!ok || !args.empty())
|
||||
KALDI_ERR << "Invalid initializer for layer of type "
|
||||
<< Type() << ": \"" << orig_args << "\"";
|
||||
|
||||
CuVector<BaseFloat> vec;
|
||||
ReadKaldiObject(filename, &vec);
|
||||
Init(vec);
|
||||
}
|
||||
|
||||
|
||||
std::string FixedBiasComponent::Info() const {
|
||||
std::stringstream stream;
|
||||
BaseFloat bias_size = static_cast<BaseFloat>(bias_.Dim()),
|
||||
bias_mean = bias_.Sum() / bias_size,
|
||||
bias_stddev = std::sqrt(VecVec(bias_, bias_) / bias_size)
|
||||
- (bias_mean * bias_mean);
|
||||
stream << Component::Info() << ", bias-mean=" << bias_mean
|
||||
<< ", bias-stddev=" << bias_stddev;
|
||||
return stream.str();
|
||||
}
|
||||
|
||||
void FixedBiasComponent::Propagate(const CuMatrixBase<BaseFloat> &in,
|
||||
int32 num_chunks,
|
||||
CuMatrix<BaseFloat> *out) const {
|
||||
*out = in;
|
||||
out->AddVecToRows(1.0, bias_, 1.0);
|
||||
}
|
||||
|
||||
void FixedBiasComponent::Backprop(const CuMatrixBase<BaseFloat> &, // in_value
|
||||
const CuMatrixBase<BaseFloat> &, // out_value
|
||||
const CuMatrixBase<BaseFloat> &out_deriv,
|
||||
int32, // num_chunks
|
||||
Component *, // to_update
|
||||
CuMatrix<BaseFloat> *in_deriv) const {
|
||||
*in_deriv = out_deriv;
|
||||
}
|
||||
|
||||
Component* FixedBiasComponent::Copy() const {
|
||||
FixedBiasComponent *ans = new FixedBiasComponent();
|
||||
ans->bias_ = bias_;
|
||||
return ans;
|
||||
}
|
||||
|
||||
|
||||
void FixedBiasComponent::Write(std::ostream &os, bool binary) const {
|
||||
WriteToken(os, binary, "<FixedBiasComponent>");
|
||||
WriteToken(os, binary, "<Bias>");
|
||||
bias_.Write(os, binary);
|
||||
WriteToken(os, binary, "</FixedBiasComponent>");
|
||||
}
|
||||
|
||||
void FixedBiasComponent::Read(std::istream &is, bool binary) {
|
||||
ExpectOneOrTwoTokens(is, binary, "<FixedBiasComponent>", "<Bias>");
|
||||
bias_.Read(is, binary);
|
||||
ExpectToken(is, binary, "</FixedBiasComponent>");
|
||||
}
|
||||
|
||||
|
||||
|
||||
|
||||
std::string DropoutComponent::Info() const {
|
||||
|
|
|
|||
|
|
@ -638,6 +638,11 @@ class AffineComponent: public UpdatableComponent {
|
|||
friend class SoftmaxComponent; // Friend declaration relates to mixing up.
|
||||
public:
|
||||
explicit AffineComponent(const AffineComponent &other);
|
||||
// The next constructor is used in converting from nnet1.
|
||||
AffineComponent(const CuMatrix<BaseFloat> &linear_params,
|
||||
const CuVector<BaseFloat> &bias_params,
|
||||
BaseFloat learning_rate);
|
||||
|
||||
virtual int32 InputDim() const { return linear_params_.NumCols(); }
|
||||
virtual int32 OutputDim() const { return linear_params_.NumRows(); }
|
||||
void Init(BaseFloat learning_rate,
|
||||
|
|
@ -1153,6 +1158,7 @@ class SpliceComponent: public Component {
|
|||
};
|
||||
|
||||
|
||||
|
||||
/// This is as SpliceComponent but outputs the max of
|
||||
/// any of the inputs (taking the max across time).
|
||||
class SpliceMaxComponent: public Component {
|
||||
|
|
@ -1442,12 +1448,16 @@ private:
|
|||
};
|
||||
|
||||
|
||||
/// PermuteComponent does a random permutation of the dimensions. Useful in
|
||||
/// conjunction with block-diagonal transforms.
