77. ./cmd.sh
88
99# general configuration
10- backend=chainer
11- stage=0 # start from 0 if you need to start from data preparation
12- gpu=-1 # use 0 when using GPU on slurm/grid engine, otherwise -1
10+ backend=pytorch
11+ stage=-1 # start from -1 if you need to start from data download
12+ gpu= # will be deprecated, please use ngpu
13+ ngpu=0 # number of gpus ("0" uses cpu, otherwise use gpu)
1314debugmode=1
1415dumpdir=dump # directory to dump full features
1516N=0 # number of minibatches to be used (mainly for debugging). "0" uses all minibatches.
@@ -21,8 +22,8 @@ do_delta=false # true when using CNN
2122
2223# network archtecture
2324# encoder related
24- etype=vggblstmp # encoder architecture type
25- elayers=6
25+ etype=blstmp # encoder architecture type
26+ elayers=8
2627eunits=320
2728eprojs=320
2829subsample=1_2_2_1_1 # skip every n frame from input to nth layers
@@ -38,7 +39,7 @@ aconv_filts=100
3839mtlalpha=0.5
3940
4041# minibatch related
41- batchsize=30
42+ batchsize=50
4243maxlen_in=800 # if input length > maxlen_in, batchsize is automatically reduced
4344maxlen_out=150 # if output length > maxlen_out, batchsize is automatically reduced
4445
@@ -47,7 +48,7 @@ opt=adadelta
4748epochs=15
4849
4950# rnnlm related
50- lm_weight=1.0
51+ lm_weight=0.3
5152
5253# decoding parameter
5354beam_size=20
@@ -57,37 +58,62 @@ minlenratio=0.0
5758ctc_weight=0.3
5859recog_model=acc.best # set a model to be used for decoding: 'acc.best' or 'loss.best'
5960
60- # data
61- wsj0=/export/corpora5/LDC/LDC93S6B
62- wsj1=/export/corpora5/LDC/LDC94S13B
61+ # Set this to somewhere where you want to put your data, or where
62+ # someone else has already put it. You'll want to change this
63+ # if you're not on the CLSP grid.
64+ datadir=/export/a15/vpanayotov/data
6365
64- # bpe
65- nbpe=500
66+ # base url for downloads.
67+ data_url=www.openslr.org/resources/12
68+
69+ # bpemode (unigram or bpe)
70+ nbpe=200
71+ bpemode=unigram
6672
6773# exp tag
6874tag=" " # tag for managing experiments.
6975
7076. utils/parse_options.sh || exit 1;
7177
72- . ./path.sh
73- . ./cmd.sh
78+ . ./path.sh
79+ . ./cmd.sh
80+
81+ # check gpu option usage
82+ if [ ! -z $gpu ]; then
83+ echo " WARNING: --gpu option will be deprecated."
84+ echo " WARNING: please use --ngpu option."
85+ if [ $gpu -eq -1 ]; then
86+ ngpu=0
87+ else
88+ ngpu=1
89+ fi
90+ fi
7491
7592# Set bash to 'debug' mode, it will exit on :
7693# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
7794set -e
7895set -u
7996set -o pipefail
8097
81- train_set=train_si284
82- train_dev=test_dev93
83- recog_set=" test_dev93 test_eval92"
98+ train_set=train_960
99+ train_dev=dev
100+ recog_set=" test_clean test_other dev_clean dev_other"
101+
102+ if [ ${stage} -le -1 ]; then
103+ echo " stage -1: Data Download"
104+ for part in dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500; do
105+ local/download_and_untar.sh ${datadir} ${data_url} ${part}
106+ done
107+ fi
84108
85109if [ ${stage} -le 0 ]; then
86110 # ## Task dependent. You have to make data the following preparation part by yourself.
87111 # ## But you can utilize Kaldi recipes in most cases
88112 echo " stage 0: Data preparation"
89- local/wsj_data_prep.sh ${wsj0} /?? -{? ,?? }.? ${wsj1} /?? -{? ,?? }.?
90- local/wsj_format_data.sh
113+ for part in dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500; do
114+ # use underscore-separated names in data directories.
