Skip to content

S
S2T
Commit 6ca43d9e by xuchen
Options

shell and yaml

parent 9d4fe566
正在显示 包含 233 行增加 和 59 行删除
egs/aishell/asr/conf/xctc.yaml 0 → 100644
xctc-weight: 0.3
share-xctc-and-embed: True
\ No newline at end of file
egs/aishell/asr/conf/xinter.yaml 0 → 100644
inter-xctc-weight: 0.2
inter-xctc-layers: 6,9
xctc-pae: none
# xctc-pae: inter_league
xctc-cross-attn: False
cross-attn-start-layer: 7
cross-attn-layer: 6
cross-attn-collaboration-mode: parallel
cross-attn-league-s1-ratio: 0.5
cross-attn-league-s2-ratio: 0.5
cross-attn-league-out-norm: False
cross-attn-league-gated: False
cross-attn-league-drop-net: False
cross-attn-league-drop-net-prob: 0.2
cross-attn-league-drop-net-mix: False
# xctc-pae-ground-truth-ratio: 0.1
# xctc-pae-ground-truth-ratio-adaptive: True
# xctc-pae-ground-truth-only-mistake: True
# pae-oracle-smooth: True
# pae-gumbel: True
# pae-distribution-hard: True
# pae-drop-prob: 0.0
# pae-distribution-cutoff: 10
# share-pae-and-xctc: True
# pae-embed-norm: True
# pae-out-norm: True
# ctc-self-distill-weight: 1
# target-ctc-self-distill-weight: 1
# ctc-self-distill-prob: 0.1
# cal-all-ctc: True
\ No newline at end of file
egs/aishell/asr/decode.sh
......@@ -2,7 +2,7 @@
gpu_num=1
data_dir=
data_tag=asr
test_subset=(dev test)
exp_name=
......@@ -11,11 +11,14 @@ if [ "$#" -eq 1 ]; then
fi
cer=1
ctc_infer=0
ctc_infer=1
n_average=10
beam_size=5
infer_ctc_weight=0.1
len_penalty=1.0
max_tokens=80000
max_tokens=50000
batch_size=1
infer_debug=0
dec_model=checkpoint_best.pt
cmd="./run.sh
......@@ -28,12 +31,16 @@ cmd="./run.sh
--ctc_infer ${ctc_infer}
--beam_size ${beam_size}
--len_penalty ${len_penalty}
--batch_size ${batch_size}
--max_tokens ${max_tokens}
--dec_model ${dec_model}
--ctc_infer ${ctc_infer}
--infer_ctc_weight ${infer_ctc_weight}
--infer_debug ${infer_debug}
"
if [[ -n ${data_dir} ]]; then
cmd="$cmd --data_dir ${data_dir}"
if [[ -n ${data_tag} ]]; then
cmd="$cmd --data_tag ${data_tag}"
fi
if [[ ${#test_subset[@]} -ne 0 ]]; then
subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
......
egs/aishell/asr/run.sh
......@@ -39,7 +39,7 @@ data_tag=asr
task=speech_to_text
vocab_type=unigram
vocab_type=char
vocab_size=5000
vocab_size=10000
speed_perturb=1
lcrm=0
tokenizer=0
......@@ -77,14 +77,17 @@ step_valid=0
dec_model=checkpoint_best.pt
cer=1
ctc_infer=0
infer_ctc_weight=0
ctc_self_ensemble=0
ctc_inter_logit=0
n_average=10
batch_size=0
beam_size=5
len_penalty=1.0
single=0
epoch_ensemble=0
best_ensemble=1
infer_debug=0
infer_score=0
# infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
......@@ -115,13 +118,6 @@ if [[ "${vocab_type}" == "char" ]]; then
data_dir=${data_dir}_char
exp_prefix=${exp_prefix}_char
fi
if [[ ! -d /mnt/bd/data-model && -d /mnt/bd/data-model2 ]]; then
sudo ln -s /mnt/bd/data-model2/ /mnt/bd/data-model
fi
if [[ ! -d ${data_dir} ]]; then
echo "No feature dir ${data_dir}"
exit
fi
# setup nccl envs
export NCCL_IB_DISABLE=0
......@@ -174,6 +170,8 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
--task asr
--src-lang ${src_lang}
--splits ${valid_split},${test_split},${train_split}
--add-src
--share
--vocab-type ${vocab_type}
--vocab-size ${vocab_size}"
......@@ -309,8 +307,8 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
mv tmp.log $log
log=${model_dir}/train.log
cmd="${cmd} 2>&1 | tee -a ${log}"
#cmd="nohup ${cmd} >> ${log} 2>&1 &"
# cmd="${cmd} 2>&1 | tee -a ${log}"
cmd="${cmd} >> ${log} 2>&1 "
if [[ $eval -eq 1 ]]; then
# tensorboard
port=6666
......@@ -326,7 +324,7 @@ fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "Stage 2: Decoding"
dec_models=
if [[ ${single} -eq 1 ]]; then
if [[ ${n_average} -eq 1 ]]; then
dec_models=${dec_model}
fi
if [[ ${n_average} -ne 1 ]]; then
......@@ -370,7 +368,7 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
fi
for dec_model in ${dec_models[@]}; do
suffix=beam${beam_size}_alpha${len_penalty}_tokens${max_tokens}
suffix=alpha${len_penalty}
model_str=`echo $dec_model | sed -e "s#checkpoint##" | sed "s#.pt##"`
suffix=${suffix}_${model_str}
if [[ -n ${cer} && ${cer} -eq 1 ]]; then
......@@ -378,6 +376,13 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
else
suffix=${suffix}_wer
fi
suffix=${suffix}_beam${beam_size}
if [[ ${batch_size} -ne 0 ]]; then
suffix=${suffix}_batch${batch_size}
else
suffix=${suffix}_tokens${max_tokens}
fi
if [[ ${ctc_infer} -eq 1 ]]; then
suffix=${suffix}_ctc
fi
......@@ -387,6 +392,12 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
if [[ ${ctc_inter_logit} -ne 0 ]]; then
suffix=${suffix}_logit${ctc_inter_logit}
fi
if (( $(echo "${infer_ctc_weight} > 0" | bc -l) )); then
suffix=${suffix}_ctc${infer_ctc_weight}
fi
if [[ ${infer_score} -eq 1 ]]; then
suffix=${suffix}_score
fi
suffix=`echo $suffix | sed -e "s#__#_#"`
result_file=${model_dir}/decode_result_${suffix}
......@@ -395,16 +406,23 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
test_subset=${test_subset//,/ }
for subset in ${test_subset[@]}; do
subset=${subset}
cmd="python3 ${code_dir}/fairseq_cli/generate.py
if [[ ${infer_debug} -ne 0 ]]; then
cmd="python3 -m debugpy --listen 0.0.0.0:5678 --wait-for-client"
else
cmd="python3 "
fi
cmd="$cmd ${code_dir}/fairseq_cli/generate.py
${data_dir}
--config-yaml ${data_config}
--gen-subset ${subset}
--task speech_to_text
--path ${model_dir}/${dec_model}
--results-path ${model_dir}
--batch-size ${batch_size}
--max-tokens ${max_tokens}
--beam ${beam_size}
--lenpen ${len_penalty}
--infer-ctc-weight ${infer_ctc_weight}
--scoring wer"
if [[ ${cer} -eq 1 ]]; then
......@@ -423,10 +441,14 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
cmd="${cmd}
--ctc-inter-logit ${ctc_inter_logit}"
fi
if [[ ${infer_score} -eq 1 ]]; then
cmd="${cmd}
--score-reference"
fi
if [[ -n ${infer_parameters} ]]; then
cmd="${cmd}
${infer_parameters}"
fi
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
......@@ -455,12 +477,14 @@ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
if [[ ${ctc_infer} -eq 1 && -f ${model_dir}/${ctc_file} ]]; then
ref_file=${model_dir}/${subset}.${src_lang}
if [[ ! -f ${ref_file} ]]; then
python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "src_text"
python3 ./local/extract_txt_from_tsv.py ${data_dir}/${subset}.tsv ${ref_file} "tgt_text"
fi
if [[ -f ${ref_file} ]]; then
ctc=$(mktemp -t temp.record.XXXXXX)
cd ./local
./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} > ${ctc}
cmd="./cal_wer.sh ${model_dir} ${subset} ${trans_file} ${ctc_file} ${ref_file} > ${ctc}"
#echo $cmd
eval $cmd
cd ..
echo "CTC WER" >> ${result_file}
......
egs/librispeech/asr/conf/basis.yaml
......@@ -18,8 +18,8 @@ best_checkpoint_metric: dec_wer
maximize_best_checkpoint_metric: False
validate-interval: 5
no-epoch-checkpoints: True
# keep-last-epochs: 10
# no-epoch-checkpoints: True
keep-last-epochs: 10
keep-best-checkpoints: 10
num-workers: 8
......
egs/librispeech/asr/conf/xctc.yaml 0 → 100644
xctc-weight: 0.3
share-xctc-and-embed: True
\ No newline at end of file
egs/librispeech/asr/conf/xinter.yaml 0 → 100644
inter-xctc-weight: 0.2
inter-xctc-layers: 6,9
xctc-pae: none
# xctc-pae: inter_league
xctc-cross-attn: False
cross-attn-start-layer: 7
cross-attn-layer: 6
cross-attn-collaboration-mode: parallel
cross-attn-league-s1-ratio: 0.5
cross-attn-league-s2-ratio: 0.5
cross-attn-league-out-norm: False
cross-attn-league-gated: False
cross-attn-league-drop-net: False
cross-attn-league-drop-net-prob: 0.2
cross-attn-league-drop-net-mix: False
# xctc-pae-ground-truth-ratio: 0.1
# xctc-pae-ground-truth-ratio-adaptive: True
# xctc-pae-ground-truth-only-mistake: True
# pae-oracle-smooth: True
# pae-gumbel: True
# pae-distribution-hard: True
# pae-drop-prob: 0.0
# pae-distribution-cutoff: 10
# share-pae-and-xctc: True
# pae-embed-norm: True
# pae-out-norm: True
# ctc-self-distill-weight: 1
# target-ctc-self-distill-weight: 1
# ctc-self-distill-prob: 0.1
# cal-all-ctc: True
\ No newline at end of file
egs/librispeech/asr/decode.sh
......@@ -2,8 +2,9 @@
gpu_num=1
data_dir=
data_tag=asr
test_subset=(dev-clean dev-other test-clean test-other all)
test_subset=(dev-clean dev-other test-clean test-other)
exp_name=
if [ "$#" -eq 1 ]; then
......@@ -13,7 +14,7 @@ fi
ctc_infer=0
n_average=10
beam_size=5
infer_ctc_weight=0.3
infer_ctc_weight=0.1
len_penalty=1.0
max_tokens=50000
batch_size=1
......@@ -36,8 +37,8 @@ cmd="./run.sh
--infer_debug ${infer_debug}
"
if [[ -n ${data_dir} ]]; then
cmd="$cmd --data_dir ${data_dir}"
if [[ -n ${data_tag} ]]; then
cmd="$cmd --data_tag ${data_tag}"
fi
if [[ ${#test_subset[@]} -ne 0 ]]; then
subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
......
egs/librispeech/asr/run.sh
......@@ -76,14 +76,15 @@ infer_ctc_weight=0
ctc_self_ensemble=0
ctc_inter_logit=0
n_average=10
batch_size=0
beam_size=5
len_penalty=1.0
single=0
epoch_ensemble=0
best_ensemble=1
epoch_ensemble=1
best_ensemble=0
infer_debug=0
infer_score=0
# infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
#infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
data_config=config.yaml
......@@ -101,13 +102,6 @@ if [[ "${vocab_type}" == "char" ]]; then
data_dir=${data_dir}_char
exp_prefix=${exp_prefix}_char
fi
if [[ ! -d /mnt/bd/data-model && -d /mnt/bd/data-model2 ]]; then
sudo ln -s /mnt/bd/data-model2/ /mnt/bd/data-model
fi
if [[ ! -d ${data_dir} ]]; then
echo "No feature dir ${data_dir}"
exit
fi
# setup nccl envs
export NCCL_IB_DISABLE=0
......@@ -278,8 +272,8 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
mv tmp.log $log
log=${model_dir}/train.log
cmd="${cmd} 2>&1 | tee -a ${log}"
#cmd="nohup ${cmd} >> ${log} 2>&1 &"
# cmd="${cmd} 2>&1 | tee -a ${log}"
cmd="${cmd} >> ${log} 2>&1 "
if [[ $eval -eq 1 ]]; then
# tensorboard
port=6666
......@@ -295,7 +289,7 @@ fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "Stage 2: Decoding"
dec_models=
if [[ ${single} -eq 1 ]]; then
if [[ ${n_average} -eq 1 ]]; then
dec_models=${dec_model}
fi
if [[ ${n_average} -ne 1 ]]; then
......
egs/mustc/asr/conf/xctc.yaml 0 → 100644
xctc-weight: 0.3
share-xctc-and-embed: True
\ No newline at end of file
egs/mustc/asr/conf/xinter.yaml 0 → 100644
inter-xctc-weight: 0.2
inter-xctc-layers: 6,9
xctc-pae: none
# xctc-pae: inter_league
xctc-cross-attn: False
cross-attn-start-layer: 7
cross-attn-layer: 6
cross-attn-collaboration-mode: parallel
cross-attn-league-s1-ratio: 0.5
cross-attn-league-s2-ratio: 0.5
cross-attn-league-out-norm: False
cross-attn-league-gated: False
cross-attn-league-drop-net: False
cross-attn-league-drop-net-prob: 0.2
cross-attn-league-drop-net-mix: False
# xctc-pae-ground-truth-ratio: 0.1
# xctc-pae-ground-truth-ratio-adaptive: True
# xctc-pae-ground-truth-only-mistake: True
# pae-oracle-smooth: True
# pae-gumbel: True
# pae-distribution-hard: True
# pae-drop-prob: 0.0
# pae-distribution-cutoff: 10
# share-pae-and-xctc: True
# pae-embed-norm: True
# pae-out-norm: True
# ctc-self-distill-weight: 1
# target-ctc-self-distill-weight: 1
# ctc-self-distill-prob: 0.1
# cal-all-ctc: True
\ No newline at end of file
egs/mustc/st/decode.sh
......@@ -2,7 +2,7 @@
gpu_num=1
data_dir=
data_tag=
test_subset=(dev tst-COMMON)
exp_name=
......@@ -38,8 +38,8 @@ cmd="./run.sh
--infer_debug ${infer_debug}
"
if [[ -n ${data_dir} ]]; then
cmd="$cmd --data_dir ${data_dir}"
if [[ -n ${data_tag} ]]; then
cmd="$cmd --data_tag ${data_tag}"
fi
if [[ ${#test_subset[@]} -ne 0 ]]; then
subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
......
egs/mustc/st/pipe2.sh 0 → 100644
./run.sh --stage 2 --tgt_lang fr --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
./run.sh --stage 2 --tgt_lang es --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
./run.sh --stage 2 --tgt_lang it --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
./run.sh --stage 2 --tgt_lang nl --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
./run.sh --stage 2 --tgt_lang pt --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
./run.sh --stage 2 --tgt_lang ro --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
./run.sh --stage 2 --tgt_lang ru --infer_ctc_weight 0.1 --batch_size 1 --exp_name big_conformer_ctc_inter_xctc_xinter_enc18_wsum0.45_unnormpae_oracle0.1
egs/mustc/st/prep.sh 0 → 100644
./run.sh --stage 0 --stop_stage 0 --tgt_lang es
./run.sh --stage 0 --stop_stage 0 --tgt_lang fr
./run.sh --stage 0 --stop_stage 0 --tgt_lang it
./run.sh --stage 0 --stop_stage 0 --tgt_lang nl
./run.sh --stage 0 --stop_stage 0 --tgt_lang pt
./run.sh --stage 0 --stop_stage 0 --tgt_lang ro
./run.sh --stage 0 --stop_stage 0 --tgt_lang ru
egs/mustc/st/run.sh
......@@ -79,6 +79,7 @@ step_valid=0
bleu_valid=0
# Decoding Settings
batch_size=0
sacrebleu=1
dec_model=checkpoint_best.pt
ctc_infer=0
......@@ -88,7 +89,7 @@ beam_size=5
len_penalty=1.0
infer_debug=0
infer_score=0
# infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
#infer_parameters="--cal-monotonic-cross-attn-weights --cal-localness --localness-window 0.1 --cal-topk-cross-attn-weights --topk-cross-attn-weights 15 --cal-entropy"
# Parsing Options
if [[ ${share_dict} -eq 1 ]]; then
......@@ -116,14 +117,6 @@ if [[ ${use_raw_audio} -eq 1 ]]; then
data_dir=${data_dir}_raw
exp_prefix=${exp_prefix}_raw
fi
if [[ ! -d /mnt/bd/data-model && -d /mnt/bd/data-model2 ]]; then
sudo ln -s /mnt/bd/data-model2/ /mnt/bd/data-model
fi
if [[ ! -d ${data_dir} ]]; then
echo "No feature dir ${data_dir}"
exit
fi
export PATH=$PATH:${code_dir}/scripts
. ./local/parse_options.sh || exit 1;
......@@ -361,8 +354,8 @@ if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
mv tmp.log $log
log=${model_dir}/train.log
cmd="${cmd} 2>&1 | tee -a ${log}"
#cmd="nohup ${cmd} >> ${log} 2>&1 &"
# cmd="${cmd} 2>&1 | tee -a ${log}"
cmd="${cmd} >> ${log} 2>&1 "
if [[ $eval -eq 1 ]]; then
# tensorboard
port=6666
......
entry.sh
......@@ -6,7 +6,6 @@ export ST_ROOT=/xuchen/st
export NCCL_DEBUG=INFO
echo "nameserver 114.114.114.114" >> /etc/resolv.conf
pip3 install espnet -i https://pypi.tuna.tsinghua.edu.cn/simple
if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then
echo "default stage: env configure"
......
entry_loop.sh
......@@ -6,7 +6,6 @@ export ST_ROOT=/xuchen/st
export NCCL_DEBUG=INFO
echo "nameserver 114.114.114.114" >> /etc/resolv.conf
pip3 install espnet -i https://pypi.tuna.tsinghua.edu.cn/simple
if [[ `pip list | grep fairseq | wc -l` -eq 0 ]]; then
echo "default stage: env configure"
......@@ -16,21 +15,46 @@ fi
all_cmd=$1
all_cmd_dir=`dirname ${all_cmd}`
echo "cmd dir: $all_cmd_dir"
echo_flag=1
pre_num=-1
cp $all_cmd ${all_cmd}.bak
while :
do
line=`head -n1 $all_cmd`
if [[ -z $line ]]; then
echo "No cmd. Waiting."
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
device=()
count=0
for dev in ${all_devices[@]}
do
item=$((dev + 2))
use=$(head -n $item ${record} | tail -1 | cut -d '|' -f3 | cut -d '/' -f1)
if [[ $use -lt 100 ]]; then
device[$count]=$dev
count=$((count + 1))
if [[ $count -eq $gpu_num ]]; then
break
fi
fi
done
if [[ $echo_flag -eq 1 ]]; then
echo "No cmd. Current free GPU: ${count}. Waiting."
echo_flag=0
fi
sleep 300s
continue
fi
echo $line
gpu_num=$(echo $line | awk '{print $1}')
shell_script=$(echo $line | awk '{print $2}')
cmd=$(echo $line | awk '{$1=""; print $0}')
echo_flag=1
while :
do
record=$(mktemp -t temp.record.XXXXXX)
......@@ -53,6 +77,10 @@ do
fi
done
if [[ ${#device[@]} -lt $gpu_num ]]; then
if [[ ${pre_num} -ne ${gpu_num} ]]; then
echo "Current free GPU: ${count}, need GPU: ${gpu_num}. Waiting."
fi
pre_num=$gpu_num
sleep 300s
else
echo "Run $cmd"
......@@ -67,7 +95,7 @@ do
export CUDA_VISIBLE_DEVICES=$avail_devices
eval $cmd &
cd ${THIS_DIR}
sleep 100s
sleep 300s
fi
break
done
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论