|
||||
/// PermuteComponent does a permutation of the dimensions (by default, a fixed
|
||||
/// random permutation, but it may be specified). Useful in conjunction with
|
||||
/// block-diagonal transforms.
|
||||
class PermuteComponent: public Component {
|
||||
public:
|
||||
void Init(int32 dim);
|
||||
void Init(const std::vector<int32> &reorder);
|
||||
PermuteComponent(int32 dim) { Init(dim); }
|
||||
PermuteComponent(const std::vector<int32> &reorder) { Init(reorder); }
|
||||
|
||||
PermuteComponent() { } // e.g. prior to Read() or Init()
|
||||
|
||||
virtual int32 InputDim() const { return reorder_.size(); }
|
||||
|
|
@ -1463,17 +1473,17 @@ class PermuteComponent: public Component {
|
|||
virtual void Propagate(const CuMatrixBase<BaseFloat> &in,
|
||||
int32 num_chunks,
|
||||
CuMatrix<BaseFloat> *out) const;
|
||||
virtual void Backprop(const CuMatrixBase<BaseFloat> &in_value, // dummy
|
||||
const CuMatrixBase<BaseFloat> &out_value, // dummy
|
||||
virtual void Backprop(const CuMatrixBase<BaseFloat> &,
|
||||
const CuMatrixBase<BaseFloat> &,
|
||||
const CuMatrixBase<BaseFloat> &out_deriv,
|
||||
int32 num_chunks,
|
||||
Component *to_update, // dummy
|
||||
Component *,
|
||||
CuMatrix<BaseFloat> *in_deriv) const;
|
||||
|
||||
private:
|
||||
KALDI_DISALLOW_COPY_AND_ASSIGN(PermuteComponent);
|
||||
std::vector<int32> reorder_; // This class sends input dimension i to
|
||||
// output dimension reorder_[i].
|
||||
// output dimension reorder_[i].
|
||||
};
|
||||
|
||||
|
||||
|
|
@ -1607,6 +1617,81 @@ class FixedAffineComponent: public Component {
|
|||
};
|
||||
|
||||
|
||||
/// FixedScaleComponent applies a fixed per-element scale; it's similar
|
||||
/// to the Rescale component in the nnet1 setup (and only needed for nnet1
|
||||
/// model conversion.
|
||||
class FixedScaleComponent: public Component {
|
||||
public:
|
||||
FixedScaleComponent() { }
|
||||
virtual std::string Type() const { return "FixedScaleComponent"; }
|
||||
virtual std::string Info() const;
|
||||
|
||||
void Init(const CuVectorBase<BaseFloat> &scales);
|
||||
|
||||
// InitFromString takes only the option scales=<string>,
|
||||
// where the string is the filename of a Kaldi-format matrix to read.
|
||||
virtual void InitFromString(std::string args);
|
||||
|
||||
virtual int32 InputDim() const { return scales_.Dim(); }
|
||||
virtual int32 OutputDim() const { return scales_.Dim(); }
|
||||
virtual void Propagate(const CuMatrixBase<BaseFloat> &in,
|
||||
int32 num_chunks,
|
||||
CuMatrix<BaseFloat> *out) const;
|
||||
virtual void Backprop(const CuMatrixBase<BaseFloat> &in_value,
|
||||
const CuMatrixBase<BaseFloat> &out_value,
|
||||
const CuMatrixBase<BaseFloat> &out_deriv,
|
||||
int32 num_chunks,
|
||||
Component *to_update, // may be identical to "this".
|
||||
CuMatrix<BaseFloat> *in_deriv) const;
|
||||
virtual bool BackpropNeedsInput() const { return false; }
|
||||
virtual bool BackpropNeedsOutput() const { return false; }
|
||||
virtual Component* Copy() const;
|
||||
virtual void Read(std::istream &is, bool binary);
|
||||
virtual void Write(std::ostream &os, bool binary) const;
|
||||
|
||||
protected:
|
||||
CuVector<BaseFloat> scales_;
|
||||
KALDI_DISALLOW_COPY_AND_ASSIGN(FixedScaleComponent);
|
||||
};
|
||||
|
||||
/// FixedBiasComponent applies a fixed per-element bias; it's similar
|
||||
/// to the AddShift component in the nnet1 setup (and only needed for nnet1
|
||||
/// model conversion.