115+ local/data_prep.sh ${datadir} /LibriSpeech/${part} data/$( echo ${part} | sed s/-/_/g)
116+ done
91117fi
92118
93119feat_tr_dir=${dumpdir} /${train_set} /delta${do_delta} ; mkdir -p ${feat_tr_dir}
@@ -98,81 +124,97 @@ if [ ${stage} -le 1 ]; then
98124 echo " stage 1: Feature Generation"
99125 fbankdir=fbank
100126 # Generate the fbank features; by default 80-dimensional fbanks with pitch on each frame
101- for x in train_si284 test_dev93 test_eval92 ; do
102- steps/make_fbank_pitch.sh --cmd " $train_cmd " 10 data/${x} exp/make_fbank/${x} ${fbankdir}
127+ for x in dev_clean test_clean dev_other test_other train_clean_100 train_clean_360 train_other_500 ; do
128+ steps/make_fbank_pitch.sh --cmd " $train_cmd " 32 data/${x} exp/make_fbank/${x} ${fbankdir}
103129 done
104130
131+ utils/combine_data.sh data/${train_set} _org data/train_clean_100 data/train_clean_360 data/train_other_500
132+ utils/combine_data.sh data/${train_dev} _org data/dev_clean data/dev_other
133+
134+ # remove utt having more than 3000 frames
135+ # remove utt having more than 400 characters
136+ remove_longshortdata.sh --maxframes 3000 --maxchars 400 data/${train_set} _org data/${train_set}
137+ remove_longshortdata.sh --maxframes 3000 --maxchars 400 data/${train_dev} _org data/${train_dev}
138+
105139 # compute global CMVN
106140 compute-cmvn-stats scp:data/${train_set} /feats.scp data/${train_set} /cmvn.ark
107141
108142 # dump features for training
109143 if [[ $( hostname -f) == * .clsp.jhu.edu ]] && [ ! -d ${feat_tr_dir} /storage ]; then
110144 utils/create_split_dir.pl \
111- /export/b{10,11,12,13 }/${USER} /espnet-data/egs/voxforge/asr1/dump/${train_set} /delta${do_delta} /storage \
145+ /export/b{14,15,16,17 }/${USER} /espnet-data/egs/voxforge/asr1/dump/${train_set} /delta${do_delta} /storage \
112146 ${feat_tr_dir} /storage
113147 fi
114148 if [[ $( hostname -f) == * .clsp.jhu.edu ]] && [ ! -d ${feat_dt_dir} /storage ]; then
115149 utils/create_split_dir.pl \
116- /export/b{10,11,12,13 }/${USER} /espnet-data/egs/voxforge/asr1/dump/${train_dev} /delta${do_delta} /storage \
150+ /export/b{14,15,16,17 }/${USER} /espnet-data/egs/voxforge/asr1/dump/${train_dev} /delta${do_delta} /storage \
117151 ${feat_dt_dir} /storage
118152 fi
119- dump.sh --cmd " $train_cmd " 32 --do_delta $do_delta \
153+ dump.sh --cmd " $train_cmd " 80 --do_delta $do_delta \
120154 data/${train_set} /feats.scp data/${train_set} /cmvn.ark exp/dump_feats/train ${feat_tr_dir}
121- dump.sh --cmd " $train_cmd " 4 --do_delta $do_delta \
155+ dump.sh --cmd " $train_cmd " 32 --do_delta $do_delta \
122156 data/${train_dev} /feats.scp data/${train_set} /cmvn.ark exp/dump_feats/dev ${feat_dt_dir}
157+ for rtask in ${recog_set} ; do
158+ feat_recog_dir=${dumpdir} /${rtask} /delta${do_delta} ; mkdir -p ${feat_recog_dir}
159+ dump.sh --cmd " $train_cmd " $do_delta \
160+ data/${rtask} /feats.scp data/${train_set} /cmvn.ark exp/dump_feats/recog/${rtask} \
161+ ${feat_recog_dir}
162+ done
123163fi
124164
125- dict=data/bpe/${train_set} _units_${nbpe} .txt
126- code=data/bpe/code_${nbpe} .txt
127- nlsyms=data/bpe/non_lang_syms.txt
165+ dict=data/lang_char/${train_set} _units.txt
166+ bpemodel=data/lang_char/${train_set}${nbpe}
128167echo " dictionary: ${dict} "
129168if [ ${stage} -le 2 ]; then
130169 # ## Task dependent. You have to check non-linguistic symbols used in the corpus.
