scripts/inference.sh

#! /bin/bash
set -eux
ROOT_DIR=$(dirname $(dirname `readlink -f $0`))

export HF_HOME="$ROOT_DIR/cache/"
export MODELSCOPE_CACHE="$ROOT_DIR/cache/"
export HF_EVALUATE_OFFLINE=1
export HF_DATASETS_OFFLINE=1

config_file=$ROOT_DIR/configs/accelerate_config.yaml
export NPROC_PER_NODE=8


# model
predict_model_dir=${1:-""}


# eval
comet_model=$ROOT_DIR/model_card/wmt22-comet-da/checkpoints/model.ckpt 
xcome_model=$ROOT_DIR/model_card/XCOMET-XXL/checkpoints/model.ckpt


lang_pair_strs=""
src_file_strs=""
ref_file_strs=""
hypo_file_strs=""

# for lang in en ja ru de ug; do
for lang in en de ru bn hi th jv sw si km; do
# for lang in en ja ko ru de fr it pt es;do
    for src in $lang zh ;do 

        if [ $src = "zh" ]; then # en2zh
            src_lang=zh
            tgt_lang=$lang 
        else  # zh2en
            src_lang=$lang
            tgt_lang=zh 
        fi

        lp=${src_lang}2${tgt_lang}
        src_file=$ROOT_DIR/data_arr/flores200/zh-${lang}/test.zh-$lang.$src_lang
        ref_file=$ROOT_DIR/data_arr/flores200/zh-${lang}/test.zh-$lang.$tgt_lang
        # test_file=$ROOT_DIR/data_arr/sft_100k_ugbomn/test.$lp.jsonl
        test_file=$ROOT_DIR/data_arr/test/test.$lp.jsonl

        output_dir=$predict_model_dir/decode_result/$lp
        mkdir -p $output_dir
      ############################!!!!!
        rm -rf $output_dir/*
        #######################
        cp $0 $output_dir

        swift infer \
            --infer_backend pt \
            --val_dataset $test_file \
            --load_from_cache_file True \
            --dataset_shuffle False \
            --val_dataset_shuffle False \
            --model $predict_model_dir \
            --torch_dtype bfloat16 \
            --max_new_tokens 1024 \
            --max_batch_size 16 \
            --num_beams 5 \
            --max_length 1024 \
            --dataset_num_proc 8 \
            --temperature 0 \
            --result_path $output_dir/generated_predictions.jsonl | tee $output_dir/train.log

        jq -r '.response' $output_dir/generated_predictions.jsonl > $output_dir/hypo.$lp.txt

        hypo_file=$output_dir/hypo.$lp.txt

        lang_pair_strs=${lang_pair_strs:+$lang_pair_strs,}$lp
        src_file_strs=${src_file_strs:+$src_file_strs,}$src_file
        ref_file_strs=${ref_file_strs:+$ref_file_strs,}$ref_file
        hypo_file_strs=${hypo_file_strs:+$hypo_file_strs,}$hypo_file
    done
done


# # metric="bleu,comet_22,xcomet_xxl" 
metric="bleu,comet_22" 
python $ROOT_DIR/src/mt_scoring.py \
    --metric $metric  \
    --comet_22_path $comet_model \
    --xcomet_xxl_path $xcome_model \
    --lang_pair $lang_pair_strs \
    --src_file $src_file_strs \
    --ref_file $ref_file_strs \
    --hypo_file $hypo_file_strs \
    --record_file "result_mt.xlsx"


# lang_pair_strs=""
# src_file_strs=""
# ref_file_strs=""
# hypo_file_strs=""

# # 
# # for lang in ja ko ru de fr it pt es;do
# # for lang in ja ru de ug; do
# for lang in mn_cn; do
#     for src in $lang en ;do 

#         if [ $src = "en" ]; then # en2zh
#             src_lang=en
#             tgt_lang=$lang 
#         else  # zh2en
#             src_lang=$lang
#             tgt_lang=en 
#         fi

#         lp=${src_lang}2${tgt_lang}
#         src_file=$ROOT_DIR/data_arr/flores200/en-${lang}/test.en-$lang.$src_lang
#         ref_file=$ROOT_DIR/data_arr/flores200/en-${lang}/test.en-$lang.$tgt_lang
#         # test_file=$ROOT_DIR/data_arr/sft_100k_ugbomn/test.$lp.jsonl
#         test_file=/mnt/nvme1/luoyingfeng/llm-mt/data_arr/merge_0701/train1/test/test.$lp.jsonl

#         output_dir=$predict_model_dir/decode_result/$lp
#         mkdir -p $output_dir
#         #############################!!!!!
#         rm -rf $output_dir/*
#         ########################
#         cp $0 $output_dir

#         #             --load_args False \
#         swift infer \
#             --infer_backend pt \
#             --val_dataset $test_file \
#             --load_from_cache_file True \
#             --dataset_shuffle False \
#             --val_dataset_shuffle False \
#             --model $predict_model_dir \
#             --torch_dtype bfloat16 \
#             --max_new_tokens 1024 \
#             --max_batch_size 8 \
#             --num_beams 5 \
#             --max_length 1024 \
#             --dataset_num_proc 8 \
#             --temperature 0 \
#             --result_path $output_dir/generated_predictions.jsonl | tee $output_dir/train.log


#         jq -r '.response' $output_dir/generated_predictions.jsonl > $output_dir/hypo.$lp.txt
        
#         hypo_file=$output_dir/hypo.$lp.txt

#         lang_pair_strs=${lang_pair_strs:+$lang_pair_strs,}$lp
#         src_file_strs=${src_file_strs:+$src_file_strs,}$src_file
#         ref_file_strs=${ref_file_strs:+$ref_file_strs,}$ref_file
#         hypo_file_strs=${hypo_file_strs:+$hypo_file_strs,}$hypo_file

#     done
# done

# # metric="bleu,comet_22,xcomet_xxl" 
# metric="bleu,comet_22" 
# python $ROOT_DIR/src/mt_scoring.py \
#     --metric $metric  \
#     --comet_22_path $comet_model \
#     --xcomet_xxl_path $xcome_model \
#     --lang_pair $lang_pair_strs \
#     --src_file $src_file_strs \
#     --ref_file $ref_file_strs \
#     --hypo_file $hypo_file_strs \
#     --record_file "result_mt.xlsx"