|
||||
class FixedBiasComponent: public Component {
|
||||
public:
|
||||
FixedBiasComponent() { }
|
||||
virtual std::string Type() const { return "FixedBiasComponent"; }
|
||||
virtual std::string Info() const;
|
||||
|
||||
void Init(const CuVectorBase<BaseFloat> &scales);
|
||||
|
||||
// InitFromString takes only the option bias=<string>,
|
||||
// where the string is the filename of a Kaldi-format matrix to read.
|
||||
virtual void InitFromString(std::string args);
|
||||
|
||||
virtual int32 InputDim() const { return bias_.Dim(); }
|
||||
virtual int32 OutputDim() const { return bias_.Dim(); }
|
||||
virtual void Propagate(const CuMatrixBase<BaseFloat> &in,
|
||||
int32 num_chunks,
|
||||
CuMatrix<BaseFloat> *out) const;
|
||||
virtual void Backprop(const CuMatrixBase<BaseFloat> &in_value,
|
||||
const CuMatrixBase<BaseFloat> &out_value,
|
||||
const CuMatrixBase<BaseFloat> &out_deriv,
|
||||
int32 num_chunks,
|
||||
Component *to_update, // may be identical to "this".
|
||||
CuMatrix<BaseFloat> *in_deriv) const;
|
||||
virtual bool BackpropNeedsInput() const { return false; }
|
||||
virtual bool BackpropNeedsOutput() const { return false; }
|
||||
virtual Component* Copy() const;
|
||||
virtual void Read(std::istream &is, bool binary);
|
||||
virtual void Write(std::ostream &os, bool binary) const;
|
||||
|
||||
protected:
|
||||
CuVector<BaseFloat> bias_;
|
||||
KALDI_DISALLOW_COPY_AND_ASSIGN(FixedBiasComponent);
|
||||
};
|
||||
|
||||
|
||||
/// This Component, if present, randomly zeroes half of
|
||||
/// the inputs and multiplies the other half by two.
|
||||
/// Typically you would use this in training but not in
|
||||
|
|
|
|||
|
|
@ -661,7 +661,6 @@ void Nnet::Collapse(bool match_updatableness) {
|
|||
KALDI_LOG << "Collapsed " << num_collapsed << " components.";
|
||||
}
|
||||
|
||||
|
||||
Nnet *GenRandomNnet(int32 input_dim,
|
||||
int32 output_dim) {
|
||||
|
||||
|
|
@ -711,6 +710,14 @@ Nnet *GenRandomNnet(int32 input_dim,
|
|||
|
||||
|
||||
|
||||
int32 Nnet::LastUpdatableComponent() const {
|
||||
int32 last_updatable_component = NumComponents();
|
||||
for (int32 i = NumComponents() - 1; i >= 0; i--)
|
||||
if (dynamic_cast<UpdatableComponent*>(components_[i]) != NULL)
|
||||
last_updatable_component = i;
|
||||
return last_updatable_component;
|
||||
}
|
||||
|
||||
} // namespace nnet2
|
||||
} // namespace kaldi
|
||||
|
||||
|
|
|
|||
|
|
@ -98,6 +98,12 @@ class Nnet {
|
|||
/// this neural net, leaving everything else fixed.
|
||||
void CopyStatsFrom(const Nnet &nnet);
|
||||
|
||||
|
||||
/// Returns the index of the last component which is updatable,
|
||||
/// or NumComponents() if none are updatable.
|
||||
int32 LastUpdatableComponent() const;
|
||||
|
||||
/// Returns the number of updatable components.
|
||||
int32 NumUpdatableComponents() const;
|
||||
|
||||
/// Scales the parameters of each of the updatable components.
|
||||
|
|
@ -192,9 +198,10 @@ class Nnet {
|
|||
/// AffineComponent learning-rate=0.01 l2-penalty=0.001 input-dim=10 output-dim=15 param-stddev=0.1
|
||||
void Init(std::istream &is);
|
||||
|
||||
/// This Init method works from a vector of components. It will take ownership
|
||||
/// of the pointers and resize the vector to zero to avoid a chance of the
|
||||
/// caller deallocating them.
|
||||
/// This Init method works from a vector of components. It will take
|
||||
/// ownership of the pointers and will resize the vector to zero to avoid a
|
||||
/// chance of the caller deallocating them (but the vector itself is not
|
||||
/// deleted).
|
||||
void Init(std::vector<Component*> *components);
|
||||
|
||||
/// Appends this component to the components already in the neural net.
|
||||
|
|
|
|||
|
|
@ -135,7 +135,8 @@ double NnetUpdater::ComputeTotAccuracy(
|
|||
|
||||
void NnetUpdater::Backprop(CuMatrix<BaseFloat> *deriv) const {
|
||||
// We assume ComputeObjfAndDeriv has already been called.
|
||||
for (int32 c = nnet_.NumComponents() - 1; c >= 0; c--) {
|
||||
for (int32 c = nnet_.NumComponents() - 1;
|
||||
c >= nnet_.LastUpdatableComponent(); c--) {
|
||||
const Component &component = nnet_.GetComponent(c);
|
||||
Component *component_to_update = (nnet_to_update_ == NULL ? NULL :
|
||||
&(nnet_to_update_->GetComponent(c)));
|
||||
|
|
|
|||
|
|
@ -26,15 +26,18 @@ BINFILES = nnet-randomize-frames nnet-am-info nnet-init \
|
|||
nnet-modify-learning-rates nnet-normalize-stddev nnet-perturb-egs \
|
||||
nnet-perturb-egs-fmllr nnet-get-weighted-egs nnet-adjust-priors \
|
||||
cuda-compiled nnet-replace-last-layers nnet-am-switch-preconditioning \
|
||||
nnet-train-simple-perturbed nnet-train-parallel-perturbed
|
||||
nnet-train-simple-perturbed nnet-train-parallel-perturbed \
|
||||
nnet1-to-raw-nnet
|
||||
|
||||
OBJFILES =
|
||||
|
||||
# Add this dependency to force cuda-compiled.o to be rebuilt when we reconfigure.
|
||||
cuda-compiled.o: ../kaldi.mk
|
||||
|
||||
|
||||
TESTFILES =
|
||||
|
||||
ADDLIBS = ../nnet2/kaldi-nnet2.a ../gmm/kaldi-gmm.a \
|
||||
ADDLIBS = ../nnet2/kaldi-nnet2.a ../nnet/kaldi-nnet.a ../gmm/kaldi-gmm.a \
|
||||
../decoder/kaldi-decoder.a ../lat/kaldi-lat.a ../hmm/kaldi-hmm.a \
|
||||
../transform/kaldi-transform.a ../tree/kaldi-tree.a ../thread/kaldi-thread.a \
|
||||
../cudamatrix/kaldi-cudamatrix.a ../matrix/kaldi-matrix.a \
|
||||
|
|
|
|||
|
|
@ -48,16 +48,20 @@ BaseFloat KlDivergence(const Vector<BaseFloat> &p,
|
|||
|
||||
void PrintPriorDiagnostics(const Vector<BaseFloat> &old_priors,
|
||||
const Vector<BaseFloat> &new_priors) {
|
||||
Vector<BaseFloat> diff_prior(new_priors);
|
||||
diff_prior.AddVec(-1.0, old_priors);
|
||||
diff_prior.ApplyAbs();
|
||||
int32 max_index;
|
||||
diff_prior.Max(&max_index);
|
||||
KALDI_LOG << "Adjusting priors: largest absolute difference was for "
|
||||
<< "pdf " << max_index << ", " << old_priors(max_index)
|
||||
<< " -> " << new_priors(max_index);
|
||||
KALDI_LOG << "Adjusting priors: K-L divergence from old to new is "
|
||||
<< KlDivergence(old_priors, new_priors);
|
||||
if (old_priors.Dim() == 0) {
|
||||
KALDI_LOG << "Model did not previously have priors attached.";
|
||||
} else {
|
||||
Vector<BaseFloat> diff_prior(new_priors);
|
||||
diff_prior.AddVec(-1.0, old_priors);
|
||||
diff_prior.ApplyAbs();
|
||||
int32 max_index;
|
||||
diff_prior.Max(&max_index);
|
||||
KALDI_LOG << "Adjusting priors: largest absolute difference was for "
|
||||
<< "pdf " << max_index << ", " << old_priors(max_index)
|
||||
<< " -> " << new_priors(max_index);
|
||||
KALDI_LOG << "Adjusting priors: K-L divergence from old to new is "
|
||||
<< KlDivergence(old_priors, new_priors);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -38,6 +38,7 @@ int main(int argc, char *argv[]) {
|
|||
"See example scripts to see how this works in practice.\n"
|
||||
"\n"
|
||||
"Usage: nnet-am-init [options] <tree-in> <topology-in> <raw-nnet-in> <nnet-am-out>\n"
|
||||
"or: nnet-am-init [options] <transition-model-in> <raw-nnet-in> <nnet-am-out>\n"
|
||||
"e.g.:\n"
|
||||
" nnet-am-init tree topo \"nnet-init nnet.config - |\" 1.mdl\n";
|
||||
|
||||
|
|
@ -48,25 +49,37 @@ int main(int argc, char *argv[]) {
|
|||
|
||||
po.Read(argc, argv);
|
||||
|
||||
if (po.NumArgs() != 4) {
|
||||
if (po.NumArgs() != 3 && po.NumArgs() != 4) {
|
||||
po.PrintUsage();
|
||||
exit(1);
|
||||
}
|
||||
|
||||
std::string tree_rxfilename = po.GetArg(1),
|
||||
topo_rxfilename = po.GetArg(2),
|
||||
raw_nnet_rxfilename = po.GetArg(3),
|
||||
nnet_wxfilename = po.GetArg(4);
|
||||
std::string raw_nnet_rxfilename, nnet_wxfilename;
|
||||
|
||||
ContextDependency ctx_dep;
|
||||
ReadKaldiObject(tree_rxfilename, &ctx_dep);
|
||||
TransitionModel *trans_model = NULL;
|
||||
|
||||
if (po.NumArgs() == 4) {
|
||||
std::string tree_rxfilename = po.GetArg(1),
|
||||
topo_rxfilename = po.GetArg(2);
|
||||
raw_nnet_rxfilename = po.GetArg(3);
|
||||
nnet_wxfilename = po.GetArg(4);
|
||||
|
||||
HmmTopology topo;
|
||||
ReadKaldiObject(topo_rxfilename, &topo);
|
||||
|
||||
// Construct the transition model from the tree and the topology file.
|
||||
TransitionModel trans_model(ctx_dep, topo);
|
||||
ContextDependency ctx_dep;
|
||||
ReadKaldiObject(tree_rxfilename, &ctx_dep);
|
||||
|
||||
HmmTopology topo;
|
||||
ReadKaldiObject(topo_rxfilename, &topo);
|
||||
|
||||
// Construct the transition model from the tree and the topology file.
|
||||
trans_model = new TransitionModel(ctx_dep, topo);
|
||||
} else {
|
||||
std::string trans_model_rxfilename = po.GetArg(1);
|
||||
raw_nnet_rxfilename = po.GetArg(2);
|
||||
nnet_wxfilename = po.GetArg(3);
|
||||
trans_model = new TransitionModel();
|
||||
ReadKaldiObject(trans_model_rxfilename, trans_model);
|
||||
}
|
||||
|
||||
AmNnet am_nnet;
|
||||
{
|
||||
Nnet nnet;
|
||||
|
|
@ -76,16 +89,17 @@ int main(int argc, char *argv[]) {
|
|||
am_nnet.Init(nnet);
|
||||
}
|
||||
|
||||
if (am_nnet.NumPdfs() != trans_model.NumPdfs())
|
||||
if (am_nnet.NumPdfs() != trans_model->NumPdfs())
|
||||
KALDI_ERR << "Mismatch in number of pdfs, neural net has "
|
||||
<< am_nnet.NumPdfs() << ", transition model has "
|
||||
<< trans_model.NumPdfs();
|
||||
<< trans_model->NumPdfs();
|
||||
|
||||
{
|
||||
Output ko(nnet_wxfilename, binary_write);
|
||||
trans_model.Write(ko.Stream(), binary_write);
|
||||
trans_model->Write(ko.Stream(), binary_write);
|
||||
am_nnet.Write(ko.Stream(), binary_write);
|
||||
}
|
||||
delete trans_model;
|
||||
KALDI_LOG << "Initialized neural net and wrote it to " << nnet_wxfilename;
|
||||
return 0;
|
||||
} catch(const std::exception &e) {
|
||||
|
|
|
|||
|
|
@ -35,7 +35,7 @@ int main(int argc, char *argv[]) {
|
|||
"Initialize the neural network from a config file with a line for each\n"
|
||||
"component. Note, this only outputs the neural net itself, not the associated\n"
|
||||
"information such as the transition-model; you'll probably want to pipe\n"
|
||||
"the output into something like am-nnet-init.\n"
|
||||
"the output into something like nnet-am-init.\n"
|
||||
"\n"
|
||||
"Usage: nnet-init [options] <config-in> <raw-nnet-out>\n"
|
||||
"e.g.:\n"
|
||||
|
|
|
|||
|
|
@ -0,0 +1,203 @@
|
|||
// nnet2bin/nnet1-to-raw-nnet.cc
|
||||
|
||||
// Copyright 2013 Johns Hopkins University (author: Daniel Povey, Hainan Xu)
|
||||
|
||||
// See ../../COPYING for clarification regarding multiple authors
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
||||
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
|
||||
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
|
||||
// MERCHANTABLITY OR NON-INFRINGEMENT.
|
||||
// See the Apache 2 License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
#include "base/kaldi-common.h"
|
||||
#include "util/common-utils.h"
|
||||
#include "hmm/transition-model.h"
|
||||
#include "nnet/nnet-nnet.h"
|
||||
#include "nnet/nnet-affine-transform.h"
|
||||
#include "nnet/nnet-activation.h"
|
||||
#include "nnet2/nnet-nnet.h"
|
||||
#include "nnet2/nnet-component.h"
|
||||
|
||||
namespace kaldi {
|
||||
|
||||
nnet2::Component *ConvertAffineTransformComponent(
|
||||
const nnet1::Component &nnet1_component) {
|
||||
const nnet1::AffineTransform *affine =
|
||||
dynamic_cast<const nnet1::AffineTransform*>(&nnet1_component);
|
||||
KALDI_ASSERT(affine != NULL);
|
||||
// default learning rate is 1.0e-05, you can use the --learning-rate or
|
||||
// --learning-rates option to nnet-am-copy to change it if you need.
|
||||
BaseFloat learning_rate = 1.0e-05;
|
||||
return new nnet2::AffineComponent(affine->GetLinearity(),
|
||||
affine->GetBias(),
|
||||
learning_rate);
|
||||
}
|
||||
|
||||
nnet2::Component *ConvertSoftmaxComponent(
|
||||
const nnet1::Component &nnet1_component) {
|
||||
const nnet1::Softmax *softmax =
|
||||
dynamic_cast<const nnet1::Softmax*>(&nnet1_component);
|
||||
KALDI_ASSERT(softmax != NULL);
|
||||
return new nnet2::SoftmaxComponent(softmax->InputDim());
|
||||
}
|
||||
|
||||
nnet2::Component *ConvertSigmoidComponent(
|
||||
const nnet1::Component &nnet1_component) {
|
||||
const nnet1::Sigmoid *sigmoid =
|
||||
dynamic_cast<const nnet1::Sigmoid*>(&nnet1_component);
|
||||
KALDI_ASSERT(sigmoid != NULL);
|
||||
return new nnet2::SigmoidComponent(sigmoid->InputDim());
|
||||
}
|
||||
|
||||
nnet2::Component *ConvertSpliceComponent(
|
||||
const nnet1::Component &nnet1_component) {
|
||||
const nnet1::Splice *splice =
|
||||
dynamic_cast<const nnet1::Splice*>(&nnet1_component);
|
||||
KALDI_ASSERT(splice != NULL);
|
||||
int32 low, high;
|
||||
std::vector<int32> frame_offsets;
|
||||
|
||||
std::ostringstream ostr;
|
||||
splice->WriteData(ostr, false);
|
||||
|
||||
std::istringstream istr(ostr.str());
|
||||
ReadIntegerVector(istr, false, &frame_offsets);
|
||||
|
||||
for (size_t i = 1; i < frame_offsets.size(); i++) {
|
||||
KALDI_ASSERT(frame_offsets[i-1] + 1 == frame_offsets[i]);
|
||||
}
|
||||
|
||||
low = frame_offsets[0];
|
||||
high = frame_offsets[frame_offsets.size() - 1];
|
||||
|
||||
nnet2::SpliceComponent *res = new nnet2::SpliceComponent();
|
||||
res->Init(splice->InputDim(), -low, high);
|
||||
return res;
|
||||
}
|
||||
|
||||
|
||||
nnet2::Component *ConvertAddShiftComponent(
|
||||
const nnet1::Component &nnet1_component) {
|
||||
const nnet1::AddShift *add_shift =
|
||||
dynamic_cast<const nnet1::AddShift*>(&nnet1_component);
|
||||
KALDI_ASSERT(add_shift != NULL);
|
||||
Vector<BaseFloat> bias;
|
||||
|
||||
add_shift->GetParams(&bias);
|
||||
CuVector<BaseFloat> cu_bias(bias);
|
||||
|
||||
nnet2::FixedBiasComponent *res = new nnet2::FixedBiasComponent();
|
||||
res->Init(cu_bias);
|
||||
return res;
|
||||
}
|
||||
|
||||
nnet2::Component *ConvertRescaleComponent(
|
||||
const nnet1::Component &nnet1_component) {
|
||||
const nnet1::Rescale *rescale =
|
||||
dynamic_cast<const nnet1::Rescale*>(&nnet1_component);
|
||||
KALDI_ASSERT(rescale != NULL);
|
||||
|
||||
Vector<BaseFloat> scale;
|
||||
rescale->GetParams(&scale);
|
||||
|
||||
CuVector<BaseFloat> cu_scale(scale);
|
||||
|
||||
nnet2::FixedScaleComponent *res = new nnet2::FixedScaleComponent();
|
||||
res->Init(cu_scale);
|
||||
return res;
|
||||
}
|
||||
|
||||
nnet2::Component *ConvertComponent(const nnet1::Component &nnet1_component) {
|
||||
nnet1::Component::ComponentType type_in = nnet1_component.GetType();
|
||||
switch (type_in) {
|
||||
case nnet1::Component::kAffineTransform:
|
||||
return ConvertAffineTransformComponent(nnet1_component);
|
||||
case nnet1::Component::kSoftmax:
|
||||
return ConvertSoftmaxComponent(nnet1_component);
|
||||
case nnet1::Component::kSigmoid:
|
||||
return ConvertSigmoidComponent(nnet1_component);
|
||||
case nnet1::Component::kSplice:
|
||||
return ConvertSpliceComponent(nnet1_component); // note, this will for now only handle the
|
||||
// special case nnet1::Component::where all splice indexes in nnet1_component are contiguous, e.g.
|
||||
// -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5 .
|
||||
case nnet1::Component::kAddShift:
|
||||
return ConvertAddShiftComponent(nnet1_component); // convert to FixedBiasComponent
|
||||
case nnet1::Component::kRescale:
|
||||
return ConvertRescaleComponent(nnet1_component); // convert to FixedScaleComponent
|
||||
default: KALDI_ERR << "Un-handled nnet1 component type "
|
||||
<< nnet1::Component::TypeToMarker(type_in);
|
||||
return NULL;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
nnet2::Nnet *ConvertNnet1ToNnet2(const nnet1::Nnet &nnet1) {
|
||||
// get a vector of nnet2::Component pointers and initialize the nnet2::Nnet with it.
