add

toolkit/build_fairseq_sharded_dataset.sh

@@ -0,0 +1,155 @@
+#!/bin/bash
+
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=40
+#SBATCH --job-name=process
+#SBATCH --nodelist=ilps-cn002
+#SBATCH --time=2-00:00:00
+#SBATCH --mem=256G
+
+pwd
+conda info --envs
+source /home/stan1/anaconda3/bin/activate fairseq
+cd /ivi/ilps/projects/ltl-mt/EC40-dataset
+
+mkdir spm_sharded
+
+
+######################## ------------ IMPORTRANT ------------ ########################
+
+######## This is an example of how to build a sharded dataset (5 shards)
+######## Before run the following code, you should have trained your sentencepiece/subword-mt tokenizer already
+######## Then you should encode the dataset using spm, and then use following code to split them to 5 shards
+
+#### For eval set, the most easiest way is to add the whole eval-set to all 5 shard fairseq data folder
+### note: ha and kab is two exceptions (because of their data-size): you will find them in *SPECIAL* 
+
+######################## ------------ IMPORTRANT ------------ ########################
+
+SHARD_SUB_DIR=('0' '1' '2' '3' '4')
+for i in "${!SHARD_SUB_DIR[@]}"; do
+    SUB_NUMBER=${SHARD_SUB_DIR[i]}
+    mkdir dataset/spm_sharded/shard${SUB_NUMBER}
+done
+
+HIGH=('de' 'nl' 'fr' 'es' 'ru' 'cs' 'hi' 'bn' 'ar' 'he')
+MED=('sv' 'da' 'it' 'pt' 'pl' 'bg' 'kn' 'mr' 'mt') #ha
+LOW=('af' 'lb' 'ro' 'oc' 'uk' 'sr' 'sd' 'gu' 'ti' 'am')
+ELOW=('no' 'is' 'ast' 'ca' 'be' 'bs' 'ne' 'ur' 'so') #kab
+
+SPM_DIR=dataset/spm
+SPM_SHARD_DIR=dataset/spm_sharded
+
+##
+
+## HIGH 5m each file -> split to 1m for one shard
+for i in "${!HIGH[@]}"; do
+    LANG=${HIGH[i]}
+    split -l 1000000 $SPM_DIR/train.en-$LANG.en -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.en.shard
+    split -l 1000000 $SPM_DIR/train.en-$LANG.$LANG -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard
+
+    for j in "${!SHARD_SUB_DIR[@]}"; do
+        SUB_NUMBER=${SHARD_SUB_DIR[j]}
+        mv $SPM_SHARD_DIR/train.en-$LANG.en.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.en
+        mv $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.$LANG
+    done
+done
+
+# MED 1m each file -> split to 200K for one shard
+for i in "${!MED[@]}"; do
+    LANG=${MED[i]}
+    split -l 200000 $SPM_DIR/train.en-$LANG.en -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.en.shard
+    split -l 200000 $SPM_DIR/train.en-$LANG.$LANG -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard
+
+    for j in "${!SHARD_SUB_DIR[@]}"; do
+        SUB_NUMBER=${SHARD_SUB_DIR[j]}
+        mv $SPM_SHARD_DIR/train.en-$LANG.en.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.en
+        mv $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.$LANG
+    done
+done
+
+# LOW 100k each file -> split to 20k for one shard
+for i in "${!LOW[@]}"; do
+    LANG=${LOW[i]}
+    split -l 20000 $SPM_DIR/train.en-$LANG.en -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.en.shard
+    split -l 20000 $SPM_DIR/train.en-$LANG.$LANG -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard
+
+    for j in "${!SHARD_SUB_DIR[@]}"; do
+        SUB_NUMBER=${SHARD_SUB_DIR[j]}
+        mv $SPM_SHARD_DIR/train.en-$LANG.en.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.en
+        mv $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.$LANG
+    done
+done
+
+## ELOW 50k each file -> split to 10k for one shard
+for i in "${!ELOW[@]}"; do
+    LANG=${ELOW[i]}
+    split -l 10000 $SPM_DIR/train.en-$LANG.en -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.en.shard
+    split -l 10000 $SPM_DIR/train.en-$LANG.$LANG -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard
+
+    for j in "${!SHARD_SUB_DIR[@]}"; do
+        SUB_NUMBER=${SHARD_SUB_DIR[j]}
+        mv $SPM_SHARD_DIR/train.en-$LANG.en.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.en
+        mv $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.$LANG
+    done
+done
+
+# SPECIAL HA 344000 -> split to 68800 for one shard 
+HA=('ha')
+for i in "${!HA[@]}"; do
+    LANG=${HA[i]}
+    split -l 68800 $SPM_DIR/train.en-$LANG.en -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.en.shard
+    split -l 68800 $SPM_DIR/train.en-$LANG.$LANG -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard
+
+    for j in "${!SHARD_SUB_DIR[@]}"; do
+        SUB_NUMBER=${SHARD_SUB_DIR[j]}
+        mv $SPM_SHARD_DIR/train.en-$LANG.en.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.en
+        mv $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.$LANG
+    done
+done
+
+# SPECIAL HA 18448 -> split to 3690 for one shard 
+KAB=('kab')
+for i in "${!KAB[@]}"; do
+    LANG=${KAB[i]}
+    split -l 3690 $SPM_DIR/train.en-$LANG.en -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.en.shard
+    split -l 3690 $SPM_DIR/train.en-$LANG.$LANG -d -a 2 $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard
+
+    for j in "${!SHARD_SUB_DIR[@]}"; do
+        SUB_NUMBER=${SHARD_SUB_DIR[j]}
+        mv $SPM_SHARD_DIR/train.en-$LANG.en.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.en
+        mv $SPM_SHARD_DIR/train.en-$LANG.$LANG.shard0${SUB_NUMBER} dataset/spm_sharded/shard${SUB_NUMBER}/train.en-$LANG.$LANG
+    done
+done
+
+# ------------------------ 4. Fairseq preparation Sharded ------------------------ #
+SPM_DATA_DIR=dataset/spm_sharded
+FAIRSEQ_DIR=dataset/fairseq-data-bin-sharded
+mkdir ${FAIRSEQ_DIR}
+
+cut -f1 dataset/spm/spm_64k.vocab | tail -n +4 | sed "s/$/ 100/g" > ${FAIRSEQ_DIR}/dict.txt
+
+SHARD_SUB_DIR=('0' '1' '2' '3' '4')
+for i in "${!SHARD_SUB_DIR[@]}"; do
+    SUB_NUMBER=${SHARD_SUB_DIR[i]}
+    mkdir $FAIRSEQ_DIR/shard${SUB_NUMBER}
+done
+
+# preprocess with mmap dataset
+for SHARD in $(seq 0 4); do
+    SRC=en
+    for TGT in bg so ca da be bs mt es uk am hi ro no ti de cs lb pt nl mr is ne ur oc ast ha sv kab gu ar fr ru it pl sr sd he af kn bn; do
+        fairseq-preprocess \
+            --dataset-impl mmap \
+            --source-lang ${SRC} \
+            --target-lang ${TGT} \
+            --trainpref ${SPM_DATA_DIR}/shard${SHARD}/train.${SRC}-${TGT} \
+            --destdir ${FAIRSEQ_DIR}/shard${SHARD} \
+            --thresholdtgt 0 \
+            --thresholdsrc 0 \
+            --workers 40 \
+            --srcdict ${FAIRSEQ_DIR}/dict.txt \
+            --tgtdict ${FAIRSEQ_DIR}/dict.txt
+    cp ${FAIRSEQ_DIR}/dict.txt ${FAIRSEQ_DIR}/shard${SHARD}/dict.txt
+    done
+done

