This repository was archived by the owner on Nov 21, 2023. It is now read-only.
PS and KM basic support #17
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,156 @@ | ||
| # Copyright (c) Facebook, Inc. and its affiliates. | ||
| # All rights reserved. | ||
| # | ||
| # This source code is licensed under the license found in the | ||
| # LICENSE file in the root directory of this source tree. | ||
| # | ||
| #!/bin/bash | ||
|
|
||
| SRC="ps" | ||
| SRCS="ps km hi fa" | ||
| TGT=en | ||
|
|
||
| BPESIZE=15000 | ||
| TRAIN_MINLEN=1 # remove sentences with <1 BPE token | ||
| TRAIN_MAXLEN=250 # remove sentences with >250 BPE tokens | ||
|
|
||
| ROOT=$(dirname "$0") | ||
| SCRIPTS=$ROOT/scripts | ||
| DATA=$ROOT/data | ||
| mkdir -p $DATA | ||
|
|
||
| SPM_TRAIN=$SCRIPTS/spm_train.py | ||
| SPM_ENCODE=$SCRIPTS/spm_encode.py | ||
|
|
||
| TRAIN_SETS_PS=( | ||
| "all-clean-ps/GNOME.en-ps" | ||
| "all-clean-ps/KDE4.en-ps" | ||
| "all-clean-ps/Tatoeba.en-ps" | ||
| "all-clean-ps/Ubuntu.en-ps" | ||
| "all-clean-ps/wikimedia.en-ps" | ||
| ) | ||
| TRAIN_SETS_KM=( | ||
| "all-clean-km/GlobalVoices.en-km" | ||
| "all-clean-km/GNOME.en-km" | ||
| "all-clean-km/KDE4.en-km" | ||
| "all-clean-km/Tatoeba.en-km" | ||
| "all-clean-km/Ubuntu.en-km" | ||
| ) | ||
| TRAIN_SETS_HI=("all-clean-hi/IITB.en-hi") | ||
| DATA_SETS_FA=( | ||
| "all-clean-fa/GlobalVoices.en-fa" | ||
| "all-clean-fa/GNOME.en-fa" | ||
| "all-clean-fa/infopankki.en-fa" | ||
| "all-clean-fa/KDE4.en-fa" | ||
| "all-clean-fa/OpenSubtitles.en-fa" | ||
| "all-clean-fa/QED.en-fa" | ||
| "all-clean-fa/Tanzil.en-fa" | ||
| "all-clean-fa/TED2013.en-fa" | ||
| "all-clean-fa/TEP.en-fa" | ||
| "all-clean-fa/Ubuntu.en-fa" | ||
| "all-clean-fa/Wikipedia.en-fa" | ||
| ) | ||
|
|
||
| echo "Aggregating splits" | ||
| for FILE in "${TRAIN_SETS_PS[@]}" ; do cat "${DATA}/$FILE.ps"; done > $DATA/undeduped-train.ps-en.ps | ||
| for FILE in "${TRAIN_SETS_PS[@]}" ; do cat "${DATA}/$FILE.en"; done > $DATA/undeduped-train.ps-en.en | ||
| for FILE in "${TRAIN_SETS_KM[@]}" ; do cat "${DATA}/$FILE.km"; done > $DATA/undeduped-train.km-en.km | ||
| for FILE in "${TRAIN_SETS_KM[@]}" ; do cat "${DATA}/$FILE.en"; done > $DATA/undeduped-train.km-en.en | ||
| for FILE in "${TRAIN_SETS_HI[@]}" ; do cat "${DATA}/$FILE.hi"; done > $DATA/undeduped-train.hi-en.hi | ||
| for FILE in "${TRAIN_SETS_HI[@]}" ; do cat "${DATA}/$FILE.en"; done > $DATA/undeduped-train.hi-en.en | ||
| for FILE in "${DATA_SETS_FA[@]}" ; do cat "${DATA}/$FILE.fa"; done | awk 'NR % 1500 > 1' > $DATA/undeduped-train.fa-en.fa | ||
| for FILE in "${DATA_SETS_FA[@]}" ; do cat "${DATA}/$FILE.en"; done | awk 'NR % 1500 > 1' > $DATA/undeduped-train.fa-en.en | ||
| for FILE in "${DATA_SETS_FA[@]}" ; do cat "${DATA}/$FILE.fa"; done | awk 'NR % 1500 == 0' > $DATA/undeduped-valid.fa-en.fa | ||
| for FILE in "${DATA_SETS_FA[@]}" ; do cat "${DATA}/$FILE.en"; done | awk 'NR % 1500 == 0' > $DATA/undeduped-valid.fa-en.en | ||
| for FILE in "${DATA_SETS_FA[@]}" ; do cat "${DATA}/$FILE.fa"; done | awk 'NR % 1500 == 1' > $DATA/undeduped-test.fa-en.fa | ||
| for FILE in "${DATA_SETS_FA[@]}" ; do cat "${DATA}/$FILE.en"; done | awk 'NR % 1500 == 1' > $DATA/undeduped-test.fa-en.en | ||
| echo "Deduplicating splits" | ||
| dedupe () { | ||
| dedupe_set=$1 | ||
| dedupe_lang=$2 | ||
| python3 scripts/deduplicate.py \ | ||
| --input-src $DATA/undeduped-$dedupe_set.$dedupe_lang-en.$dedupe_lang \ | ||
| --input-tgt $DATA/undeduped-$dedupe_set.$dedupe_lang-en.en \ | ||
| --output-src $DATA/$dedupe_set.$dedupe_lang-en.$dedupe_lang \ | ||
| --output-tgt $DATA/$dedupe_set.$dedupe_lang-en.en | ||
| } | ||
| dedupe "train" "ps" | ||
| dedupe "train" "km" | ||
| dedupe "train" "hi" | ||
| dedupe "train" "fa" | ||
| dedupe "valid" "fa" | ||
| dedupe "test" "fa" | ||
| cat "${DEVTEST_PSKM}/ps-en.dev.ps" > $DATA/valid.ps-en.ps | ||
| cat "${DEVTEST_PSKM}/ps-en.dev.en" > $DATA/valid.ps-en.en | ||
| cat "${DEVTEST_PSKM}/ps-en.devtest.ps" > $DATA/test.ps-en.ps | ||
| cat "${DEVTEST_PSKM}/ps-en.devtest.en" > $DATA/test.ps-en.en | ||
| cat "${DEVTEST_PSKM}/km-en.dev.km" > $DATA/valid.km-en.km | ||
| cat "${DEVTEST_PSKM}/km-en.dev.en" > $DATA/valid.km-en.en | ||
| cat "${DEVTEST_PSKM}/km-en.devtest.km" > $DATA/test.km-en.km | ||
| cat "${DEVTEST_PSKM}/km-en.devtest.en" > $DATA/test.km-en.en | ||
| cat "${DEVTEST_HI}/dev.hi" > $DATA/valid.hi-en.hi | ||
| cat "${DEVTEST_HI}/dev.en" > $DATA/valid.hi-en.en | ||
| cat "${DEVTEST_HI}/test.hi" > $DATA/test.hi-en.hi | ||
| cat "${DEVTEST_HI}/test.en" > $DATA/test.hi-en.en | ||
|
There was a problem hiding this comment.
|
||
| echo "Creating abridged train sets for BPE, so all languages are similarly well-represented" | ||
| cat $DATA/train.ps-en.ps | head -100000 > $DATA/abridged.ps | ||
| cat $DATA/train.km-en.km | head -100000 > $DATA/abridged.km | ||
| cat $DATA/train.hi-en.hi | head -100000 > $DATA/abridged.hi | ||
| cat $DATA/train.fa-en.fa | head -100000 > $DATA/abridged.fa | ||
| cat $DATA/train.ps-en.en | head -100000 > $DATA/abridged.en | ||
| cat $DATA/train.km-en.en | head -100000 >> $DATA/abridged.en | ||
| cat $DATA/train.hi-en.en | head -100000 >> $DATA/abridged.en | ||
| cat $DATA/train.fa-en.en | head -100000 >> $DATA/abridged.en | ||
| echo "Concatenating target data" | ||
| cat $DATA/train.ps-en.en > $DATA/train-concatenated.en | ||
| cat $DATA/train.km-en.en >> $DATA/train-concatenated.en | ||
| cat $DATA/train.hi-en.en >> $DATA/train-concatenated.en | ||
| cat $DATA/train.fa-en.en >> $DATA/train-concatenated.en | ||
| cat $DATA/valid.ps-en.en > $DATA/valid-concatenated.en | ||
| cat $DATA/valid.km-en.en >> $DATA/valid-concatenated.en | ||
| cat $DATA/valid.hi-en.en >> $DATA/valid-concatenated.en | ||
| cat $DATA/valid.fa-en.en >> $DATA/valid-concatenated.en | ||
| cat $DATA/test.ps-en.en > $DATA/test-concatenated.en | ||
| cat $DATA/test.km-en.en >> $DATA/test-concatenated.en | ||
| cat $DATA/test.hi-en.en >> $DATA/test-concatenated.en | ||
| cat $DATA/test.fa-en.en >> $DATA/test-concatenated.en | ||
| echo "Learning BPE with sentencepiece" | ||
| python $SPM_TRAIN \ | ||
| --input=$DATA/abridged.ps,$DATA/abridged.km,$DATA/abridged.hi,$DATA/abridged.fa,$DATA/abridged.en \ | ||
| --model_prefix=$DATA/sentencepiece.bpe \ | ||
| --vocab_size=$BPESIZE \ | ||
| --character_coverage=1.0 \ | ||
| --model_type=bpe | ||
| echo "Encoding" | ||
| for SOMESRC in $SRCS; do | ||
| python $SPM_ENCODE \ | ||
| --model $DATA/sentencepiece.bpe.model \ | ||
| --output_format=piece \ | ||
| --inputs $DATA/train.$SOMESRC-en.$SOMESRC $DATA/train.$SOMESRC-en.en \ | ||
| --outputs $DATA/train.bpe.$SOMESRC-en.$SOMESRC $DATA/train.bpe.$SOMESRC-en.en \ | ||
| --min-len $TRAIN_MINLEN --max-len $TRAIN_MAXLEN | ||
| for SPLIT in "valid" "test"; do | ||
| python $SPM_ENCODE \ | ||
| --model $DATA/sentencepiece.bpe.model \ | ||
| --output_format=piece \ | ||
| --inputs $DATA/$SPLIT.$SOMESRC-en.$SOMESRC $DATA/$SPLIT.$SOMESRC-en.en \ | ||
| --outputs $DATA/$SPLIT.bpe.$SOMESRC-en.$SOMESRC $DATA/$SPLIT.bpe.$SOMESRC-en.en | ||
| done | ||
| done | ||
| echo "Converting sentencepiece vocabulary into fairseq dictionary" | ||
| tail -n +4 $DATA/sentencepiece.bpe.vocab | awk -F'\t' 'BEGIN{OFS=" "} {print $1, 100}' > $DATA/vocab | ||
| echo "Preprocessing" | ||
| for SOMESRC in $SRCS; do | ||
| echo "Binarizing ${SOMESRC}" | ||
| fairseq-preprocess \ | ||
| --source-lang $SOMESRC --target-lang en \ | ||
| --destdir $DATA \ | ||
| --joined-dictionary \ | ||
| --workers 4 \ | ||
| --trainpref $DATA/train.bpe.$SOMESRC-en \ | ||
| --validpref $DATA/valid.bpe.$SOMESRC-en \ | ||
| --testpref $DATA/test.bpe.$SOMESRC-en \ | ||
| --srcdict $DATA/vocab | ||
| mv "${DATA}/dict.en.txt" "${DATA}/dict.en.txt-moved" | ||
| done | ||
| mv "${DATA}/dict.en.txt-moved" "${DATA}/dict.en.txt" | ||
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We don't have Farsi in the original flores. Is there a reason why are you including it in the target?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I believe the intent was to improve the Pashto performance of a multilingual model by training for Farsi as well, being a related language. The Pashto-English parallel corpus was very limited so being able to make use of the Farsi-English corpus would hopefully be a boon. I don't recall whether we tested this or if we did whether it had an impact.