131170 echo " stage 2: Dictionary and Json Data Preparation"
132- mkdir -p data/bpe
171+ mkdir -p data/lang_char/
133172 echo " <unk> 1" > ${dict} # <unk> must be 1, 0 will be used for "blank" in CTC
134- cut -f 2- -d" " ${train_set} /text | learn_bpe.py -s ${nbpe} > ${code}
135- cut -f 2- -d " " data/${train_set} /text | apply_bpe.py -c ${code} \
136- | tr ' ' ' \n' | sort | uniq | awk ' {print $0 " " NR+1}' >> ${dict}
173+ cut -f 2- -d" " ${train_set} /text > data/lang_char/input.txt
174+ spm_train --input= data/lang_char/input.txt --vocab_size= ${nbpe} --model_type= ${bpemode} --model_prefix= ${bpemodel} --input_sentence_size=100000000
175+ spm_encode --model= ${bpemodel} .model --output_format=piece < data/lang_char/input.txt | tr ' ' ' \n' | sort | uniq | awk ' {print $0 " " NR+1}' >> ${dict}
137176 wc -l ${dict}
138177
139- echo " make a non-linguistic symbol list"
140- cut -f 2- data/${train_set} /text | tr " " " \n" | sort | uniq | grep " <" > ${nlsyms}
141- cat ${nlsyms}
142-
143- echo " make json files"
144- data2json.sh --feat ${feat_tr_dir} /feats.scp --bpecode ${code} \
145- data/${train_set} ${dict} > ${feat_tr_dir} /data_${nbpe} .json
146- data2json.sh --feat ${feat_dt_dir} /feats.scp --bpecode ${code} \
178+ # make json labels
179+ data2json.sh --feat ${feat_tr_dir} /feats.scp --bpecode ${bpemodel} .model \
180+ data/${train_set} ${dict} > ${feat_tr_dir} /data_${nbpe} .json
181+ data2json.sh --feat ${feat_dt_dir} /feats.scp --bpecode ${bpemodel} .model \
147182 data/${train_dev} ${dict} > ${feat_dt_dir} /data_${nbpe} .json
148183fi
149184
150- # you can skip this and remove --rnnlm option in the recognition (stage 5)
151- lmexpdir=exp/train_rnnlm_2layer_bs2048_bpe ${nbpe}
185+ # You can skip this and remove --rnnlm option in the recognition (stage 5)
186+ lmexpdir=exp/train_rnnlm_2layer_bs256
152187mkdir -p ${lmexpdir}
153188if [ ${stage} -le 3 ]; then
154189 echo " stage 3: LM Preparation"
155190 lmdatadir=data/local/lm_train
156191 mkdir -p ${lmdatadir}
157- cut -f 2- -d" " ${train_set} /text | apply_bpe.py -c ${code} | perl -pe ' s/\n/ <eos> /g'
158- > ${lmdatadir} /train_trans_${nbpe} .txt
159- zcat ${wsj1} /13-32.1/wsj1/doc/lng_modl/lm_train/np_data/{87,88,89}/* .z | grep -v " <" | tr [a-z] [A-Z] \
160- | apply_bpe.py -c ${code} | perl -pe ' s/\n/ <eos> /g' >> ${lmdatadir} /train_others_${nbpe} .txt
161- cat ${lmdatadir} /train_trans_${nbpe} .txt ${lmdatadir} /train_others_${nbpe} .txt | tr ' \n' ' ' > ${lmdatadir} /train_${nbpe} .txt
162- cut -f 2- -d" " ${train_dev} /text | apply_bpe.py -c ${code} | perl -pe ' s/\n/ <eos> /g'
163- > ${lmdatadir} /valid_${nbpe} .txt
192+ spm_encode --model=${bpemodel} .model --output_format=piece < data/lang_char/input.txt | perl -pe ' s/\n/ <eos> /g'
193+ > ${lmdatadir} /train.txt
194+ cut -f 2- -d" " ${train_set} /text | spm_encode --model=${bpemodel} .model --output_format=piece | perl -pe ' s/\n/ <eos> /g'
195+ > ${lmdatadir} /valid.txt
196+ # use only 1 gpu
197+ if [ ${ngpu} -gt 1 ]; then
198+ echo " LM training does not support multi-gpu. signle gpu will be used."
199+ lmngpu=1
200+ else
201+ lmngpu=${ngpu}
202+ fi
164203 ${cuda_cmd} ${lmexpdir} /train.log \
165204 lm_train.py \
166- --gpu ${gpu} \
205+ --ngpu ${lmngpu} \
206+ --backend ${backend} \
167207 --verbose 1 \
168208 --outdir ${lmexpdir} \
169- --train-label ${lmdatadir} /train_${nbpe} .txt \
170- --valid-label ${lmdatadir} /valid_${nbpe} .txt \
209+ --train-label ${lmdatadir} /train.txt \
210+ --valid-label ${lmdatadir} /valid.txt \
211+ --epoch 60 \
212+ --batchsize 256 \
171213 --dict ${dict}
172214fi
173215
174216if [ -z ${tag} ]; then
175- expdir=exp/${train_set} _${etype} _e${elayers} _subsample${subsample} _unit${eunits} _proj${eprojs} _d${dlayers} _unit${dunits} _${atype} _aconvc${aconv_chans} _aconvf${aconv_filts} _mtlalpha${mtlalpha} _${opt} _bs${batchsize} _mli${maxlen_in} _mlo${maxlen_out} _bpe${nbpe}
