get_mapping.sh

#!/bin/bash

set -e

mkdir -p cache

GPU=0
L1=en
L2=es
MODEL=bert-base-multilingual-cased
FAST_ALIGN=./tools/fast_align/build/
OUTPUT_PATH=./
CACHE=./cache
PARA=europarl-v7.$L1-$L2.30k

# download and prep parallel data
wget https://www.statmt.org/europarl/v7/$L2-$L1.tgz
tar -zxvf $L2-$L1.tgz -C $OUTPUT_PATH/
rm -f $L2-$L1.tgz
python prep_parallel.py --bert_model $MODEL --lang1 $OUTPUT_PATH/europarl-v7.$L2-$L1.$L1 --lang2 $OUTPUT_PATH/europarl-v7.$L2-$L1.$L2 --size 30000 --output $OUTPUT_PATH/$PARA

# get word alignment
$FAST_ALIGN/fast_align -i $OUTPUT_PATH/$PARA.uncased -d -o -v > $OUTPUT_PATH/forward.align.$PARA.uncased
$FAST_ALIGN/fast_align -i $OUTPUT_PATH/$PARA.uncased -d -o -v -r > $OUTPUT_PATH/reverse.align.$PARA.uncased
$FAST_ALIGN/atools -i $OUTPUT_PATH/forward.align.$PARA.uncased -j $OUTPUT_PATH/reverse.align.$PARA.uncased -c grow-diag-final-and > $OUTPUT_PATH/sym.align.$PARA.uncased

# produce alignment matrix (for the last 4 layers)
for l in 9 10 11 12; do
    CUDA_VISIBLE_DEVICES=$GPU python get_aligned_features_avgbpe.py \
			--para_file $OUTPUT_PATH/$PARA.cased \
			--align_file $OUTPUT_PATH/sym.align.$PARA.uncased \
			--bert_model $MODEL \
			--cache_dir $CACHE \
			--layer $l \
			--output_file $OUTPUT_PATH/$PARA.$l
    
    python train_mapping.py --emb_file $OUTPUT_PATH/$PARA.$l --output_file $OUTPUT_PATH/$PARA.$l --orthogonal
done