fix the bugs and use float32 for softmax

egs/iwslt2022/asr/binary.sh

+set -e
+
+eval=1
+
+lcrm=1
+tokenizer=0
+
+vocab_type=unigram
+vocab_size=5000
+use_raw_audio=0
+speed_perturb=0
+
+dataset=iwslt2022
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/asr
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+src_lang=en
+tgt_lang=zh
+subsets=(train_covost)
+
+mkdir -p $data_dir
+splits=$(echo ${subsets[*]} | sed 's/ /_/g')
+cmd="python ${code_dir}/examples/speech_to_text/prep_audio_data.py
+    --data-root ${org_data_dir}
+    --output-root ${data_dir}
+    --task asr
+    --src-lang ${src_lang}
+    --splits ${splits}
+    --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}.* ${data_dir}
+    cmd="$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}

egs/iwslt2022/asr/conf/base.yaml

+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
+
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/asr/conf/basis.yaml

+train-subset: train
+valid-subset: dev
+
+max-epoch: 100
+max-update: 100000
+patience: 20
+best_checkpoint_metric: loss
+maximize_best_checkpoint_metric: False
+
+no-epoch-checkpoints: True
+#keep-last-epochs: 10
+keep-best-checkpoints: 10
+
+num-workers: 8
+no-progress-bar: True
+log-interval: 100
+seed: 1
+report-accuracy: True
+skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/iwslt2022/asr/conf/big.yaml

+arch: s2t_transformer_m
+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
+
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 2048
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:

egs/iwslt2022/asr/conf/conformer.yaml

+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 31
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish

egs/iwslt2022/asr/conf/ctc.yaml

+ctc-weight: 0.3
+post-process: sentencepiece

egs/iwslt2022/asr/conf/dlcl.yaml

+use-enc-dlcl: True
+use-dec-dlcl: True

egs/iwslt2022/asr/conf/inter.yaml

+ctc-weight: 0.2
+intermedia-ctc-layers: 6,9
+intermedia-adapter: league
+intermedia-ctc-weight: 0.1
+ctc-self-distill-weight: 0
+post-process: sentencepiece
\ No newline at end of file

egs/iwslt2022/asr/conf/local_attn.yaml

+encoder-attention-type: local
+hard-mask-window: 0
+gauss-mask-sigma: 3
+init-mask-weight: 0
\ No newline at end of file

egs/iwslt2022/asr/conf/pds_base.yaml

+arch: pdss2t_transformer_s_8
+pds-fusion: True
+ctc-layer: 12
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/asr/conf/pds_base_16.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 256
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 2_2_6_2
+pds-ratios: 2_2_2_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/asr/conf/pds_base_32.yaml

+arch: pdss2t_transformer_s_32
+
+encoder-embed-dim: 256
+pds-stages: 5
+ctc-layer: 12
+pds-layers: 2_2_3_3_2
+pds-ratios: 2_2_2_2_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1_1
+pds-kernel-sizes: 5_5_5_5_5
+pds-ffn-ratios: 8_8_8_8_8
+pds-attn-heads: 4_4_4_4_4
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/asr/conf/pds_base_8.yaml

+arch: pdss2t_transformer_s_8
+
+encoder-embed-dim: 256
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:

egs/iwslt2022/asr/conf/pds_big_8.yaml

+arch: pdss2t_transformer_m_8
+
+encoder-embed-dim: 512
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 512_512_512_512
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 4_4_4_4
+pds-attn-heads: 8_8_8_8
+
+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: 0.15
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:

egs/iwslt2022/asr/conf/purectc.yaml

+arch: s2t_ctc
+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: ctc
+zero_infinity: True
+post-process: sentencepiece
+
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.1
+attention-dropout: 0.1
+activation-dropout: 0.1
+
+activation-fn: relu
+encoder-embed-dim: 256
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+encoder-attention-heads: 4
+
+#load-pretrained-encoder-from:
\ No newline at end of file

egs/iwslt2022/asr/conf/rpr.yaml

+encoder-attention-type: rel_selfattn
+#encoder-attention-type: relative
+#max-encoder-relative-length: 100

egs/iwslt2022/asr/decode.sh

+#! /bin/bash
+
+gpu_num=1
+
+data_dir=
+test_subset=(dev tst-COMMON)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+cer=0
+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}
+    --cer ${cer}
+    --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/iwslt2022/asr/local/monitor.sh

+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/iwslt2022/asr/local/parse_options.sh

+#!/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/iwslt2022/asr/local/utils.sh

+
+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/iwslt2022/asr/run.sh

+#! /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=$(date "+%m%d_%H%M")
+
+stage=0
+stop_stage=0
+
+######## hardware ########
+# devices
+#device=()
+gpu_num=8
+update_freq=1
+
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+pwd_dir=$PWD
+
+# dataset
+src_lang=en
+tgt_lang=de
+lang=${src_lang}-${tgt_lang}
+
+dataset=mustc
+task=speech_to_text
+vocab_type=unigram
+vocab_size=5000
+speed_perturb=0
+lcrm=0
+tokenizer=0
+use_raw_audio=0
+
+use_specific_dict=1
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/asr
+train_split=train
+valid_split=dev
+test_split=tst-COMMON
+test_subset=tst-COMMON
+
+# 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
+cer=0
+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
+    feature_zip=fbank80.zip
+    if [[ ${speed_perturb} -eq 1 ]]; then
+        feature_zip=fbank80_sp.zip
+    fi
+    if [[ ! -f ${data_dir}/${feature_zip} && -f ${data_dir}/../feature_zip ]]; then
+        ln -s ${data_dir}/../feature_zip ${data_dir}
+    fi
+
+    cmd="python ${code_dir}/examples/speech_to_text/prep_audio_data.py
+        --data-root ${org_data_dir}
+        --output-root ${data_dir}
+        --task asr
+        --src-lang ${src_lang}
+        --splits ${valid_split},${test_split},${train_split}
+        --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}.* ${data_dir}
+        cmd="$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}
+
+    if [[ -f ${data_dir}/${feature_zip} && ! -f ${data_dir}/../${feature_zip} ]]; then
+        mv ${data_dir}/${feature_zip} ${data_dir}/..
+        ln -s ${data_dir}/../${feature_zip} ${data_dir}
+    fi
+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 ${model_dir}
+
+    extra_parameter="${extra_parameter}
+        --train-config ${pwd_dir}/conf/basis.yaml"
+    cp ${pwd_dir}/conf/basis.yaml ${model_dir}
+    config_list="${train_config//,/ }"
+    idx=1
+    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} ${model_dir}
+
+        extra_parameter="${extra_parameter}
+        --train-config${idx} ${config_path}"
+        idx=$((idx + 1))
+    done
+
+    cmd="python3 -u ${code_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
+
+        if [[ ! -f ${model_dir}/${dec_model} ]]; then
+            cmd="python ${code_dir}/scripts/average_checkpoints.py
+            --inputs ${model_dir}
+            --num-best-checkpoints ${n_average}
+            --output ${model_dir}/${dec_model}"
+            echo -e "\033[34mRun command: \n${cmd} \033[0m"
+            [[ $eval -eq 1 ]] && eval $cmd
+        fi
+	else
+		dec_model=${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}
+  		cmd="python ${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}
+        --max-tokens ${max_tokens}
+        --beam ${beam_size}
+        --lenpen ${len_penalty}
+        --scoring wer
+        --wer-tokenizer 13a
+        --wer-lowercase
+        --wer-remove-punct
+        "
+
+        if [[ ${cer} -eq 1 ]]; then
+            cmd="${cmd}
+        --wer-char-level"
+        fi
+
+    	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/iwslt2022/asr/train.sh

+#! /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 ctc)
+config_list=(purectc)
+#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/iwslt2022/mt/binary.sh

+set -e
+
+eval=1
+lcrm=0
+
+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/iwslt2022/mt/conf/base.yaml

+arch: transformer
+share-all-embeddings: 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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/mt/conf/basis.yaml

+train-subset: train
+valid-subset: valid
+
+max-epoch: 50
+max-update: 100000
+patience: 20
+best_checkpoint_metric: loss
+maximize_best_checkpoint_metric: False
+
+no-epoch-checkpoints: True
+#keep-last-epochs: 10
+keep-best-checkpoints: 10
+
+num-workers: 8
+no-progress-bar: True
+log-interval: 100
+seed: 1
+report-accuracy: True
+skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/iwslt2022/mt/conf/dlcl.yaml

+use-enc-dlcl: True
+use-dec-dlcl: True

egs/iwslt2022/mt/conf/rpr.yaml

+#encoder-attention-type: rel_selfattn
+encoder-attention-type: relative
+decoder-attention-type: relative
+max-encoder-relative-length: 20
+max-decoder-relative-length: 20

egs/iwslt2022/mt/conf/small.yaml

+arch: transformer
+share-all-embeddings: 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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/mt/decode.sh

+#! /bin/bash
+
+gpu_num=1
+
+data_dir=
+test_subset=(test)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+sacrebleu=1
+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}
+    --sacrebleu ${sacrebleu}
+    --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/iwslt2022/mt/local/lower_rm.py

+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/iwslt2022/mt/local/monitor.sh

+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/iwslt2022/mt/local/parse_options.sh

+#!/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/iwslt2022/mt/local/utils.sh

+
+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/iwslt2022/mt/train.sh

+#! /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/iwslt2022/st/binary.sh

+set -e
+
+eval=1
+
+lcrm=1
+tokenizer=0
+
+vocab_type=unigram
+vocab_size=5000
+use_raw_audio=0
+speed_perturb=0
+
+dataset=mustc
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/st
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
+asr_vocab_prefix=spm_unigram10000_st_share
+
+src_lang=en
+tgt_lang=zh
+subsets=(2019)
+
+splits=$(echo ${splits[*]} | sed 's/ /_/g')
+cmd="python ${code_dir}/examples/speech_to_text/prep_audio_data.py
+    --data-root ${org_data_dir}
+    --output-root ${data_dir}
+    --task asr
+    --src-lang ${src_lang}
+    --tgt-lang ${tgt_lang}
+    --splits ${splits}
+    --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}.* ${data_dir}
+    cmd="$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}

egs/iwslt2022/st/conf/base.yaml

+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
+
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/basis.yaml

+train-subset: train
+valid-subset: dev
+
+max-epoch: 100
+max-update: 100000
+patience: 20
+best_checkpoint_metric: loss
+maximize_best_checkpoint_metric: False
+
+no-epoch-checkpoints: True
+#keep-last-epochs: 10
+keep-best-checkpoints: 10
+
+num-workers: 8
+no-progress-bar: True
+log-interval: 100
+seed: 1
+report-accuracy: True
+skip-invalid-size-inputs-valid-test: True
\ No newline at end of file

egs/iwslt2022/st/conf/conformer.yaml

+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 31
+encoder-attention-type: rel_pos
+encoder-activation-fn: swish
\ No newline at end of file

egs/iwslt2022/st/conf/ctc.yaml

+ctc-weight: 0.3
+post-process: sentencepiece
\ No newline at end of file

egs/iwslt2022/st/conf/dlcl.yaml

+use-enc-dlcl: True
+use-dec-dlcl: True

egs/iwslt2022/st/conf/inter.yaml

+ctc-weight: 0.2
+intermedia-ctc-layers: 6,9
+intermedia-adapter: league
+intermedia-ctc-weight: 0.1
+ctc-self-distill-weight: 0
+post-process: sentencepiece
\ No newline at end of file

egs/iwslt2022/st/conf/local_attn.yaml

+encoder-attention-type: local
+hard-mask-window: 0
+gauss-mask-sigma: 3
+init-mask-weight: 0
\ No newline at end of file

egs/iwslt2022/st/conf/pds_base.yaml

+arch: pdss2t_transformer_s_8
+pds-fusion: True
+ctc-layer: 12
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/pds_base_16.yaml

+arch: pdss2t_transformer_s_16
+
+encoder-embed-dim: 256
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 2_2_6_2
+pds-ratios: 2_2_2_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/pds_base_32.yaml

+arch: pdss2t_transformer_s_32
+
+encoder-embed-dim: 256
+pds-stages: 5
+ctc-layer: 12
+pds-layers: 2_2_3_3_2
+pds-ratios: 2_2_2_2_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1_1
+pds-kernel-sizes: 5_5_5_5_5
+pds-ffn-ratios: 8_8_8_8_8
+pds-attn-heads: 4_4_4_4_4
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/pds_base_8.yaml

+arch: pdss2t_transformer_s_8
+
+pds-ctc: 1_1_1_1
+intermedia-adapter: league
+intermedia-ctc-weight: 0.15
+
+encoder-embed-dim: 256
+pds-stages: 4
+ctc-layer: 12
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+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
+
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/rpr.yaml

+encoder-attention-type: rel_selfattn
+#encoder-attention-type: relative
+#decoder-attention-type: relative
+#max-encoder-relative-length: 100
+#max-decoder-relative-length: 20

egs/iwslt2022/st/conf/sate.yaml

+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)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+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
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: transformer
+adapter: league
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/sate_big.yaml

+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: 1e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 2048
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.15
+activation-fn: relu
+encoder-embed-dim: 512
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: transformer
+adapter: league
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/sate_big_pds.yaml

+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: 1e-3
+adam_betas: (0.9,0.98)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 2048
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
+dropout: 0.15
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 8
+
+decoder-embed-dim: 512
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 8
+
+acoustic-encoder: pds
+adapter: league
+
+encoder-embed-dim: 512
+ctc-layer: 12
+pds-stages: 4
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 512_512_512_512
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 4_4_4_4
+pds-attn-heads: 8_8_8_8
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/conf/sate_pds.yaml

+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)
+
+criterion: label_smoothed_cross_entropy_with_ctc
+label_smoothing: 0.1
+
+encoder-normalize-before: True
+decoder-normalize-before: True
+dropout: 0.1
+activation-fn: relu
+encoder-ffn-embed-dim: 2048
+encoder-layers: 12
+text-encoder-layers: 6
+decoder-layers: 6
+encoder-attention-heads: 4
+
+decoder-embed-dim: 256
+decoder-ffn-embed-dim: 2048
+decoder-attention-heads: 4
+
+acoustic-encoder: pds
+adapter: league
+
+encoder-embed-dim: 256
+ctc-layer: 12
+pds-stages: 4
+pds-layers: 3_3_3_3
+pds-ratios: 2_2_1_2
+pds-fusion: True
+pds-fusion-method: all_conv
+pds-embed-dims: 256_256_256_256
+pds-ds-method: conv
+pds-embed-norm: True
+pds-position-embed: 1_1_1_1
+pds-kernel-sizes: 5_5_5_5
+pds-ffn-ratios: 8_8_8_8
+pds-attn-heads: 4_4_4_4
+
+#load-pretrained-encoder-from:
+#load-pretrained-acoustic-encoder-from:
+#load-pretrained-text-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/iwslt2022/st/decode.sh

+#! /bin/bash
+
+gpu_num=1
+
+data_dir=
+test_subset=(dev tst-COMMON)
+
+exp_name=
+if [ "$#" -eq 1 ]; then
+    exp_name=$1
+fi
+
+sacrebleu=1
+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}
+    --sacrebleu ${sacrebleu}
+    --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/iwslt2022/st/ensemble.sh

+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/iwslt2022/st/local/monitor.sh

+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/iwslt2022/st/local/parse_options.sh

+#!/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/iwslt2022/st/local/utils.sh

+
+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/iwslt2022/st/train.sh

+#! /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/libri_trans/asr/conf/debug.yaml

@@ -5,6 +5,7 @@ clip-norm: 10.0
 lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-updates: 10000
+weight-decay: 1e-6
 lr: 2e-3
 adam_betas: (0.9,0.98)
 
@@ -13,12 +14,21 @@ label_smoothing: 0.1
 
 subsampling-type: conv1d
 subsmapling-layers: 2
-subsampling-filter: 512
+subsampling-filter: 1024
 subsampling-kernel: 5
 subsampling-stride: 2
 subsampling-norm: none
 subsampling-activation: glu
 
+
+ctc-weight: 0.2
+intermedia-ctc-layers: 6,9
+intermedia-adapter: league
+intermedia-ctc-weight: 0.1
+intermedia-drop-prob: 0.5
+ctc-self-distill-weight: 0
+post-process: sentencepiece
+
 dropout: 0.1
 activation-fn: relu
 encoder-embed-dim: 256
@@ -33,8 +43,5 @@ decoder-attention-heads: 4
 attention-dropout: 0.1
 activation-dropout: 0.1
 
-macaron-style: True
-use-cnn-module: True
-cnn-module-kernel: 31
-encoder-activation-fn: swish
-encoder-attention-type: rel_pos_legacy
+#load-pretrained-encoder-from:
+#load-pretrained-decoder-from:
\ No newline at end of file

egs/mustc/asr/binary.sh

@@ -2,43 +2,61 @@ set -e
 
 eval=1
 
-lcrm=0
+lcrm=1
 tokenizer=0
 
-root_dir=~/st/Fairseq-S2T
-data_dir=~/st/data/test
-vocab_dir=~/st/data/mustc/st/en-de
+vocab_type=unigram
+vocab_size=5000
+use_raw_audio=0
+speed_perturb=0
+
+dataset=mustc
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/asr
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
 asr_vocab_prefix=spm_unigram10000_st_share
 
 src_lang=en
-tgt_lang=de
+tgt_lang=zh
 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 ${data_dir}
+splits=$(echo ${subsets[*]} | sed 's/ /_/g')
+cmd="python ${code_dir}/examples/speech_to_text/prep_audio_data.py
+    --data-root ${org_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"
+    --splits ${splits}
+    --vocab-type ${vocab_type}
+    --vocab-size ${vocab_size}"
 
-    if [[ ${lcrm} -eq 1 ]]; then
+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}.* ${data_dir}
+    cmd="$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
+fi
+if [[ ${tokenizer} -eq 1 ]]; then
     cmd="$cmd
     --tokenizer"
-    fi
+fi
 
 echo -e "\033[34mRun command: \n${cmd} \033[0m"
 [[ $eval -eq 1 ]] && eval ${cmd}

egs/mustc/st/binary.sh

@@ -5,44 +5,59 @@ 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
+vocab_type=unigram
+vocab_size=5000
+use_raw_audio=0
+speed_perturb=0
+
+dataset=mustc
+root_dir=~/st
+code_dir=${root_dir}/Fairseq-S2T
+org_data_dir=${root_dir}/data/${dataset}
+data_dir=${root_dir}/data/${dataset}/st
+
+use_specific_dict=0
+specific_prefix=st
+specific_dir=${root_dir}/data/mustc/st
 asr_vocab_prefix=spm_unigram10000_st_share
-st_vocab_prefix=spm_unigram10000_st_share
 
 src_lang=en
-tgt_lang=de
-splits=(2019)
+tgt_lang=zh
+subsets=(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 ${data_dir}
+splits=$(echo ${subsets[*]} | sed 's/ /_/g')
+cmd="python ${code_dir}/examples/speech_to_text/prep_audio_data.py
+    --data-root ${org_data_dir}
     --output-root ${data_dir}
-    --splits ${splits}
-    --task st
+    --task asr
     --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"
+    --splits ${splits}
+    --vocab-type ${vocab_type}
+    --vocab-size ${vocab_size}"
 
-    if [[ ${lcrm} -eq 1 ]]; then
+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}.* ${data_dir}
+    cmd="$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
+fi
+if [[ ${tokenizer} -eq 1 ]]; then
     cmd="$cmd
     --tokenizer"
-    fi
+fi
 
 echo -e "\033[34mRun command: \n${cmd} \033[0m"
 [[ $eval -eq 1 ]] && eval ${cmd}

egs/mustc/st/run.sh

@@ -129,11 +129,12 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     if [[ ! -e ${data_dir} ]]; then
         mkdir -p ${data_dir}
     fi
-    if [[ ! -f ${data_dir}/fbank80.zip && -f ${data_dir}/../fbank80.zip ]]; then
-        ln -s ${data_dir}/../fbank80.zip ${data_dir}
+    feature_zip=fbank80.zip
+    if [[ ${speed_perturb} -eq 1 ]]; then
+        feature_zip=fbank80_sp.zip
     fi
-    if [[ ! -f ${data_dir}/fbank80_sp.zip && -f ${data_dir}/../fbank80_sp.zip ]]; then
-        ln -s ${data_dir}/../fbank80_sp.zip ${data_dir}
+    if [[ ! -f ${data_dir}/${feature_zip} && -f ${data_dir}/../feature_zip ]]; then
+        ln -s ${data_dir}/../feature_zip ${data_dir}
     fi
 
     # create ASR vocabulary if necessary
@@ -204,13 +205,9 @@ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
     echo -e "\033[34mRun command: \n${cmd} \033[0m"
     [[ $eval -eq 1 ]] && eval ${cmd}
 
-    if [[ ! -f ${data_dir}/../fbank80.zip ]]; then
-        mv ${data_dir}/fbank80.zip ${data_dir}/..
-        ln -s ${data_dir}/../fbank80.zip ${data_dir}
-    fi
-    if [[ ! -f ${data_dir}/../fbank80_sp.zip ]]; then
-        mv ${data_dir}/fbank80_sp.zip ${data_dir}/..
-        ln -s ${data_dir}/../fbank80_sp.zip ${data_dir}
+    if [[ -f ${data_dir}/${feature_zip} && ! -f ${data_dir}/../${feature_zip} ]]; then
+        mv ${data_dir}/${feature_zip} ${data_dir}/..
+        ln -s ${data_dir}/../${feature_zip} ${data_dir}
     fi
 fi
 

examples/speech_to_text/data_utils.py

@@ -26,7 +26,8 @@ PAD_TOKEN, PAD_TOKEN_ID = "<pad>", 1
 
 def gen_vocab(
     input_path: Path, output_path_prefix: Path, model_type="bpe",
-    vocab_size=1000, special_symbols: Optional[List[str]] = None
+    vocab_size=1000, special_symbols: Optional[List[str]] = None,
+    normalization_rule_name=None
 ):
     # Train SentencePiece Model
     arguments = [
@@ -43,6 +44,8 @@ def gen_vocab(
         f"--eos_id={EOS_TOKEN_ID}",
         f"--pad_id={PAD_TOKEN_ID}",
     ]
+    if normalization_rule_name is not None:
+        arguments.append(f"--normalization_rule_name={normalization_rule_name}")
     if special_symbols is not None:
         _special_symbols = ",".join(special_symbols)
         arguments.append(f"--user_defined_symbols={_special_symbols}")

examples/speech_to_text/prep_audio_data.py

@@ -68,10 +68,6 @@ class AudioDataset(Dataset):
 
                 if 0 < self.size < total_length:
                     segments = segments[:self.size]
-                # for idx, seg in enumerate(content):
-                #     segments[idx] = seg
-                #     if 0 < self.size < idx:
-                #         break
         else:
             self.mode = "easy"
 
@@ -141,7 +137,7 @@ class AudioDataset(Dataset):
                 for i, segment in enumerate(seg_group):
                     offset = int(float(segment["offset"]) * sample_rate)
                     n_frames = int(float(segment["duration"]) * sample_rate)
-                    _id = f"{wav_path.stem}_{i}"
+                    _id = f"{split}_{wav_path.stem}_{i}"
 
                     item = dict()
                     item["audio"] = wav_path.as_posix()
@@ -240,7 +236,7 @@ def process(args):
         if not Path.exists(zip_path) or args.overwrite:
             gen_feature_flag = True
 
-    if gen_feature_flag:
+    if True and gen_feature_flag:
         if args.speed_perturb:
             feature_root = output_root / "fbank80_sp"
         else:
@@ -265,7 +261,7 @@ def process(args):
             for idx in tqdm(range(len(dataset))):
                 item = dataset[idx]
 
-                utt_id = item["id"]
+                utt_id = item['id']
                 features_path = (feature_root / f"{utt_id}.npy").as_posix()
 
                 if os.path.exists(features_path):
@@ -291,7 +287,7 @@ def process(args):
         create_zip(feature_root, zip_path)
 
         # Clean up
-        shutil.rmtree(feature_root)
+        # shutil.rmtree(feature_root)
 
     gen_manifest_flag = False
     for split in splits:

examples/speech_to_text/prep_mt_data.py

@@ -115,6 +115,7 @@ def process(args):
             output_root / tgt_spm_filename_prefix,
             args.vocab_type,
             args.vocab_size,
+            normalization_rule_name="identity" if tgt_lang == "zh" else None
         )
 
     if not args.share:
@@ -126,6 +127,7 @@ def process(args):
                 output_root / src_spm_filename_prefix,
                 args.vocab_type,
                 args.vocab_size,
+                normalization_rule_name="identity" if tgt_lang == "zh" else None
             )
 
     # Generate config YAML

fairseq/criterions/ctc.py

@@ -235,8 +235,8 @@ class CtcCriterion(FairseqCriterion):
 
                 ctc_self_distill_num += 1
                 loss = F.kl_div(
-                    F.log_softmax(inter_ctc_logit, dim=-1),
-                    F.softmax(ctc_logit, dim=-1),
+                    F.log_softmax(inter_ctc_logit, dim=-1, dtype=torch.float32),
+                    F.softmax(ctc_logit, dim=-1, dtype=torch.float32),
                     reduction="none",
                 )
                 loss = loss.sum(-1).transpose(0, 1).masked_fill_(~non_padding_mask, 0.0)

fairseq/data/audio/speech_to_text_dataset.py

@@ -360,6 +360,7 @@ class SpeechToTextDataset(FairseqDataset):
                 target = torch.cat((torch.LongTensor([lang_tag_idx]), target), 0)
 
         transcript = None
+
         if self.src_dict is not None and self.src_texts is not None and self.src_bpe_tokenizer is not None:
             tokenized = self.tokenize_text(self.src_texts[index], True)
             transcript = self.src_dict.encode_line(
@@ -443,6 +444,7 @@ class SpeechToTextDataset(FairseqDataset):
             "ntokens": ntokens,
             "nsentences": len(samples),
         }
+
         return out
 
     def num_tokens(self, index):

fairseq/models/speech_to_text/modules/adapter.py

@@ -49,14 +49,16 @@ class CTCCompressStrategy:
             for t_idx, same in enumerate(pred):
                 new_processed_inputs_cnt = processed_inputs_cnt + same[1]
                 # Get the probabilities of the prediction for the different time steps as weight
-                weights = F.softmax(prob_ctc[b_idx, processed_inputs_cnt:new_processed_inputs_cnt, same[0]])
+                weights = F.softmax(
+                    prob_ctc[b_idx, processed_inputs_cnt:new_processed_inputs_cnt, same[0]], dtype=torch.float32
+                )
                 weights_matrix[b_idx, processed_inputs_cnt:new_processed_inputs_cnt, t_idx] = \
                     weights / weights.sum()
                 processed_inputs_cnt = new_processed_inputs_cnt
         return weights_matrix
 
 
-class InterAdapter(nn.Module):
+class Adapter(nn.Module):
     def __init__(self, dim, adapter_type, dictionary, embed_tokens=None, strategy=None):
         super().__init__()
 
@@ -84,6 +86,7 @@ class InterAdapter(nn.Module):
             self.gate_linear2 = nn.Linear(dim, dim)
 
         if self.adapter_type == "shrink":
+            assert strategy is not None
             self.ctc_compress = getattr(CTCCompressStrategy, strategy)
             logger.info("CTC Compress Strategy: %s" % strategy)
         elif self.adapter_type == "league":
@@ -94,7 +97,7 @@ class InterAdapter(nn.Module):
     def forward(self, x, padding):
 
         representation, distribution = x
-        dim1, dim2, dim = representation.size()
+        seq_len, bsz, dim = representation.size()
         org_distribution = distribution
         lengths = (~padding).long().sum(-1)
 
@@ -103,7 +106,9 @@ class InterAdapter(nn.Module):
 
         elif self.adapter_type == "context":
             distribution = distribution.view(-1, distribution.size(-1))
-            out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
+            out = torch.mm(
+                distribution, self.embed_adapter.weight.float()
+            ).view(seq_len, bsz, -1).type_as(representation)
 
         elif self.adapter_type == "league":
             linear_out = self.linear_adapter(representation)
@@ -112,19 +117,25 @@ class InterAdapter(nn.Module):
                 threshold = distribution.sort(dim=-1, descending=True)[0][:, :, cutoff:cutoff+1]
                 distribution = torch.where(distribution > threshold, distribution, torch.zeros_like(distribution))
             distribution = distribution.view(-1, distribution.size(-1))
-            soft_out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
+            soft_out = torch.mm(
+                distribution, self.embed_adapter.weight.float()
+            ).view(seq_len, bsz, -1).type_as(representation)
             out = linear_out + soft_out
 
         elif self.adapter_type == "gated_league":
             linear_out = self.linear_adapter(representation)
             distribution = distribution.view(-1, distribution.size(-1))
-            soft_out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
+            soft_out = torch.mm(
+                distribution, self.embed_adapter.weight.float()
+            ).view(seq_len, bsz, -1).type_as(representation)
             coef = (self.gate_linear(torch.cat([linear_out, soft_out], dim=-1))).sigmoid()
             out = coef * linear_out + (1 - coef) * soft_out
 
         elif self.adapter_type == "inter_league":
             distribution = distribution.view(-1, distribution.size(-1))
-            soft_out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
+            soft_out = torch.mm(
+                distribution, self.embed_adapter.weight.float()
+            ).view(seq_len, bsz, -1).type_as(representation)
             out = representation + soft_out
 
         elif self.adapter_type == "none":
@@ -142,10 +153,11 @@ class InterAdapter(nn.Module):
 
                 new_lengths = [len(p) for p in batch_predicted]
                 weights_matrix = self.ctc_compress(prob_ctc, batch_predicted, new_lengths,
-                                                   representation.dtype, representation.device)
+                                                   prob_ctc.dtype, prob_ctc.device)
 
             # x is T x B x C -> B x C x T; weights_matrix is B x T x T'
-            compressed_output = representation.permute(1, 2, 0).bmm(weights_matrix)  # B x C x T'
+            representation = representation.permute(1, 2, 0)
+            compressed_output = representation.float().bmm(weights_matrix).type_as(representation)  # B x C x T'
             out = compressed_output.permute(2, 0, 1)
 
             out_lengths = lengths.new(new_lengths)

fairseq/models/speech_to_text/modules/ctc.py

@@ -38,10 +38,10 @@ class CTC(nn.Module):
         return x
 
     def softmax(self, x, temperature=1.0):
-        return F.softmax(self.ctc_projection(x) / temperature, dim=-1)
+        return F.softmax(self.ctc_projection(x) / temperature, dim=-1, dtype=torch.float32)
 
     def log_softmax(self, x, temperature=1.0):
-        return F.log_softmax(self.ctc_projection(x) / temperature, dim=-1)
+        return F.log_softmax(self.ctc_projection(x) / temperature, dim=-1, dtype=torch.float32)
 
     def argmax(self, x):
         return torch.argmax(self.ctc_projection(x), dim=-1)

fairseq/models/speech_to_text/pdss2t_transformer.py

@@ -13,7 +13,7 @@ from fairseq.models import (
     register_model_architecture,
 )
 from fairseq.models.speech_to_text import S2TTransformerModel
-from fairseq.models.speech_to_text.modules import CTC, InterAdapter
+from fairseq.models.speech_to_text.modules import CTC, Adapter
 
 from fairseq.modules import (
     FairseqDropout,
@@ -673,13 +673,13 @@ class PDSS2TTransformerEncoder(FairseqEncoder):
                         strategy = None
                         if args.intermedia_adapter == "shrink":
                             strategy = getattr(args, "ctc_compress_strategy", "avg")
-                        adapter = InterAdapter(embed_dim, args.intermedia_adapter,
+                        adapter = Adapter(embed_dim, args.intermedia_adapter,
                                           task.source_dictionary, strategy=strategy)
                         inter_adapter = adapter
                     else:
                         adapter = inter_adapter
                 else:
-                    adapter = InterAdapter(embed_dim, "none",
+                    adapter = Adapter(embed_dim, "none",
                                       task.source_dictionary)
             else:
                 ctc = None
@@ -830,7 +830,7 @@ class PDSS2TTransformerEncoder(FairseqEncoder):
                 logit = ctc(x.clone())
                 intermedia_ctc_logits.append([logit, encoder_padding_mask])
 
-                prob = F.softmax(logit, dim=-1)
+                prob = utils.softmax(logit, dim=-1)
                 x, encoder_padding_mask = adapter([x, prob], encoder_padding_mask)
 
         if self.fusion_stages_num != 0:

fairseq/models/speech_to_text/s2t_ctc.py

@@ -14,7 +14,7 @@ from fairseq.models import (
     register_model,
     register_model_architecture,
 )
-from fairseq.models.speech_to_text.modules import InterAdapter, CTC
+from fairseq.models.speech_to_text.modules import Adapter, CTC
 from fairseq.modules import (
     FairseqDropout,
     LayerNorm,
@@ -382,6 +382,12 @@ class S2TCTCModel(FairseqEncoderModel):
             type=int,
             help="cutoff of the distribution",
         )
+        parser.add_argument(
+            "--intermedia-drop-prob",
+            default=0,
+            type=float,
+            help="probability of dropping the followed layers",
+        )
         pass
 
     @classmethod
@@ -424,9 +430,9 @@ class S2TCTCModel(FairseqEncoderModel):
         else:
             logits = net_output["ctc_logit"][0]
         if log_probs:
-            return utils.log_softmax(logits.float(), dim=-1)
+            return utils.log_softmax(logits, dim=-1)
         else:
-            return utils.softmax(logits.float(), dim=-1)
+            return utils.softmax(logits, dim=-1)
 
     def forward(self, src_tokens, src_lengths, prev_output_tokens=None):
         """
@@ -513,11 +519,12 @@ class S2TCTCEncoder(FairseqEncoder):
 
             strategy = None
             if args.intermedia_adapter == "shrink":
-                strategy = getattr(args, "ctc_compress_strategy", None)
+                strategy = getattr(args, "ctc_compress_strategy", "avg")
             elif args.intermedia_adapter == "league":
                 strategy = getattr(args, "intermedia_distribution_cutoff", -1)
-            self.adapter = InterAdapter(dim, args.intermedia_adapter,
+            self.adapter = Adapter(dim, args.intermedia_adapter,
                                    task.source_dictionary, strategy=strategy)
+            self.intermedia_drop_prob = getattr(args, "intermedia_drop_prob", 0)
 
     def add_to_dict(self, x, dis, idx):
         sim = 0
@@ -595,11 +602,16 @@ class S2TCTCEncoder(FairseqEncoder):
 
             # interleave CTC
             if layer_idx in self.intermedia_ctc_layers:
+                if self.intermedia_drop_prob > 0:
+                    p = torch.rand(1).uniform_()
+                    if p < self.intermedia_drop_prob:
+                        break
+
                 norm_x = self.layer_norm(x)
                 logit = self.ctc(norm_x)
                 intermedia_ctc_logits.append(logit)
 
-                prob = F.softmax(logit, dim=-1)
+                prob = F.softmax(logit, dim=-1, dtype=torch.float32)
                 x, encoder_padding_mask = self.adapter([x, prob], encoder_padding_mask)
 
             # gather cosine similarity
@@ -702,7 +714,7 @@ class CTCDecoder(object):
             model_path=self.lm_model,
             alpha=self.lm_weight,
             beta=0,
-            cutoff_top_n=self.vocab_size,
+            cutoff_top_n=40,
             cutoff_prob=1.0,
             beam_width=self.beam_size,
             num_processes=20,
@@ -725,7 +737,9 @@ class CTCDecoder(object):
                                   src_lengths=src_lengths)
 
         ctc_logit = encoder_outs["ctc_logit"][0].transpose(0, 1)
-        beam_results, beam_scores, timesteps, out_lens = self.ctc_decoder.decode(F.softmax(ctc_logit, -1), src_lengths)
+        beam_results, beam_scores, time_steps, out_lens = self.ctc_decoder.decode(
+            utils.softmax(ctc_logit, -1), src_lengths
+        )
 
         finalized = []
         for idx in range(bsz):

fairseq/models/speech_to_text/s2t_transformer.py

@@ -2,6 +2,7 @@ import logging
 import math
 from typing import Dict, List, Optional, Tuple
 
+import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
@@ -13,7 +14,7 @@ from fairseq.models import (
     register_model,
     register_model_architecture,
 )
-from fairseq.models.speech_to_text.modules import InterAdapter, CTC
+from fairseq.models.speech_to_text.modules import Adapter, CTC
 from fairseq.models.transformer import Embedding, TransformerDecoder
 from fairseq.modules import (
     FairseqDropout,
@@ -388,6 +389,12 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
             type=int,
             help="cutoff of the distribution",
         )
+        parser.add_argument(
+            "--intermedia-drop-prob",
+            default=0,
+            type=float,
+            help="probability of dropping the followed layers",
+        )
         pass
 
     @classmethod
@@ -494,6 +501,7 @@ class S2TTransformerEncoder(FairseqEncoder):
         self.padding_idx = 1
 
         self.subsample = subsampling(args)
+        self.linear = nn.Linear(dim, dim)
 
         self.attn_type = getattr(args, "encoder_attention_type", "selfattn")
 
@@ -573,8 +581,9 @@ class S2TTransformerEncoder(FairseqEncoder):
                 strategy = getattr(args, "ctc_compress_strategy", None)
             elif args.intermedia_adapter == "league":
                 strategy = getattr(args, "intermedia_distribution_cutoff", -1)
-            self.adapter = InterAdapter(dim, args.intermedia_adapter,
+            self.adapter = Adapter(dim, args.intermedia_adapter,
                                    task.source_dictionary, strategy=strategy)
+            self.intermedia_drop_prob = getattr(args, "intermedia_drop_prob", 0)
 
     @staticmethod
     def pooling_ratio():
@@ -631,6 +640,7 @@ class S2TTransformerEncoder(FairseqEncoder):
             x += positions
             positions = None
 
+        x = self.linear(x)
         x = self.dropout_module(x)
 
         # add emb into history
@@ -662,12 +672,17 @@ class S2TTransformerEncoder(FairseqEncoder):
 
             # interleave CTC
             if layer_idx in self.intermedia_ctc_layers:
+                if self.intermedia_drop_prob > 0:
+                    p = torch.rand(1).uniform_()
+                    if p < self.intermedia_drop_prob:
+                        break
+
                 norm_x = self.layer_norm(x)
                 logit = self.ctc(norm_x)
-                intermedia_ctc_logits.append(logit)
 
+                intermedia_ctc_logits.append(logit)
                 # prob = self.ctc.softmax(norm_x)
-                prob = F.softmax(logit, dim=-1)
+                prob = utils.softmax(logit, dim=-1)
                 x, encoder_padding_mask = self.adapter([x, prob], encoder_padding_mask)
 
             # gather cosine similarity
@@ -872,6 +887,7 @@ def base_architecture(args):
     # intermedia CTC
     args.intermedia_ctc_layers = getattr(args, "intermedia_ctc_layers", None)
     args.intermedia_adapter = getattr(args, "intermedia_adapter", None)
+    args.intermedia_drop_prob = getattr(args, "intermedia_drop_prob", 0)
 
 
 @register_model_architecture("s2t_transformer", "s2t_transformer_s")

fairseq/models/transformer.py

@@ -478,6 +478,7 @@ class TransformerEncoder(FairseqEncoder):
 
         self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(embed_dim)
 
+        self.attn_type = getattr(args, "encoder_attention_type", "selfattn")
         if self.attn_type == "rel_pos":
             self.embed_positions = RelPositionalEncoding(
                 args.max_source_positions, args.encoder_embed_dim

fairseq/modules/dlcl.py

@@ -121,7 +121,7 @@ class DynamicLinearCombination(nn.Module):
                 self.weight_mask = self.weight_mask.cuda()
             if self.normalize_learned_weight:
                 weight = self.weight.masked_fill((self.weight_mask == 0).unsqueeze(2), float('-inf'))
-                self.normalized_weight = F.softmax(weight, dim=1)
+                self.normalized_weight = F.softmax(weight, dim=1, dtype=torch.float32)
             return
 
         # following layer

fairseq/modules/espnet_multihead_attention.py

@@ -9,6 +9,7 @@
 import math
 import torch
 from torch import nn
+import torch.nn.functional as F
 from fairseq.modules.rotary_positional_embedding import (
     RotaryPositionalEmbedding,
     apply_rotary_pos_emb,
@@ -73,10 +74,8 @@ class ESPNETMultiHeadedAttention(nn.Module):
                 mask.unsqueeze(1).unsqueeze(2).to(bool),
                 float("-inf"),  # (batch, head, time1, time2)
             )
-            self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
-
-        else:
-            self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+        self.attn = F.softmax(scores, dim=-1, dtype=torch.float32).type_as(scores)  # (batch, head, time1, time2)
+        # self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
         p_attn = self.dropout(self.attn)
         x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
         x = (
@@ -91,7 +90,7 @@ class ESPNETMultiHeadedAttention(nn.Module):
             query (torch.Tensor): Query tensor T X B X C
             key (torch.Tensor): Key tensor T X B X C
             value (torch.Tensor): Value tensor T X B X C
-            mask (torch.Tensor): Mask tensor T X B
+            key_padding_mask (torch.Tensor): Mask tensor T X B
         Returns:
             torch.Tensor: Output tensor T X B X D.
         """

fairseq/modules/multihead_attention.py

@@ -151,6 +151,7 @@ class MultiheadAttention(nn.Module):
         assert list(query.size()) == [tgt_len, bsz, embed_dim]
 
         if (
+            False and
             not self.onnx_trace
             and not is_tpu  # don't use PyTorch version on TPUs
             and incremental_state is None
@@ -349,9 +350,7 @@ class MultiheadAttention(nn.Module):
         if before_softmax:
             return attn_weights, v
 
-        attn_weights_float = utils.softmax(
-            attn_weights, dim=-1, onnx_trace=self.onnx_trace
-        )
+        attn_weights_float = F.softmax(attn_weights, dim=-1, dtype=torch.float32)
         attn_weights = attn_weights_float.type_as(attn_weights)
         attn_probs = self.dropout_module(attn_weights)
 

fairseq/modules/relative_multihead_attention.py

@@ -205,6 +205,7 @@ class RelativeMultiheadAttention(MultiheadAttention):
             # In this branch incremental_state is never None
             assert incremental_state is not None
             incremental_state = self._set_input_buffer(incremental_state, saved_state)
+
         assert k is not None
         src_len = k.size(1)
 
@@ -271,7 +272,7 @@ class RelativeMultiheadAttention(MultiheadAttention):
                 )
             else:
                 attn_weights = attn_weights.transpose(0, 2)
-                attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf"))
+                attn_weights = attn_weights.float().masked_fill(key_padding_mask, float("-inf")).type_as(attn_weights)
                 attn_weights = attn_weights.transpose(0, 2)
             attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
 
@@ -292,7 +293,6 @@ class RelativeMultiheadAttention(MultiheadAttention):
         else:
             attn = self._relative_attention_inner(attn_probs, v, relation_values, transpose=False)
 
-        # attn = torch.bmm(attn_probs, v)
         assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
         if self.onnx_trace and attn.size(1) == 1:
             # when ONNX tracing a single decoder step (sequence length == 1)
@@ -316,11 +316,11 @@ class RelativeMultiheadAttention(MultiheadAttention):
     def _generate_relative_positions_matrix(length, max_relative_length, device, incremental_state):
         if not incremental_state:
             # training process
-            range_vec = torch.arange(length).to(device)
+            range_vec = torch.arange(length, device=device)
             range_mat = range_vec.repeat(length, 1)
             distance_mat = range_mat - range_mat.transpose(0, 1)
         else:
-            distance_mat = torch.arange(-length + 1, 1).view(1, -1).to(device)
+            distance_mat = torch.arange(-length + 1, 0, device=device).view(1, -1)
 
         distance_mat_clipped = torch.clamp(distance_mat, -max_relative_length, max_relative_length)
 
@@ -337,7 +337,7 @@ class RelativeMultiheadAttention(MultiheadAttention):
 
         Args:
           x: Tensor with shape [batch_size*heads, length, length or depth].
-          y: Tensor with shap e [batch_size*heads, length, depth].
+          y: Tensor with shape [batch_size*heads, length, depth].
           z: Tensor with shape [length, length, depth].
           transpose: Whether to transpose inner matrices of y and z. Should be true if
               last dimension of x is depth, not length.

fairseq/modules/s2t_transformer_layer.py

@@ -161,7 +161,13 @@ class S2TTransformerEncoderLayer(nn.Module):
             else:
                 print("The maximum encoder relative length %d can not be -1!" % max_relative_length)
                 exit(1)
-        elif self.attn_type in ["rel_pos", "rel_pos_legacy"]:
+        elif self.attn_type == "rel_pos":
+            return RelPositionMultiHeadedAttention(
+                embed_dim,
+                attention_heads,
+                dropout=dropout,
+            )
+        elif self.attn_type == "rel_pos_legacy":
             return LegacyRelPositionMultiHeadedAttention(
                 embed_dim,
                 attention_heads,
@@ -236,8 +242,8 @@ class S2TTransformerEncoderLayer(nn.Module):
         # Note that we cannot use -inf here, because at some edge cases,
         # the attention weight (before softmax) for some padded element in query
         # will become -inf, which results in NaN in model parameters
-        # if attn_mask is not None:
-        #     attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8)
+        if attn_mask is not None:
+            attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8)
 
         # whether to use macaron style
         if self.macaron_norm is not None:

fairseq/modules/speech_to_text/subsampling.py

@@ -142,9 +142,6 @@ class Conv1dSubsampling(nn.Module):
             get_activation_class(act, dim=1)
         ) for layer_id in range(num_layers)])
 
-        out_dim = filters[-1]
-        self.linear = nn.Linear(out_dim, out_dim)
-
     def forward(self, x, x_len):
 
         # (B, T, D) -> (B, D, T)
@@ -157,7 +154,6 @@ class Conv1dSubsampling(nn.Module):
             if x_len is not None:
                 x_len = torch.div(x_len - 1, 2, rounding_mode='floor') + 1
         x = x.transpose(1, 2)
-        x = self.linear(x)
         return x, x_len
 
 

fairseq/modules/transformer_layer.py

@@ -169,7 +169,7 @@ class TransformerEncoderLayer(nn.Module):
                 `attn_mask[tgt_i, src_j] = 1` means that when calculating the
                 embedding for `tgt_i`, we exclude (mask out) `src_j`. This is
                 useful for strided self-attention.
-            positions (Tensor): the position embedding for relative position encoding
+            pos_emb (Tensor): the position embedding for relative position encoding
 
         Returns:
             encoded output of shape `(seq_len, batch, embed_dim)`
@@ -180,7 +180,9 @@ class TransformerEncoderLayer(nn.Module):
         # the attention weight (before softmax) for some padded element in query
         # will become -inf, which results in NaN in model parameters
         if attn_mask is not None:
-            attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8)
+            attn_mask = attn_mask.masked_fill(
+                attn_mask.to(torch.bool), -1e8 if x.dtype == torch.float32 else -1e4
+            )
 
         residual = x
         if self.normalize_before: