Skip to content

F
Fairseq-S2T
Commit 093098e4 by xuchen
Options

Support the raw audio and online speed perturbation! 

I have wanted to do this for a long time!
parent d220c040
正在显示 包含 2529 行增加 和 17 行删除
egs/libri_trans/asr/binary.sh 0 → 100644
set -e
eval=1
lcrm=0
tokenizer=0
root_dir=~/st/Fairseq-S2T
data_dir=~/st/data/test
vocab_dir=~/st/data/mustc/st/en-de
asr_vocab_prefix=spm_unigram10000_st_share
src_lang=en
tgt_lang=de
subsets=(2019)
cp -r ${vocab_dir}/${asr_vocab_prefix}.* ${data_dir}/${src_lang}-${tgt_lang}
rm -rf ${data_dir}/${src_lang}-${tgt_lang}/fbank80.zip
splits=$(echo ${subsets[*]} | sed 's/ /,/g')
cmd="python ${root_dir}/examples/speech_to_text/prep_st_data.py
--data-root ${data_dir}
--output-root ${data_dir}
--splits ${splits}
--task asr
--src-lang ${src_lang}
--tgt-lang ${tgt_lang}
--add-src
--share
--asr-prefix ${asr_vocab_prefix}
--cmvn-type utterance"
if [[ ${lcrm} -eq 1 ]]; then
cmd="$cmd
--lowercase-src
--rm-punc-src"
fi
if [[ ${tokenizer} -eq 1 ]]; then
cmd="$cmd
--tokenizer"
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
egs/libri_trans/asr/conf/base.yaml 0 → 100644
train-subset: train
valid-subset: dev
max-epoch: 100
max-update: 100000
num-workers: 0
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
arch: s2t_transformer_s
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
conv-kernel-sizes: 5,5
conv-channels: 1024
dropout: 0.1
activation-fn: relu
encoder-embed-dim: 256
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
attention-dropout: 0.1
activation-dropout: 0.1
egs/libri_trans/asr/conf/conformer.yaml 0 → 100644
macaron-style: True
use-cnn-module: True
cnn-module-kernel: 31
egs/libri_trans/asr/conf/dlcl.yaml 0 → 100644
use-enc-dlcl: True
use-dec-dlcl: True
egs/libri_trans/asr/conf/local_attn.yaml 0 → 100644
encoder-attention-type: local
hard-mask-window: 0
gauss-mask-sigma: 3
init-mask-weight: 0
\ No newline at end of file
egs/libri_trans/asr/conf/pds_base.yaml 0 → 100644
arch: pdss2t_transformer_s_8
train-subset: train_asr
valid-subset: dev_asr
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/asr/conf/pds_base_16.yaml 0 → 100644
arch: pdss2t_transformer_s_16
encoder-embed-dim: 256
pyramid-stages: 4
#pyramid-dropout: 0
pyramid-layers: 2_2_6_2
pyramid-ratios: 2_2_2_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1
pyramid-kernel-sizes: 5_5_5_5
pyramid-ffn-ratios: 8_8_8_8
pyramid-attn-heads: 4_4_4_4
train-subset: train_asr
valid-subset: dev_asr
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/asr/conf/pds_base_32.yaml 0 → 100644
arch: pdss2t_transformer_s_32
encoder-embed-dim: 256
pyramid-stages: 5
#pyramid-dropout: 0
pyramid-layers: 2_2_3_3_2
pyramid-ratios: 2_2_2_2_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1_1
pyramid-kernel-sizes: 5_5_5_5_5
pyramid-ffn-ratios: 8_8_8_8_8
pyramid-attn-heads: 4_4_4_4_4
train-subset: train_asr
valid-subset: dev_asr
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/asr/conf/pds_base_8.yaml 0 → 100644
arch: pdss2t_transformer_s_8
encoder-embed-dim: 256
pyramid-stages: 4
#pyramid-dropout: 0
pyramid-layers: 3_3_3_3
pyramid-ratios: 2_2_1_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1
pyramid-kernel-sizes: 5_5_5_5
pyramid-ffn-ratios: 8_8_8_8
pyramid-attn-heads: 4_4_4_4
train-subset: train_asr
valid-subset: dev_asr
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/asr/conf/rpr.yaml 0 → 100644
encoder-attention-type: rel_selfattn
#encoder-attention-type: relative
#max-encoder-relative-length: 100
egs/libri_trans/asr/decode.sh 0 → 100755
#! /bin/bash
gpu_num=1
data_dir=
test_subset=(test)
exp_name=
if [ "$#" -eq 1 ]; then
exp_name=$1
fi
n_average=10
beam_size=5
len_penalty=1.0
max_tokens=80000
dec_model=checkpoint_best.pt
cmd="./run.sh
--stage 2
--stop_stage 2
--gpu_num ${gpu_num}
--exp_name ${exp_name}
--n_average ${n_average}
--beam_size ${beam_size}
--len_penalty ${len_penalty}
--max_tokens ${max_tokens}
--dec_model ${dec_model}
"
if [[ -n ${data_dir} ]]; then
cmd="$cmd --data_dir ${data_dir}"
fi
if [[ ${#test_subset[@]} -ne 0 ]]; then
subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
cmd="$cmd --test_subset ${subsets}"
fi
echo $cmd
eval $cmd
egs/libri_trans/asr/local/monitor.sh 0 → 100755
gpu_num=4
cmd="sh train.sh"
while :
do
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
count=0
for dev in ${all_devices[@]}
do
line=$((dev + 2))
use=$(head -n $line ${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 [[ ${#device[@]} -lt $gpu_num ]]; then
sleep 60s
else
echo "Run $cmd"
eval $cmd
sleep 10s
exit
fi
done
egs/libri_trans/asr/local/parse_options.sh 0 → 100755
#!/usr/bin/env bash
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey);
# Arnab Ghoshal, Karel Vesely
# 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.
# Parse command-line options.
# To be sourced by another script (as in ". parse_options.sh").
# Option format is: --option-name arg
# and shell variable "option_name" gets set to value "arg."
# The exception is --help, which takes no arguments, but prints the
# $help_message variable (if defined).
###
### The --config file options have lower priority to command line
### options, so we need to import them first...
###
# Now import all the configs specified by command-line, in left-to-right order
for ((argpos=1; argpos<$#; argpos++)); do
if [ "${!argpos}" == "--config" ]; then
argpos_plus1=$((argpos+1))
config=${!argpos_plus1}
[ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
. $config # source the config file.
fi
done
###
### Now we process the command line options
###
while true; do
[ -z "${1:-}" ] && break; # break if there are no arguments
case "$1" in
# If the enclosing script is called with --help option, print the help
# message and exit. Scripts should put help messages in $help_message
--help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
else printf "$help_message\n" 1>&2 ; fi;
exit 0 ;;
--*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
exit 1 ;;
# If the first command-line argument begins with "--" (e.g. --foo-bar),
# then work out the variable name as $name, which will equal "foo_bar".
--*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
# Next we test whether the variable in question is undefned-- if so it's
# an invalid option and we die. Note: $0 evaluates to the name of the
# enclosing script.
# The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
# is undefined. We then have to wrap this test inside "eval" because
# foo_bar is itself inside a variable ($name).
eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
oldval="`eval echo \\$$name`";
# Work out whether we seem to be expecting a Boolean argument.
if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
was_bool=true;
else
was_bool=false;
fi
# Set the variable to the right value-- the escaped quotes make it work if
# the option had spaces, like --cmd "queue.pl -sync y"
eval $name=\"$2\";
# Check that Boolean-valued arguments are really Boolean.
if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
exit 1;
fi
shift 2;
;;
*) break;
esac
done
# Check for an empty argument to the --cmd option, which can easily occur as a
# result of scripting errors.
[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
true; # so this script returns exit code 0.
egs/libri_trans/asr/local/utils.sh 0 → 100755
get_devices(){
gpu_num=$1
use_cpu=$2
device=()
while :
do
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
count=0
for dev in ${all_devices[@]}
do
line=$((dev + 2))
use=$(head -n $line ${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 [[ ${#device[@]} -lt $gpu_num ]]; then
if [[ $use_cpu -eq 1 ]]; then
device=(-1)
else
sleep 60s
fi
else
break
fi
done
echo ${device[*]} | sed 's/ /,/g'
return $?