217+ expdir=exp/${train_set} _${etype} _e${elayers} _subsample${subsample} _unit${eunits} _proj${eprojs} _d${dlayers} _unit${dunits} _${atype} _aconvc${aconv_chans} _aconvf${aconv_filts} _mtlalpha${mtlalpha} _${opt} _bs${batchsize} _mli${maxlen_in} _mlo${maxlen_out} _bpe${nbpe} _bpemode ${bpemode}
176218 if ${do_delta} ; then
177219 expdir=${expdir} _delta
178220 fi
@@ -181,32 +223,21 @@ else
181223fi
182224mkdir -p ${expdir}
183225
184- # switch backend
185- if [[ ${backend} == chainer ]]; then
186- train_script=asr_train.py
187- decode_script=asr_recog.py
188- else
189- train_script=asr_train_th.py
190- decode_script=asr_recog_th.py
191- fi
192-
193226if [ ${stage} -le 4 ]; then
194227 echo " stage 4: Network Training"
195-
196- ${cuda_cmd} ${expdir} /train.log \
197- ${train_script }
198- --gpu ${gpu } \
228+ ${cuda_cmd} --gpu ${ngpu} ${expdir} /train.log \
229+ asr_train.py \
230+ --ngpu ${ngpu }
231+ --backend ${backend } \
199232 --outdir ${expdir} /results \
200233 --debugmode ${debugmode} \
201234 --dict ${dict} \
202235 --debugdir ${expdir} \
203236 --minibatches ${N} \
204237 --verbose ${verbose} \
205238 --resume ${resume} \
206- --train-feat scp:${feat_tr_dir} /feats.scp \
207- --valid-feat scp:${feat_dt_dir} /feats.scp \
208- --train-label ${feat_tr_dir} /data_${nbpe} .json \
209- --valid-label ${feat_dt_dir} /data_${nbpe} .json \
239+ --train-json ${feat_tr_dir} /data_${nbpe} .json \
240+ --valid-json ${feat_dt_dir} /data_${nbpe} .json \
210241 --etype ${etype} \
211242 --elayers ${elayers} \
212243 --eunits ${eunits} \
@@ -232,29 +263,27 @@ if [ ${stage} -le 5 ]; then
232263 for rtask in ${recog_set} ; do
233264 (
234265 decode_dir=decode_${rtask} _beam${beam_size} _e${recog_model} _p${penalty} _len${minlenratio} -${maxlenratio} _ctcw${ctc_weight} _rnnlm${lm_weight}
266+ feat_recog_dir=${dumpdir} /${rtask} /delta${do_delta}
235267
236268 # split data
237269 data=data/${rtask}
238270 split_data.sh --per-utt ${data} ${nj} ;
239271 sdata=${data} /split${nj} utt;
240272
241- # feature extraction
242- feats=" ark,s,cs:apply-cmvn --norm-vars=true data/${train_set} /cmvn.ark scp:${sdata} /JOB/feats.scp ark:- |"
243- if ${do_delta} ; then
244- feats=" $feats add-deltas ark:- ark:- |"
245- fi
246-
247273 # make json labels for recognition
248- data2json.sh --bpecode ${code} ${data} ${dict} > ${data} /data_${nbpe} .json
274+ for j in ` seq 1 ${nj} ` ; do
275+ data2json.sh --bpecode ${bpemodel} .model --feat ${feat_recog_dir} /feats.scp \
276+ ${sdata} /${j} ${dict} > ${sdata} /${j} /data_${nbpe} .json
277+ done
249278
250279 # ### use CPU for decoding
251- gpu=-1
280+ ngpu=0
252281
253282 ${decode_cmd} JOB=1:${nj} ${expdir} /${decode_dir} /log/decode.JOB.log \
254- ${decode_script} \
255- --gpu ${gpu } \
256- --recog-feat " $feats "
257- --recog-label ${data} /data_${nbpe} .json \
283+ asr_recog.py \
284+ --ngpu ${ngpu } \
285+ --backend ${backend} \
286+ --recog-json ${sdata} /JOB /data_${nbpe} .json \
258287 --result-label ${expdir} /${decode_dir} /data.JOB.json \
259288 --model ${expdir} /results/model.${recog_model} \
260289 --model-conf ${expdir} /results/model.conf \
@@ -264,11 +293,12 @@ if [ ${stage} -le 5 ]; then
264293 --minlenratio ${minlenratio} \
265294 --ctc-weight ${ctc_weight} \
266295 --rnnlm ${lmexpdir} /rnnlm.model.best \
267- --lm-weight ${lm_weight} &
296+ --lm-weight ${lm_weight} \
297+ &
268298 wait
269299
270- score_sclite.sh --bpe true --nlsyms ${nlsyms} ${expdir} /${decode_dir} ${dict}
271-
300+ score_sclite.sh --bpe ${nbpe} --bpemodel ${bpemodel} .model --wer true ${expdir} /${decode_dir} ${dict}
301+
272302 ) &
273303 done
274304 wait
0 commit comments