toolkit/train-EC40-mTrans-large.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+#SBATCH --partition=gpu
+#SBATCH --gres=gpu:4
+#SBATCH --ntasks=4
+#SBATCH --cpus-per-task=11
+#SBATCH --job-name=m2m_base_shard
+#SBATCH --nodelist=ilps-cn116
+#SBATCH --time=20-00:00:00
+#SBATCH --mem=250G
+#SBATCH -o /ivi/ilps/personal/...m2m_baseline_monitor_new_large.o
+#SBATCH -e /ivi/ilps/personal/...m2m_baseline_monitor_new_large.e
+
+pwd
+conda info --envs
+source /home/stan1/anaconda3/bin/activate fairseq
+
+
+fairseq-train fairseq-data-bin-sharded/shard0:fairseq-data-bin-sharded/shard1:fairseq-data-bin-sharded/shard2:fairseq-data-bin-sharded/shard3:fairseq-data-bin-sharded/shard4 \
+    --langs en,de,nl,sv,da,is,af,lb,no,fr,es,it,pt,ro,oc,ast,ca,ru,cs,pl,bg,uk,sr,be,bs,hi,bn,kn,mr,sd,gu,ne,ur,ar,he,ha,mt,ti,am,kab,so \
+    --lang-pairs en-de,en-nl,en-sv,en-da,en-is,en-af,en-lb,en-no,en-fr,en-es,en-it,en-pt,en-ro,en-oc,en-ast,en-ca,en-ru,en-cs,en-pl,en-bg,en-uk,en-sr,en-be,en-bs,en-hi,en-bn,en-kn,en-mr,en-sd,en-gu,en-ne,en-ur,en-ar,en-he,en-ha,en-mt,en-ti,en-am,en-kab,en-so,de-en,nl-en,sv-en,da-en,is-en,af-en,lb-en,no-en,fr-en,es-en,it-en,pt-en,ro-en,oc-en,ast-en,ca-en,ru-en,cs-en,pl-en,bg-en,uk-en,sr-en,be-en,bs-en,hi-en,bn-en,kn-en,mr-en,sd-en,gu-en,ne-en,ur-en,ar-en,he-en,ha-en,mt-en,ti-en,am-en,kab-en,so-en \
+    --encoder-langtok tgt \
+    --arch transformer_vaswani_wmt_en_de_big \
+    --encoder-normalize-before --decoder-normalize-before --layernorm-embedding \
+    --encoder-layers 12 --decoder-layers 12 \
+    --sampling-method temperature --sampling-temperature 5 \
+    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+    --skip-invalid-size-inputs-valid-test \
+    --max-tokens 10240 --update-freq 21 --max-update 900000 \
+    --share-all-embeddings \
+    --max-source-positions 256 --max-target-positions 256 \
+    --lr 0.0005 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+    --seed 1234 --patience 10 \
+    --optimizer adam --adam-betas '(0.9, 0.98)' --adam-eps 1e-06 --weight-decay 0.0 \
+    --dropout 0.1 --attention-dropout 0.1 \
+    --fp16 --ddp-backend no_c10d \
+    --wandb-project 'EC40' \
+    --checkpoint-suffix _m2m_ --save-dir checkpoints/m2m_base_monitor_shard_new_large \
+    --save-interval-updates 2000 --keep-interval-updates 5 --no-epoch-checkpoints --log-interval 100 \
+    --distributed-world-size 4 --distributed-num-procs 44 --ddp-comm-hook fp16