}
egs/libri_trans/asr/run.sh 0 → 100755
#! /bin/bash
# Processing ASR Datasets
# Copyright 2021 Natural Language Processing Laboratory
# Xu Chen (xuchenneu@163.com)
# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
set -e
#set -u
set -o pipefail
export PYTHONIOENCODING=UTF-8
eval=1
time=$(date "+%m%d_%H%M")
stage=0
stop_stage=0
######## hardware ########
# devices
#device=()
gpu_num=0
update_freq=1
s2t_dir=~/Code/st
root_dir=${s2t_dir}/Fairseq-S2T
pwd_dir=$PWD
# dataset
src_lang=en
tgt_lang=fr
lang=${src_lang}-${tgt_lang}
dataset=libri_trans
task=speech_to_text
vocab_type=unigram
vocab_size=1000
speed_perturb=0
lcrm=1
tokenizer=0
use_raw_audio=1
use_specific_dict=0
specific_prefix=st
specific_dir=${s2t_dir}/data/mustc/st/en-de
asr_vocab_prefix=spm_unigram10000_st_share
org_data_dir=${s2t_dir}/data/${dataset}
data_dir=${s2t_dir}/data/${dataset}/asr
train_split=train
valid_split=dev
test_split=test
test_subset=test
# exp
exp_prefix=$(date "+%m%d")
extra_tag=
extra_parameter=
exp_tag=baseline
exp_name=
# config
train_config=base
data_config=config.yaml
# training setting
fp16=1
max_tokens=40000
step_valid=0
# decoding setting
dec_model=checkpoint_best.pt
n_average=10
beam_size=5
len_penalty=1.0
if [[ ${speed_perturb} -eq 1 ]]; then
data_dir=${data_dir}_sp
exp_prefix=${exp_prefix}_sp
fi
if [[ ${lcrm} -eq 1 ]]; then
data_dir=${data_dir}_lcrm
exp_prefix=${exp_prefix}_lcrm
fi
if [[ ${use_specific_dict} -eq 1 ]]; then
data_dir=${data_dir}_${specific_prefix}
exp_prefix=${exp_prefix}_${specific_prefix}
fi
if [[ ${tokenizer} -eq 1 ]]; then
data_dir=${data_dir}_tok
exp_prefix=${exp_prefix}_tok
fi
if [[ ${use_raw_audio} -eq 1 ]]; then
data_dir=${data_dir}_raw
exp_prefix=${exp_prefix}_raw
fi
. ./local/parse_options.sh || exit 1;
if [[ -z ${exp_name} ]]; then
config_string=${train_config//,/_}
exp_name=${exp_prefix}_${config_string}_${exp_tag}
if [[ -n ${extra_tag} ]]; then
exp_name=${exp_name}_${extra_tag}
fi
fi
model_dir=$root_dir/../checkpoints/$dataset/asr/${exp_name}
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
echo "stage -1: Data Download"
# pass
fi
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
### Task dependent. You have to make data the following preparation part by yourself.
### But you can utilize Kaldi recipes in most cases
echo "stage 0: ASR Data Preparation"
if [[ ! -e ${data_dir} ]]; then
mkdir -p ${data_dir}
fi
cmd="python ${root_dir}/examples/speech_to_text/prep_audio_data.py
--data-root ${org_data_dir}
--output-root ${data_dir}
--task asr
--splits ${train_split},${valid_split},${test_split}
--src-lang ${src_lang}
--vocab-type ${vocab_type}
--vocab-size ${vocab_size}"
if [[ ${use_raw_audio} -eq 1 ]]; then
cmd="$cmd
--raw"
fi
if [[ ${use_specific_dict} -eq 1 ]]; then
cp -r ${specific_dir}/${asr_vocab_prefix}.* ${data_dir}/${lang}
cmd="$cmd
--asr-prefix ${asr_vocab_prefix}"
fi
if [[ ${speed_perturb} -eq 1 ]]; then
cmd="$cmd
--speed-perturb"
fi
if [[ ${lcrm} -eq 1 ]]; then
cmd="$cmd
--lowercase-src
--rm-punc-src"
fi
if [[ ${tokenizer} -eq 1 ]]; then
cmd="$cmd
--tokenizer"
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "stage 1: ASR Network Training"
[[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
if [[ ${gpu_num} -eq 0 ]]; then
device=""
else
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
fi
echo -e "dev=${device} data=${data_dir} model=${model_dir}"
if [[ ! -d ${model_dir} ]]; then
mkdir -p ${model_dir}
else
echo "${model_dir} exists."
fi
cp ${BASH_SOURCE[0]} ${model_dir}
cp ${PWD}/train.sh ${model_dir}
config_list="${train_config//,/ }"
idx=0
for config in ${config_list[@]}
do
config_path=$pwd_dir/conf/${config}.yaml
if [[ ! -f ${config_path} ]]; then
echo "No config file ${config_path}"
exit
fi
cp ${config_path} ${model_dir}
if [[ idx -eq 0 ]]; then
extra_parameter="${extra_parameter}
--train-config ${config_path}"
else
extra_parameter="${extra_parameter}
--train-config${idx} ${config_path}"
fi
idx=$((idx + 1))
done
cmd="python3 -u ${root_dir}/fairseq_cli/train.py
${data_dir}
--config-yaml ${data_config}
--task ${task}
--max-tokens ${max_tokens}
--skip-invalid-size-inputs-valid-test
--update-freq ${update_freq}
--log-interval 100
--save-dir ${model_dir}
--tensorboard-logdir ${model_dir}"
if [[ -n ${extra_parameter} ]]; then
cmd="${cmd}
${extra_parameter}"
fi
if [[ ${gpu_num} -gt 0 ]]; then
cmd="${cmd}
--distributed-world-size $gpu_num
--ddp-backend no_c10d"
fi
if [[ $fp16 -eq 1 ]]; then
cmd="${cmd}
--fp16"
fi
if [[ $step_valid -eq 1 ]]; then
validate_interval=1
save_interval=1
keep_last_epochs=10
no_epoch_checkpoints=0
save_interval_updates=500
keep_interval_updates=10
else
validate_interval=1
keep_last_epochs=10
fi
if [[ -n $no_epoch_checkpoints && $no_epoch_checkpoints -eq 1 ]]; then
cmd="$cmd
--no-epoch-checkpoints"
fi
if [[ -n $validate_interval ]]; then
cmd="${cmd}
--validate-interval $validate_interval "
fi
if [[ -n $save_interval ]]; then
cmd="${cmd}
--save-interval $save_interval "
fi
if [[ -n $keep_last_epochs ]]; then
cmd="${cmd}
--keep-last-epochs $keep_last_epochs "
fi
if [[ -n $save_interval_updates ]]; then
cmd="${cmd}
--save-interval-updates $save_interval_updates"
if [[ -n $keep_interval_updates ]]; then
cmd="${cmd}
--keep-interval-updates $keep_interval_updates"
fi
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
# save info
log=./history.log
echo "${time} | ${device} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
tail -n 50 ${log} > tmp.log
mv tmp.log $log
export CUDA_VISIBLE_DEVICES=${device}
cmd="nohup ${cmd} >> ${model_dir}/train.log 2>&1 &"
if [[ $eval -eq 1 ]]; then
eval $cmd
sleep 2s
tail -n "$(wc -l ${model_dir}/train.log | awk '{print $1+1}')" -f ${model_dir}/train.log
fi
fi
wait
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "stage 2: ASR Decoding"
if [[ ${n_average} -ne 1 ]]; then
# Average models
dec_model=avg_${n_average}_checkpoint.pt
cmd="python ${root_dir}/scripts/average_checkpoints.py
--inputs ${model_dir}
--num-epoch-checkpoints ${n_average}
--output ${model_dir}/${dec_model}"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval $cmd
else
dec_model=${dec_model}
fi
if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
if [[ ${gpu_num} -eq 0 ]]; then
device=""
else
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
fi
export CUDA_VISIBLE_DEVICES=${device}
result_file=${model_dir}/decode_result
[[ -f ${result_file} ]] && rm ${result_file}
test_subset=${test_subset//,/ }
for subset in ${test_subset[@]}; do
subset=${subset}_asr
cmd="python ${root_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}
--max-tokens ${max_tokens}
--beam ${beam_size}
--lenpen ${len_penalty}
--scoring wer
--wer-tokenizer 13a
--wer-lowercase
--wer-remove-punct
"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
if [[ $eval -eq 1 ]]; then
eval $cmd
tail -n 1 ${model_dir}/generate-${subset}.txt >> ${result_file}
fi
done
cat ${result_file}
fi
egs/libri_trans/asr/train.sh 0 → 100755
#! /bin/bash
# training the model
gpu_num=8
update_freq=1
max_tokens=40000
extra_tag=
extra_parameter=
#extra_tag="${extra_tag}"
#extra_parameter="${extra_parameter} "
exp_tag=
#config_list=(base)
#config_list=(ctc)
#config_list=(base conformer)
#config_list=(pds_base_16)
config_list=(pds_base_16 conformer rpr)
# exp full name
exp_name=
train_config=$(echo ${config_list[*]} | sed 's/ /,/g')
cmd="./run.sh
--stage 1
--stop_stage 1
--gpu_num ${gpu_num}
--update_freq ${update_freq}
--train_config ${train_config}
--max_tokens ${max_tokens}
"
if [[ -n ${exp_name} ]]; then
cmd="$cmd --exp_name ${exp_name}"
fi
if [[ -n ${exp_tag} ]]; then
cmd="$cmd --exp_tag ${exp_tag}"
fi
if [[ -n ${extra_tag} ]]; then
cmd="$cmd --extra_tag ${extra_tag}"
fi
if [[ -n ${extra_parameter} ]]; then
cmd="$cmd --extra_parameter \"${extra_parameter}\""
fi
echo ${cmd}
eval ${cmd}
egs/libri_trans/mt/binary.sh 0 → 100755
set -e
eval=1
root_dir=~/st/Fairseq-S2T
data_dir=/home/xuchen/st/data/wmt/test
vocab_dir=/home/xuchen/st/data/wmt/mt/en-de/unigram32000_share
src_vocab_prefix=spm_unigram32000_share
tgt_vocab_prefix=spm_unigram32000_share
src_lang=en
tgt_lang=de
tokenize=1
splits=(newstest2014 newstest2016)
for split in ${splits[@]}; do
src_file=${data_dir}/${split}.${src_lang}
tgt_file=${data_dir}/${split}.${tgt_lang}
if [[ ${tokenize} -eq 1 ]]; then
cmd="tokenizer.perl -l ${src_lang} --threads 8 -no-escape < ${src_file} > ${src_file}.tok"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
cmd="tokenizer.perl -l ${tgt_lang} --threads 8 -no-escape < ${tgt_file} > ${tgt_file}.tok"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
src_file=${src_file}.tok
tgt_file=${tgt_file}.tok
fi
cmd="cat ${src_file}"
if [[ ${lcrm} -eq 1 ]]; then
cmd="python local/lower_rm.py ${src_file}"
fi
cmd="${cmd}
| spm_encode --model ${vocab_dir}/${src_vocab_prefix}.model
--output_format=piece
> ${src_file}.spm"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
cmd="spm_encode
--model ${vocab_dir}/${tgt_vocab_prefix}.model
--output_format=piece
< ${tgt_file}
> ${tgt_file}.spm"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
src_file=${src_file}.spm
tgt_file=${tgt_file}.spm
mkdir -p ${data_dir}/final
cmd="cp ${src_file} ${data_dir}/final/${split}.${src_lang}"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
cmd="cp ${tgt_file} ${data_dir}/final/${split}.${tgt_lang}"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
done
n_set=${#splits[*]}
for ((i=0;i<$n_set;i++)); do
dataset[$i]=${data_dir}/final/${splits[$i]}
done
pref=`echo ${dataset[*]} | sed 's/ /,/g'`
cmd="python ${root_dir}/fairseq_cli/preprocess.py
--source-lang ${src_lang}
--target-lang ${tgt_lang}
--testpref ${pref}
--destdir ${data_dir}/data-bin
--srcdict ${vocab_dir}/${src_vocab_prefix}.txt
--tgtdict ${vocab_dir}/${tgt_vocab_prefix}.txt
--workers 64"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
\ No newline at end of file
egs/libri_trans/mt/conf/base.yaml 0 → 100644
train-subset: train
valid-subset: valid
max-epoch: 50
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
skip-invalid-size-inputs-valid-test: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
arch: transformer
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 8000
lr: 1e-3
adam_betas: (0.9,0.997)
criterion: label_smoothed_cross_entropy
label_smoothing: 0.1
dropout: 0.1
attention-dropout: 0.1
activation-dropout: 0.1
activation-fn: relu
encoder-normalize-before: True
decoder-normalize-before: True
encoder-embed-dim: 512
encoder-ffn-embed-dim: 2048
encoder-layers: 6
decoder-layers: 6
encoder-attention-heads: 8
decoder-embed-dim: 512
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 8
egs/libri_trans/mt/conf/base_s.yaml 0 → 100644
train-subset: train
valid-subset: valid
max-epoch: 50
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
skip-invalid-size-inputs-valid-test: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
arch: transformer
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 8000
lr: 1e-3
adam_betas: (0.9,0.997)
criterion: label_smoothed_cross_entropy
label_smoothing: 0.1
dropout: 0.1
attention-dropout: 0.1
activation-dropout: 0.1
activation-fn: relu
encoder-normalize-before: True
decoder-normalize-before: True
encoder-embed-dim: 256
encoder-ffn-embed-dim: 2048
encoder-layers: 6
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/mt/conf/dlcl.yaml 0 → 100644
use-enc-dlcl: True
use-dec-dlcl: True
egs/libri_trans/mt/conf/rpr.yaml 0 → 100644
#encoder-attention-type: rel_selfattn
encoder-attention-type: relative
decoder-attention-type: relative
max-encoder-relative-length: 20
max-decoder-relative-length: 20
\ No newline at end of file
egs/libri_trans/mt/decode.sh 0 → 100755
#! /bin/bash
gpu_num=1
data_dir=
test_subset=(test)
exp_name=
if [ "$#" -eq 1 ]; then
exp_name=$1
fi
n_average=10
beam_size=5
len_penalty=1.0
max_tokens=80000
dec_model=checkpoint_best.pt
cmd="./run.sh
--stage 2
--stop_stage 2
--gpu_num ${gpu_num}
--exp_name ${exp_name}
--n_average ${n_average}
--beam_size ${beam_size}
--len_penalty ${len_penalty}
--max_tokens ${max_tokens}
--dec_model ${dec_model}
"
if [[ -n ${data_dir} ]]; then
cmd="$cmd --data_dir ${data_dir}"
fi
if [[ -n ${test_subset} ]]; then
test_subset=`echo ${test_subset[*]} | sed 's/ /,/g'`
cmd="$cmd --test_subset ${test_subset}"
fi
echo $cmd
eval $cmd
egs/libri_trans/mt/local/lower_rm.py 0 → 100644
import sys
import string
in_file = sys.argv[1]
with open(in_file, "r", encoding="utf-8") as f:
for line in f.readlines():
line = line.strip().lower()
for w in string.punctuation:
line = line.replace(w, "")
line = line.replace(" ", "")
print(line)
egs/libri_trans/mt/local/monitor.sh 0 → 100755
gpu_num=4
cmd="sh train.sh"
while :
do
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
count=0
for dev in ${all_devices[@]}
do
line=$((dev + 2))
use=$(head -n $line ${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 [[ ${#device[@]} -lt $gpu_num ]]; then
sleep 60s
else
echo "Run $cmd"
eval $cmd
sleep 10s
exit
fi
done
egs/libri_trans/mt/local/parse_options.sh 0 → 100755
#!/usr/bin/env bash
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey);
# Arnab Ghoshal, Karel Vesely
# 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.
# Parse command-line options.
# To be sourced by another script (as in ". parse_options.sh").
# Option format is: --option-name arg
# and shell variable "option_name" gets set to value "arg."
# The exception is --help, which takes no arguments, but prints the
# $help_message variable (if defined).
###
### The --config file options have lower priority to command line
### options, so we need to import them first...
###
# Now import all the configs specified by command-line, in left-to-right order
for ((argpos=1; argpos<$#; argpos++)); do
if [ "${!argpos}" == "--config" ]; then
argpos_plus1=$((argpos+1))
config=${!argpos_plus1}
[ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
. $config # source the config file.
fi
done
###
### Now we process the command line options
###
while true; do
[ -z "${1:-}" ] && break; # break if there are no arguments
case "$1" in
# If the enclosing script is called with --help option, print the help
# message and exit. Scripts should put help messages in $help_message
--help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
else printf "$help_message\n" 1>&2 ; fi;
exit 0 ;;
--*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
exit 1 ;;
# If the first command-line argument begins with "--" (e.g. --foo-bar),
# then work out the variable name as $name, which will equal "foo_bar".
--*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
# Next we test whether the variable in question is undefned-- if so it's
# an invalid option and we die. Note: $0 evaluates to the name of the
# enclosing script.
# The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
# is undefined. We then have to wrap this test inside "eval" because
# foo_bar is itself inside a variable ($name).
eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
oldval="`eval echo \\$$name`";
# Work out whether we seem to be expecting a Boolean argument.
if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
was_bool=true;
else
was_bool=false;
fi
# Set the variable to the right value-- the escaped quotes make it work if
# the option had spaces, like --cmd "queue.pl -sync y"
eval $name=\"$2\";
# Check that Boolean-valued arguments are really Boolean.
if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
exit 1;
fi
shift 2;
;;
*) break;
esac
done
# Check for an empty argument to the --cmd option, which can easily occur as a
# result of scripting errors.
[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
true; # so this script returns exit code 0.
egs/libri_trans/mt/local/utils.sh 0 → 100755
get_devices(){
gpu_num=$1
use_cpu=$2
device=()
while :
do
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
count=0
for dev in ${all_devices[@]}
do
line=$((dev + 2))
use=$(head -n $line ${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 [[ ${#device[@]} -lt $gpu_num ]]; then
if [[ $use_cpu -eq 1 ]]; then
device=(-1)
else
sleep 60s
fi
else
break
fi
done
echo ${device[*]} | sed 's/ /,/g'
return $?
}
egs/libri_trans/mt/train.sh 0 → 100755
#! /bin/bash
# training the model
gpu_num=1
update_freq=1
max_tokens=8192
exp_tag=baseline
config_list=(base)
# exp full name
exp_name=
extra_tag=
extra_parameter=
#extra_tag="${extra_tag}"
#extra_parameter="${extra_parameter} "
train_config=$(echo ${config_list[*]} | sed 's/ /,/g')
cmd="./run.sh
--stage 1
--stop_stage 1
--gpu_num ${gpu_num}
--update_freq ${update_freq}
--train_config ${train_config}
--max_tokens ${max_tokens}
"
if [[ -n ${exp_name} ]]; then
cmd="$cmd --exp_name ${exp_name}"
fi
if [[ -n ${exp_tag} ]]; then
cmd="$cmd --exp_tag ${exp_tag}"
fi
if [[ -n ${extra_tag} ]]; then
cmd="$cmd --extra_tag ${extra_tag}"
fi
if [[ -n ${extra_parameter} ]]; then
cmd="$cmd --extra_parameter \"${extra_parameter}\""
fi
echo ${cmd}
eval ${cmd}
egs/libri_trans/st/binary.sh 0 → 100644
set -e
eval=1
lcrm=1
tokenizer=0
root_dir=~/st/Fairseq-S2T
data_dir=/home/xuchen/st/data/test
vocab_dir=/home/xuchen/st/data/mustc/st_lcrm/en-de
asr_vocab_prefix=spm_unigram10000_st_share
st_vocab_prefix=spm_unigram10000_st_share
src_lang=en
tgt_lang=de
splits=(2019)
splits=$(echo ${splits[*]} | sed 's/ /_/g')
cp -r ${vocab_dir}/${asr_vocab_prefix}.* ${data_dir}/${src_lang}-${tgt_lang}
cp -r ${vocab_dir}/${st_vocab_prefix}.* ${data_dir}/${src_lang}-${tgt_lang}
rm -rf ${data_dir}/${src_lang}-${tgt_lang}/fbank80.zip
cmd="python ${root_dir}/examples/speech_to_text/prep_st_data.py
--data-root ${data_dir}
--output-root ${data_dir}
--splits ${splits}
--task st
--src-lang ${src_lang}
--tgt-lang ${tgt_lang}
--add-src
--share
--asr-prefix ${asr_vocab_prefix}
--st-spm-prefix ${st_vocab_prefix}
--cmvn-type utterance"
if [[ ${lcrm} -eq 1 ]]; then
cmd="$cmd
--lowercase-src
--rm-punc-src"
fi
if [[ ${tokenizer} -eq 1 ]]; then
cmd="$cmd
--tokenizer"
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
egs/libri_trans/st/conf/base.yaml 0 → 100644
train-subset: train_st
valid-subset: dev_st
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
arch: s2t_transformer_s
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
conv-kernel-sizes: 5,5
conv-channels: 1024
dropout: 0.1
activation-fn: relu
encoder-embed-dim: 256
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
attention-dropout: 0.1
activation-dropout: 0.1
egs/libri_trans/st/conf/conformer.yaml 0 → 100644
macaron-style: True
use-cnn-module: True
cnn-module-kernel: 31
egs/libri_trans/st/conf/ctc.yaml 0 → 100644
ctc-weight: 0.3
\ No newline at end of file
egs/libri_trans/st/conf/dlcl.yaml 0 → 100644
use-enc-dlcl: True
use-dec-dlcl: True
egs/libri_trans/st/conf/local_attn.yaml 0 → 100644
encoder-attention-type: local
hard-mask-window: 0
gauss-mask-sigma: 3
init-mask-weight: 0
\ No newline at end of file
egs/libri_trans/st/conf/pds_base.yaml 0 → 100644
arch: pdss2t_transformer_s_8
train-subset: train_st
valid-subset: dev_st
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/st/conf/pds_base_16.yaml 0 → 100644
arch: pdss2t_transformer_s_16
encoder-embed-dim: 256
pyramid-stages: 4
#pyramid-dropout: 0
pyramid-layers: 2_2_6_2
pyramid-ratios: 2_2_2_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1
pyramid-kernel-sizes: 5_5_5_5
pyramid-ffn-ratios: 8_8_8_8
pyramid-attn-heads: 4_4_4_4
train-subset: train_st
valid-subset: dev_st
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/st/conf/pds_base_32.yaml 0 → 100644
arch: pdss2t_transformer_s_32
encoder-embed-dim: 256
pyramid-stages: 5
#pyramid-dropout: 0
pyramid-layers: 2_2_3_3_2
pyramid-ratios: 2_2_2_2_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1_1
pyramid-kernel-sizes: 5_5_5_5_5
pyramid-ffn-ratios: 8_8_8_8_8
pyramid-attn-heads: 4_4_4_4_4
train-subset: train_st
valid-subset: dev_st
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/st/conf/pds_base_8.yaml 0 → 100644
arch: pdss2t_transformer_s_8
encoder-embed-dim: 256
pyramid-stages: 4
#pyramid-dropout: 0
pyramid-layers: 3_3_3_3
pyramid-ratios: 2_2_1_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1
pyramid-kernel-sizes: 5_5_5_5
pyramid-ffn-ratios: 8_8_8_8
pyramid-attn-heads: 4_4_4_4
train-subset: train_st
valid-subset: dev_st
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-decoder-from:
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
dropout: 0.1
activation-fn: relu
encoder-ffn-embed-dim: 2048
encoder-layers: 12
decoder-layers: 6
encoder-attention-heads: 4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/st/conf/rpr.yaml 0 → 100644
encoder-attention-type: rel_selfattn
#encoder-attention-type: relative
#decoder-attention-type: relative
#max-encoder-relative-length: 100
#max-decoder-relative-length: 20
egs/libri_trans/st/conf/sate_ctc.yaml 0 → 100644
train-subset: train_st
valid-subset: dev_st
max-epoch: 100
max-update: 100000
num-workers: 8
patience: 10
no-progress-bar: True
log-interval: 100
seed: 1
report-accuracy: True
#load-pretrained-encoder-from:
#load-pretrained-acoustic-encoder-from:
#load-pretrained-text-encoder-from:
#load-pretrained-decoder-from:
arch: s2t_sate
share-decoder-input-output-embed: True
optimizer: adam
clip-norm: 10.0
lr-scheduler: inverse_sqrt
warmup-init-lr: 1e-7
warmup-updates: 10000
lr: 2e-3
#adam_betas: (0.9,0.98)
ctc-weight: 0.3
criterion: label_smoothed_cross_entropy_with_ctc
label_smoothing: 0.1
encoder-normalize-before: True
decoder-normalize-before: True
conv-kernel-sizes: 5,5
conv-channels: 1024
dropout: 0.1
activation-fn: relu
encoder-embed-dim: 256
encoder-ffn-embed-dim: 2048
encoder-layers: 12
text-encoder-layers: 6
decoder-layers: 6
encoder-attention-heads: 4
#macaron-style: True
#use-cnn-module: True
#cnn-module-kernel: 31
#acoustic-encoder: pds
acoustic-encoder: transformer
adapter: league
encoder-embed-dim: 256
pyramid-stages: 4
#pyramid-dropout: 0
pyramid-layers: 3_3_3_3
pyramid-ratios: 2_2_1_2
pyramid-fusion: True
pyramid-fusion-method: all_conv
pyramid-embed-dims: 256_256_256_256
pyramid-ds-method: conv
pyramid-embed-norm: True
pyramid-position-embed: 1_1_1_1
pyramid-kernel-sizes: 5_5_5_5
pyramid-ffn-ratios: 8_8_8_8
pyramid-attn-heads: 4_4_4_4
decoder-embed-dim: 256
decoder-ffn-embed-dim: 2048
decoder-attention-heads: 4
egs/libri_trans/st/decode.sh 0 → 100755
#! /bin/bash
gpu_num=1
data_dir=
test_subset=(tst-COMMON)
exp_name=
if [ "$#" -eq 1 ]; then
exp_name=$1
fi
n_average=10
beam_size=5
len_penalty=1.0
max_tokens=80000
dec_model=checkpoint_best.pt
cmd="./run.sh
--stage 2
--stop_stage 2
--gpu_num ${gpu_num}
--exp_name ${exp_name}
--n_average ${n_average}
--beam_size ${beam_size}
--len_penalty ${len_penalty}
--max_tokens ${max_tokens}
--dec_model ${dec_model}
"
if [[ -n ${data_dir} ]]; then
cmd="$cmd --data_dir ${data_dir}"
fi
if [[ ${#test_subset[@]} -eq 0 ]]; then
subsets=$(echo ${test_subset[*]} | sed 's/ /,/g')
cmd="$cmd --test_subset ${subsets}"
fi
echo $cmd
eval $cmd
egs/libri_trans/st/ensemble.sh 0 → 100644
set -e
gpu_num=1
root_dir=/home/xuchen/st/Fairseq-S2T
ckpt=/home/xuchen/st/checkpoints/mustc-v2/st
model_txt=$1
set=$2
test_subset=$3
#data_dir=/home/xuchen/st/data/mustc-v2/st_lcrm/en-de
#test_subset=(tst-COMMON)
data_dir=/media/data/tst/$set/en-de
#test_subset=(office)
#test_subset=(webrtc1)
#test_subset=(adap2)
data_config=config_st_share.yaml
result_file=./result
beam_size=5
lenpen=0.6
max_tokens=10000
models=()
i=0
for line in `cat $model_txt`; do
i=`expr $i + 1`
model_dir=$ckpt/$line
[[ ! -d $model_dir ]] && echo $model_dir && exit 1;
if [[ -f $model_dir/avg_10_checkpoint.pt ]]; then
model=$model_dir/avg_10_checkpoint.pt
else
model=$model_dir/checkpoint_best.pt
fi
[[ ! -f $model ]] && echo $model && exit 1;
models[$i]=$model
done
models=`echo ${models[*]} | sed 's/ /:/g'`
res_dir=$ckpt/ensemble/$set
i=0
while :
do
if [[ -d $res_dir/$i ]]; then
i=`expr $i + 1`
else
res_dir=$res_dir/$i
break
fi
done
mkdir -p $res_dir
cp $model_txt $res_dir
if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
if [[ ${gpu_num} -eq 0 ]]; then
device=()
else
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
fi
export CUDA_VISIBLE_DEVICES=${device}
for subset in ${test_subset[@]}; do
subset=${subset}_st
cmd="python ${root_dir}/fairseq_cli/generate.py
${data_dir}
--config-yaml ${data_config}
--gen-subset ${subset}
--task speech_to_text
--path ${models}
--results-path ${res_dir}
--skip-invalid-size-inputs-valid-test
--max-tokens ${max_tokens}
--beam ${beam_size}
--lenpen ${lenpen}
--scoring sacrebleu"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
eval $cmd
tail -n 1 ${res_dir}/generate-${subset}.txt
cd $res_dir
evaluate.sh translation-${subset}.txt $set
cd -
done
egs/libri_trans/st/local/monitor.sh 0 → 100755
gpu_num=4
cmd="sh train.sh"
while :
do
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
count=0
for dev in ${all_devices[@]}
do
line=$((dev + 2))
use=$(head -n $line ${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 [[ ${#device[@]} -lt $gpu_num ]]; then
sleep 60s
else
echo "Run $cmd"
eval $cmd
sleep 10s
exit
fi
done
egs/libri_trans/st/local/parse_options.sh 0 → 100755
#!/usr/bin/env bash
# Copyright 2012 Johns Hopkins University (Author: Daniel Povey);
# Arnab Ghoshal, Karel Vesely
# 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.
# Parse command-line options.
# To be sourced by another script (as in ". parse_options.sh").
# Option format is: --option-name arg
# and shell variable "option_name" gets set to value "arg."
# The exception is --help, which takes no arguments, but prints the
# $help_message variable (if defined).
###
### The --config file options have lower priority to command line
### options, so we need to import them first...
###
# Now import all the configs specified by command-line, in left-to-right order
for ((argpos=1; argpos<$#; argpos++)); do
if [ "${!argpos}" == "--config" ]; then
argpos_plus1=$((argpos+1))
config=${!argpos_plus1}
[ ! -r $config ] && echo "$0: missing config '$config'" && exit 1
. $config # source the config file.
fi
done
###
### Now we process the command line options
###
while true; do
[ -z "${1:-}" ] && break; # break if there are no arguments
case "$1" in
# If the enclosing script is called with --help option, print the help
# message and exit. Scripts should put help messages in $help_message
--help|-h) if [ -z "$help_message" ]; then echo "No help found." 1>&2;
else printf "$help_message\n" 1>&2 ; fi;
exit 0 ;;
--*=*) echo "$0: options to scripts must be of the form --name value, got '$1'"
exit 1 ;;
# If the first command-line argument begins with "--" (e.g. --foo-bar),
# then work out the variable name as $name, which will equal "foo_bar".
--*) name=`echo "$1" | sed s/^--// | sed s/-/_/g`;
# Next we test whether the variable in question is undefned-- if so it's
# an invalid option and we die. Note: $0 evaluates to the name of the
# enclosing script.
# The test [ -z ${foo_bar+xxx} ] will return true if the variable foo_bar
# is undefined. We then have to wrap this test inside "eval" because
# foo_bar is itself inside a variable ($name).
eval '[ -z "${'$name'+xxx}" ]' && echo "$0: invalid option $1" 1>&2 && exit 1;
oldval="`eval echo \\$$name`";
# Work out whether we seem to be expecting a Boolean argument.
if [ "$oldval" == "true" ] || [ "$oldval" == "false" ]; then
was_bool=true;
else
was_bool=false;
fi
# Set the variable to the right value-- the escaped quotes make it work if
# the option had spaces, like --cmd "queue.pl -sync y"
eval $name=\"$2\";
# Check that Boolean-valued arguments are really Boolean.
if $was_bool && [[ "$2" != "true" && "$2" != "false" ]]; then
echo "$0: expected \"true\" or \"false\": $1 $2" 1>&2
exit 1;
fi
shift 2;
;;
*) break;
esac
done
# Check for an empty argument to the --cmd option, which can easily occur as a
# result of scripting errors.
[ ! -z "${cmd+xxx}" ] && [ -z "$cmd" ] && echo "$0: empty argument to --cmd option" 1>&2 && exit 1;
true; # so this script returns exit code 0.
egs/libri_trans/st/local/utils.sh 0 → 100755
get_devices(){
gpu_num=$1
use_cpu=$2
device=()
while :
do
record=$(mktemp -t temp.record.XXXXXX)
gpustat > $record
all_devices=$(seq 0 "$(sed '1,2d' ${record} | wc -l)");
count=0
for dev in ${all_devices[@]}
do
line=$((dev + 2))
use=$(head -n $line ${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 [[ ${#device[@]} -lt $gpu_num ]]; then
if [[ $use_cpu -eq 1 ]]; then
device=(-1)
else
sleep 60s
fi
else
break
fi
done
echo ${device[*]} | sed 's/ /,/g'
return $?
}
egs/libri_trans/st/run.sh 0 → 100755
#! /bin/bash
# Processing MuST-C Datasets
# Copyright 2021 Natural Language Processing Laboratory
# Xu Chen (xuchenneu@163.com)
# Set bash to 'debug' mode, it will exit on :
# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
set -e
#set -u
set -o pipefail
export PYTHONIOENCODING=UTF-8
eval=1
time=$(date "+%m%d_%H%M")
stage=0
stop_stage=0
######## hardware ########
# devices
#device=()
gpu_num=8
update_freq=1
root_dir=~/st/Fairseq-S2T
pwd_dir=$PWD
# dataset
src_lang=en
tgt_lang=de
lang=${src_lang}-${tgt_lang}
dataset=st
task=speech_to_text
vocab_type=unigram
asr_vocab_size=5000
vocab_size=10000
share_dict=1
speed_perturb=0
lcrm=0
tokenizer=0
use_specific_dict=0
specific_prefix=valid
specific_dir=/home/xuchen/st/data/mustc/st_lcrm/en-de
asr_vocab_prefix=spm_unigram10000_st_share
st_vocab_prefix=spm_unigram10000_st_share
org_data_dir=~/st/data/${dataset}
data_dir=~/st/data/${dataset}/st
test_subset=tst-COMMON
# exp
exp_prefix=$(date "+%m%d")
extra_tag=
extra_parameter=
exp_tag=baseline
exp_name=
# config
train_config=ctc
# training setting
fp16=1
max_tokens=40000
step_valid=0
bleu_valid=0
# decoding setting
dec_model=checkpoint_best.pt
n_average=10
beam_size=5
len_penalty=1.0
if [[ ${share_dict} -eq 1 ]]; then
data_config=config_st_share.yaml
else
data_config=config_st.yaml
fi
if [[ ${speed_perturb} -eq 1 ]]; then
data_dir=${data_dir}_sp
exp_prefix=${exp_prefix}_sp
fi
if [[ ${lcrm} -eq 1 ]]; then
data_dir=${data_dir}_lcrm
exp_prefix=${exp_prefix}_lcrm
fi
if [[ ${use_specific_dict} -eq 1 ]]; then
data_dir=${data_dir}_${specific_prefix}
exp_prefix=${exp_prefix}_${specific_prefix}
fi
if [[ ${tokenizer} -eq 1 ]]; then
data_dir=${data_dir}_tok
exp_prefix=${exp_prefix}_tok
fi
. ./local/parse_options.sh || exit 1;
if [[ -z ${exp_name} ]]; then
config_string=${train_config//,/_}
exp_name=${exp_prefix}_${config_string}_${exp_tag}
if [[ -n ${extra_tag} ]]; then
exp_name=${exp_name}_${extra_tag}
fi
fi
model_dir=$root_dir/../checkpoints/$dataset/st/${exp_name}
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
echo "stage -1: Data Download"
# pass
fi
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
### Task dependent. You have to make data the following preparation part by yourself.
### But you can utilize Kaldi recipes in most cases
echo "stage 0: ASR Data Preparation"
if [[ ! -e ${data_dir}/${lang} ]]; then
mkdir -p ${data_dir}/${lang}
fi
cmd="python ${root_dir}/examples/speech_to_text/prep_asr_data.py
--data-root ${org_data_dir}
--output-root ${data_dir}
--task asr
--vocab-type ${vocab_type}
--vocab-size ${asr_vocab_size}"
if [[ ${speed_perturb} -eq 1 ]]; then
cmd="$cmd
--speed-perturb"
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 && ${share_dict} -ne 1 && ${use_specific_dict} -ne 1 ]] && eval $cmd
asr_prefix=spm_${vocab_type}${asr_vocab_size}_asr
echo "stage 0: ST Data Preparation"
cmd="python ${root_dir}/examples/speech_to_text/prep_st_data.py
--data-root ${org_data_dir}
--output-root ${data_dir}
--task st
--add-src
--cmvn-type utterance
--vocab-type ${vocab_type}
--vocab-size ${vocab_size}"
if [[ ${use_specific_dict} -eq 1 ]]; then
cp -r ${specific_dir}/${asr_vocab_prefix}.* ${data_dir}/${lang}
cp -r ${specific_dir}/${st_vocab_prefix}.* ${data_dir}/${lang}
if [[ $share_dict -eq 1 ]]; then
cmd="$cmd
--share
--st-spm-prefix ${st_vocab_prefix}"
else
cmd="$cmd
--st-spm-prefix ${st_vocab_prefix}
--asr-prefix ${asr_vocab_prefix}"
fi
else
if [[ $share_dict -eq 1 ]]; then
cmd="$cmd
--share"
else
cmd="$cmd
--asr-prefix ${asr_prefix}"
fi
fi
if [[ ${speed_perturb} -eq 1 ]]; then
cmd="$cmd
--speed-perturb"
fi
if [[ ${lcrm} -eq 1 ]]; then
cmd="$cmd
--lowercase-src
--rm-punc-src"
fi
if [[ ${tokenizer} -eq 1 ]]; then
cmd="$cmd
--tokenizer"
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval ${cmd}
fi
data_dir=${data_dir}/${lang}
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "stage 1: ST Network Training"
[[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
if [[ ${gpu_num} -eq 0 ]]; then
device=""
else
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
fi
echo -e "dev=${device} data=${data_dir} model=${model_dir}"
if [[ ! -d ${model_dir} ]]; then
mkdir -p ${model_dir}
else
echo "${model_dir} exists."
fi
cp ${BASH_SOURCE[0]} ${model_dir}
cp ${PWD}/train.sh ${model_dir}
config_list="${train_config//,/ }"
idx=0
for config in ${config_list[@]}
do
config_path=$pwd_dir/conf/${config}.yaml
if [[ ! -f ${config_path} ]]; then
echo "No config file ${config_path}"
exit
fi
cp ${config_path} ${model_dir}
if [[ idx -eq 0 ]]; then
extra_parameter="${extra_parameter}
--train-config ${config_path}"
else
extra_parameter="${extra_parameter}
--train-config${idx} ${config_path}"
fi
idx=$((idx + 1))
done
cmd="python3 -u ${root_dir}/fairseq_cli/train.py
${data_dir}
--config-yaml ${data_config}
--task ${task}
--max-tokens ${max_tokens}
--skip-invalid-size-inputs-valid-test
--update-freq ${update_freq}
--log-interval 100
--save-dir ${model_dir}
--tensorboard-logdir ${model_dir}"
if [[ -n ${extra_parameter} ]]; then
cmd="${cmd}
${extra_parameter}"
fi
if [[ ${gpu_num} -gt 0 ]]; then
cmd="${cmd}
--distributed-world-size $gpu_num
--ddp-backend no_c10d"
fi
if [[ $fp16 -eq 1 ]]; then
cmd="${cmd}
--fp16"
fi
if [[ $step_valid -eq 1 ]]; then
validate_interval=1
save_interval=1
keep_last_epochs=10
no_epoch_checkpoints=0
save_interval_updates=500
keep_interval_updates=10
else
validate_interval=1
keep_last_epochs=10
fi
if [[ $bleu_valid -eq 1 ]]; then
cmd="$cmd
--eval-bleu
--eval-bleu-args '{\"beam\": 1}'
--eval-tokenized-bleu
--eval-bleu-remove-bpe
--best-checkpoint-metric bleu
--maximize-best-checkpoint-metric"
fi
if [[ -n $no_epoch_checkpoints && $no_epoch_checkpoints -eq 1 ]]; then
cmd="$cmd
--no-epoch-checkpoints"
fi
if [[ -n $validate_interval ]]; then
cmd="${cmd}
--validate-interval $validate_interval "
fi
if [[ -n $save_interval ]]; then
cmd="${cmd}
--save-interval $save_interval "
fi
if [[ -n $keep_last_epochs ]]; then
cmd="${cmd}
--keep-last-epochs $keep_last_epochs "
fi
if [[ -n $save_interval_updates ]]; then
cmd="${cmd}
--save-interval-updates $save_interval_updates"
if [[ -n $keep_interval_updates ]]; then
cmd="${cmd}
--keep-interval-updates $keep_interval_updates"
fi
fi
echo -e "\033[34mRun command: \n${cmd} \033[0m"
# save info
log=./history.log
echo "${time} | ${device} | ${data_dir} | ${exp_name} | ${model_dir} " >> $log
tail -n 50 ${log} > tmp.log
mv tmp.log $log
export CUDA_VISIBLE_DEVICES=${device}
cmd="nohup ${cmd} >> ${model_dir}/train.log 2>&1 &"
if [[ $eval -eq 1 ]]; then
eval $cmd
sleep 2s
tail -n "$(wc -l ${model_dir}/train.log | awk '{print $1+1}')" -f ${model_dir}/train.log
fi
fi
wait
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
echo "stage 2: ST Decoding"
if [[ ${n_average} -ne 1 ]]; then
# Average models
dec_model=avg_${n_average}_checkpoint.pt
cmd="python ${root_dir}/scripts/average_checkpoints.py
--inputs ${model_dir}
--num-epoch-checkpoints ${n_average}
--output ${model_dir}/${dec_model}"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
[[ $eval -eq 1 ]] && eval $cmd
else
dec_model=${dec_model}
fi
if [[ -z ${device} || ${#device[@]} -eq 0 ]]; then
if [[ ${gpu_num} -eq 0 ]]; then
device=""
else
source ./local/utils.sh
device=$(get_devices $gpu_num 0)
fi
fi
export CUDA_VISIBLE_DEVICES=${device}
result_file=${model_dir}/decode_result
[[ -f ${result_file} ]] && rm ${result_file}
test_subset=${test_subset//,/ }
for subset in "${test_subset[@]}"; do
subset=${subset}_st
cmd="python ${root_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}
--max-tokens ${max_tokens}
--beam ${beam_size}
--lenpen ${len_penalty}
--scoring sacrebleu"
echo -e "\033[34mRun command: \n${cmd} \033[0m"
if [[ $eval -eq 1 ]]; then
eval $cmd
tail -n 1 ${model_dir}/generate-${subset}.txt >> ${result_file}
fi
done
cat ${result_file}
fi
egs/libri_trans/st/train.sh 0 → 100755
#! /bin/bash
# training the model
gpu_num=8
update_freq=1
max_tokens=40000
extra_tag=
extra_parameter=
#extra_tag="${extra_tag}"
#extra_parameter="${extra_parameter} "
exp_tag=
#config_list=(base)
config_list=(ctc)
#config_list=(sate_ctc)
#config_list=(ctc conformer rpr)
#config_list=(base sate)
#config_list=(pds_base)
#config_list=(pds_base conformer)
# exp full name
exp_name=
train_config=$(echo ${config_list[*]} | sed 's/ /,/g')
cmd="./run.sh
--stage 1
--stop_stage 1
--gpu_num ${gpu_num}
--update_freq ${update_freq}
--train_config ${train_config}
--max_tokens ${max_tokens}
"
if [[ -n ${exp_name} ]]; then
cmd="$cmd --exp_name ${exp_name}"
fi
if [[ -n ${exp_tag} ]]; then
cmd="$cmd --exp_tag ${exp_tag}"
fi
if [[ -n ${extra_tag} ]]; then
cmd="$cmd --extra_tag ${extra_tag}"
fi
if [[ -n ${extra_parameter} ]]; then
cmd="$cmd --extra_parameter \"${extra_parameter}\""
fi
echo ${cmd}
eval ${cmd}
egs/mustc/asr/run.sh
...@@ -39,6 +39,7 @@ vocab_size=5000 ...@@ -39,6 +39,7 @@ vocab_size=5000
speed_perturb=0 speed_perturb=0
lcrm=0 lcrm=0
tokenizer=0 tokenizer=0
use_raw_audio=0
use_specific_dict=0 use_specific_dict=0
specific_prefix=st specific_prefix=st
...@@ -87,6 +88,10 @@ if [[ ${tokenizer} -eq 1 ]]; then ...@@ -87,6 +88,10 @@ if [[ ${tokenizer} -eq 1 ]]; then
data_dir=${data_dir}_tok data_dir=${data_dir}_tok
exp_prefix=${exp_prefix}_tok exp_prefix=${exp_prefix}_tok
fi fi
if [[ ${use_raw_audio} -eq 1 ]]; then
data_dir=${data_dir}_raw
exp_prefix=${exp_prefix}_raw
fi
. ./local/parse_options.sh || exit 1; . ./local/parse_options.sh || exit 1;
...@@ -109,9 +114,6 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then ...@@ -109,9 +114,6 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
### Task dependent. You have to make data the following preparation part by yourself. ### Task dependent. You have to make data the following preparation part by yourself.
### But you can utilize Kaldi recipes in most cases ### But you can utilize Kaldi recipes in most cases
echo "stage 0: ASR Data Preparation" echo "stage 0: ASR Data Preparation"
if [[ ! -e ${data_dir}/${lang} ]]; then
mkdir -p ${data_dir}/${lang}
fi
cmd="python ${root_dir}/examples/speech_to_text/prep_mustc_data.py cmd="python ${root_dir}/examples/speech_to_text/prep_mustc_data.py
--data-root ${org_data_dir} --data-root ${org_data_dir}
...@@ -120,6 +122,10 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then ...@@ -120,6 +122,10 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
--vocab-type ${vocab_type} --vocab-type ${vocab_type}
--vocab-size ${vocab_size}" --vocab-size ${vocab_size}"
if [[ ${use_raw_audio} -eq 1 ]]; then
cmd="$cmd
--raw"
fi
if [[ ${use_specific_dict} -eq 1 ]]; then if [[ ${use_specific_dict} -eq 1 ]]; then
cp -r ${specific_dir}/${asr_vocab_prefix}.* ${data_dir}/${lang} cp -r ${specific_dir}/${asr_vocab_prefix}.* ${data_dir}/${lang}
cmd="$cmd cmd="$cmd
...@@ -143,8 +149,6 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then ...@@ -143,8 +149,6 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
[[ $eval -eq 1 ]] && eval ${cmd} [[ $eval -eq 1 ]] && eval ${cmd}
fi fi
data_dir=${data_dir}/${lang}
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
echo "stage 1: ASR Network Training" echo "stage 1: ASR Network Training"
[[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1; [[ ! -d ${data_dir} ]] && echo "The data dir ${data_dir} is not existing!" && exit 1;
......
egs/template/asr/run.sh
...@@ -36,7 +36,7 @@ task=speech_to_text ...@@ -36,7 +36,7 @@ task=speech_to_text
vocab_type=unigram vocab_type=unigram
vocab_size=5000 vocab_size=5000
speed_perturb=0 speed_perturb=0
lcrm=0 lcrm=1
tokenizer=0 tokenizer=0
use_specific_dict=0 use_specific_dict=0
...@@ -47,7 +47,7 @@ asr_vocab_prefix=spm_unigram10000_st_share ...@@ -47,7 +47,7 @@ asr_vocab_prefix=spm_unigram10000_st_share
org_data_dir=~/st/data/${dataset} org_data_dir=~/st/data/${dataset}
data_dir=~/st/data/${dataset}/asr data_dir=~/st/data/${dataset}/asr
train_split=train train_split=train
valid_split=valid valid_split=dev
test_split=test test_split=test
test_subset=test test_subset=test
......
examples/speech_to_text/prep_mt_data.py
...@@ -33,7 +33,7 @@ class MTDataset(Dataset): ...@@ -33,7 +33,7 @@ class MTDataset(Dataset):
""" """
def __init__(self, root: str, src_lang, tgt_lang: str, split: str) -> None: def __init__(self, root: str, src_lang, tgt_lang: str, split: str) -> None:
_root = Path(root) / f"{src_lang}-{tgt_lang}" / "data" _root = Path(root) / "data"
txt_root = _root txt_root = _root
assert _root.is_dir() and txt_root.is_dir(), (_root, txt_root) assert _root.is_dir() and txt_root.is_dir(), (_root, txt_root)
# Load source and target text # Load source and target text
......
examples/translation/prepare-iwslt14.sh 100644 → 100755
File mode changed from 100644 to 100755
examples/translation/prepare-iwslt17-multilingual.sh 100644 → 100755
File mode changed from 100644 to 100755
examples/translation/prepare-wmt14en2de.sh 100644 → 100755
File mode changed from 100644 to 100755
examples/translation/prepare-wmt14en2fr.sh 100644 → 100755
File mode changed from 100644 to 100755
fairseq/data/audio/audio_utils.py
...@@ -2,7 +2,8 @@ import os.path as op ...@@ -2,7 +2,8 @@ import os.path as op
from typing import BinaryIO, Optional, Tuple, Union from typing import BinaryIO, Optional, Tuple, Union
import numpy as np import numpy as np
import torch
import torchaudio
def get_waveform( def get_waveform(
path_or_fp: Union[str, BinaryIO], normalization=True path_or_fp: Union[str, BinaryIO], normalization=True
...@@ -18,14 +19,12 @@ def get_waveform( ...@@ -18,14 +19,12 @@ def get_waveform(
if ext not in {".flac", ".wav"}: if ext not in {".flac", ".wav"}:
raise ValueError(f"Unsupported audio format: {ext}") raise ValueError(f"Unsupported audio format: {ext}")
try: waveform, sample_rate = torchaudio.load(path_or_fp)
import soundfile as sf waveform = waveform.squeeze().numpy()
except ImportError:
raise ImportError("Please install soundfile to load WAV/FLAC file")
waveform, sample_rate = sf.read(path_or_fp, dtype="float32")
if not normalization: if not normalization:
waveform *= 2 ** 15 # denormalized to 16-bit signed integers waveform *= 2 ** 15 # denormalized to 16-bit signed integers
return waveform, sample_rate return waveform, sample_rate
...@@ -54,10 +53,10 @@ def _get_kaldi_fbank(waveform, sample_rate, n_bins=80) -> Optional[np.ndarray]: ...@@ -54,10 +53,10 @@ def _get_kaldi_fbank(waveform, sample_rate, n_bins=80) -> Optional[np.ndarray]:
def _get_torchaudio_fbank(waveform, sample_rate, n_bins=80) -> Optional[np.ndarray]: def _get_torchaudio_fbank(waveform, sample_rate, n_bins=80) -> Optional[np.ndarray]:
"""Get mel-filter bank features via TorchAudio.""" """Get mel-filter bank features via TorchAudio."""
try: try:
import torch
import torchaudio.compliance.kaldi as ta_kaldi import torchaudio.compliance.kaldi as ta_kaldi
import torchaudio.sox_effects as ta_sox import torchaudio.sox_effects as ta_sox
if not isinstance(waveform, torch.Tensor):
waveform = torch.from_numpy(waveform) waveform = torch.from_numpy(waveform)
if len(waveform.shape) == 1: if len(waveform.shape) == 1:
# Mono channel: D -> 1 x D # Mono channel: D -> 1 x D
...@@ -91,3 +90,21 @@ def get_fbank(path_or_fp: Union[str, BinaryIO], n_bins=80) -> np.ndarray: ...@@ -91,3 +90,21 @@ def get_fbank(path_or_fp: Union[str, BinaryIO], n_bins=80) -> np.ndarray:
) )
return features return features
def get_fbank_with_perturb(waveform, sample_rate=16000, n_bins=80):
import random
speed = random.choice([0.9, 1.0, 1.1])
if speed != 1.0:
waveform = torch.from_numpy(waveform).float().unsqueeze(0)
waveform, _ = torchaudio.sox_effects.apply_effects_tensor(
waveform, sample_rate,
[['speed', str(speed)], ['rate', str(sample_rate)]])
waveform = waveform.squeeze()
features = _get_kaldi_fbank(waveform, sample_rate)
if features is None:
features = _get_torchaudio_fbank(waveform, sample_rate, n_bins)
return features
fairseq/data/audio/speech_to_text_dataset.py
...@@ -19,7 +19,7 @@ from fairseq.data import ( ...@@ -19,7 +19,7 @@ from fairseq.data import (
ResamplingDataset, ResamplingDataset,
data_utils as fairseq_data_utils, data_utils as fairseq_data_utils,
) )
from fairseq.data.audio.audio_utils import get_fbank, get_waveform from fairseq.data.audio.audio_utils import get_fbank, get_waveform, get_fbank_with_perturb
from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform from fairseq.data.audio.feature_transforms import CompositeAudioFeatureTransform
...@@ -117,6 +117,12 @@ class S2TDataConfig(object): ...@@ -117,6 +117,12 @@ class S2TDataConfig(object):
return self.config.get("use_audio_input", False) return self.config.get("use_audio_input", False)
@property @property
def speed_perturb(self):
"""Needed by the dataset loader to see if the model requires
speed perturbation."""
return self.config.get("speed_perturb", False)
@property
def audio_root(self): def audio_root(self):
"""Audio paths in the manifest TSV can be relative and this provides """Audio paths in the manifest TSV can be relative and this provides
the root path. Set this to empty string when using absolute paths.""" the root path. Set this to empty string when using absolute paths."""
...@@ -253,6 +259,7 @@ class SpeechToTextDataset(FairseqDataset): ...@@ -253,6 +259,7 @@ class SpeechToTextDataset(FairseqDataset):
): ):
self.split, self.is_train_split = split, is_train_split self.split, self.is_train_split = split, is_train_split
self.data_cfg = data_cfg self.data_cfg = data_cfg
self.speed_perturb = data_cfg.speed_perturb
self.audio_paths, self.n_frames = audio_paths, n_frames self.audio_paths, self.n_frames = audio_paths, n_frames
self.n_samples = len(audio_paths) self.n_samples = len(audio_paths)
if data_cfg.share_src_and_tgt: if data_cfg.share_src_and_tgt:
...@@ -317,8 +324,12 @@ class SpeechToTextDataset(FairseqDataset): ...@@ -317,8 +324,12 @@ class SpeechToTextDataset(FairseqDataset):
self, index: int self, index: int
) -> Tuple[int, torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]: ) -> Tuple[int, torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
source = get_features_or_waveform( source = get_features_or_waveform(
self.audio_paths[index], need_waveform=self.data_cfg.use_audio_input self.audio_paths[index],
need_waveform=self.data_cfg.use_audio_input or (self.is_train_split and self.speed_perturb)
) )
if self.speed_perturb:
source = get_fbank_with_perturb(source[0], source[1])
if self.feature_transforms is not None: if self.feature_transforms is not None:
assert not self.data_cfg.use_audio_input assert not self.data_cfg.use_audio_input
source = self.feature_transforms(source) source = self.feature_transforms(source)
......
fairseq/tasks/speech_to_text.py
...@@ -32,6 +32,12 @@ class SpeechToTextTask(LegacyFairseqTask): ...@@ -32,6 +32,12 @@ class SpeechToTextTask(LegacyFairseqTask):
help="Configuration YAML filename (under manifest root)", help="Configuration YAML filename (under manifest root)",
) )
parser.add_argument( parser.add_argument(
"--speed-perturb",
default=False,
action="store_true",
help="use online speed perturbation",
)
parser.add_argument(
"--max-source-positions", "--max-source-positions",
default=6000, default=6000,
type=int, type=int,
...@@ -50,7 +56,9 @@ class SpeechToTextTask(LegacyFairseqTask): ...@@ -50,7 +56,9 @@ class SpeechToTextTask(LegacyFairseqTask):
super().__init__(args) super().__init__(args)
self.src_dict = src_dict self.src_dict = src_dict
self.tgt_dict = tgt_dict self.tgt_dict = tgt_dict
self.speed_perturb = args.speed_perturb
self.data_cfg = S2TDataConfig(op.join(args.data, args.config_yaml)) self.data_cfg = S2TDataConfig(op.join(args.data, args.config_yaml))
self.data_cfg.config["speed_perturb"] = self.speed_perturb
@classmethod @classmethod
def setup_task(cls, args, **kwargs): def setup_task(cls, args, **kwargs):
......
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