|
||||
size_t size = nnet1.NumComponents();
|
||||
std::vector<nnet2::Component*> *components = new std::vector<nnet2::Component*>();
|
||||
components->resize(size);
|
||||
for (size_t i = 0; i < size; i++) {
|
||||
(*components)[i] = ConvertComponent(nnet1.GetComponent(i));
|
||||
}
|
||||
|
||||
nnet2::Nnet *res = new nnet2::Nnet();
|
||||
res->Init(components);
|
||||
// not de-allocate the memory for components
|
||||
// since the nnet takes the ownership
|
||||
return res;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
|
||||
int main(int argc, char *argv[]) {
|
||||
try {
|
||||
using namespace kaldi;
|
||||
typedef kaldi::int32 int32;
|
||||
|
||||
const char *usage =
|
||||
"Convert nnet1 neural net to nnet2 'raw' neural net\n"
|
||||
""
|
||||
"\n"
|
||||
"Usage: nnet1-to-raw-nnet [options] <nnet1-in> <nnet2-out>\n"
|
||||
"e.g.:\n"
|
||||
" nnet1-to-raw-nnet srcdir/final.nnet - | nnet-am-init dest/tree dest/topo - dest/0.mdl\n";
|
||||
|
||||
bool binary_write = true;
|
||||
int32 srand_seed = 0;
|
||||
|
||||
ParseOptions po(usage);
|
||||
po.Register("binary", &binary_write, "Write output in binary mode");
|
||||
|
||||
po.Read(argc, argv);
|
||||
srand(srand_seed);
|
||||
|
||||
if (po.NumArgs() != 2) {
|
||||
po.PrintUsage();
|
||||
exit(1);
|
||||
}
|
||||
|
||||
std::string nnet1_rxfilename = po.GetArg(1),
|
||||
raw_nnet2_wxfilename = po.GetArg(2);
|
||||
|
||||
nnet1::Nnet nnet1;
|
||||
ReadKaldiObject(nnet1_rxfilename, &nnet1);
|
||||
nnet2::Nnet *nnet2 = ConvertNnet1ToNnet2(nnet1);
|
||||
WriteKaldiObject(*nnet2, raw_nnet2_wxfilename, binary_write);
|
||||
KALDI_LOG << "Converted nnet1 neural net to raw nnet2 and wrote it to "
|
||||
<< PrintableWxfilename(raw_nnet2_wxfilename);
|
||||
delete nnet2;
|
||||
return 0;
|
||||
} catch(const std::exception &e) {
|
||||
std::cerr << e.what() << '\n';
|
||||
return -1;
|
||||
}
|
||||
}
|
||||
|
|
@ -34,7 +34,7 @@ int main(int argc, char *argv[]) {
|
|||
"Concatenate two 'raw' neural nets, e.g. as output by nnet-init or\n"
|
||||
"nnet-to-raw-nnet\n"
|
||||
"\n"
|
||||
"Usage: raw-nnet-concat [options] <raw-nnet1-in> <raw-nnet2-in> <raw-nnet-out>\n"
|
||||
"Usage: raw-nnet-concat [options] <raw-nnet-in1> <raw-nnet-in2> <raw-nnet-out>\n"
|
||||
"e.g.:\n"
|
||||
" raw-nnet-concat nnet1 nnet2 nnet_concat\n";
|
||||
|
||||
|
|
|
|||
|
|
@ -124,15 +124,11 @@ int main(int argc, char *argv[]) {
|
|||
<< ", " << mat.NumRows() << "frm";
|
||||
|
||||
//check for NaN/inf
|
||||
for (int32 r = 0; r<mat.NumRows(); r++) {
|
||||
for (int32 c = 0; c<mat.NumCols(); c++) {
|
||||
BaseFloat val = mat(r,c);
|
||||
if (val != val) KALDI_ERR << "NaN in features of : " << feature_reader.Key();
|
||||
if (val == std::numeric_limits<BaseFloat>::infinity())
|
||||
KALDI_ERR << "inf in features of : " << feature_reader.Key();
|
||||
}
|
||||
BaseFloat sum = mat.Sum();
|
||||
if (!KALDI_ISFINITE(sum)) {
|
||||
KALDI_ERR << "NaN or inf found in features of " << feature_reader.Key();
|
||||
}
|
||||
|
||||
|
||||
// push it to gpu
|
||||
feats = mat;
|
||||
// fwd-pass
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче