Deep AM-FM

Inference Code associated with deep AM-FM

Installation

conda env create -f environment.yml
conda activate cuda101

Example commands

setting

The following is for variable settings in run.sh.

export CUDA_VISIBLE_DEVICES=0                                                                                                                                                                                                         
stage=1
stop_stage=5

# config files
# Data related
d_root=sample-data
SRC_LANG=en
TGT_LANG=km
TASK=ALT
SYSID=bad
# for translation from other languages to english
# uncomment the following line.
# am_model_path=./models/${TASK}/${TGT_LANG}-${SRC_LANG}_am
# for translation to other languages
# uncomment the following line
am_model_path=./models/${TASK}/${SRC_LANG}-${TGT_LANG}_am

fm_model_path=./models/${TASK}/${TGT_LANG}_lm
hyp_path=${d_root}/${TASK}/${TGT_LANG}/${SYSID}_hyp.txt
hyp_fm_output_path=${d_root}/${TASK}/${TGT_LANG}/${SYSID}_hyp.fm.prob
ref_path=${d_root}/${TASK}/${TGT_LANG}/${SYSID}_ref.txt
ref_fm_output_path=${d_root}/${TASK}/${TGT_LANG}/${SYSID}_ref.fm.prob
num_test_cases=200

Stage 1 - Compute AM Score

The following computes the AM scores for bad system on km language for ALT task.

if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
    echo "stage 1: Compute AM Score"
    python calc_am.py \
        --hyp_file=${hyp_path} \
        --ref_file=${ref_path} \
        --num_test=${num_test_cases} \
        --save_path=${result_path}/${TASK}_${SYSID}_${TGT_LANG}_am.score \
        --model_path=${am_model_path}
fi

Stage 2 to 4 - Compute FM Score

The following computes the FM scores for bad system on km language for ALT task.

if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
    echo "stage 2: Compute hypothesis sentence-level probability"
    python compute_ppl.py \
        --model_type=xlm-roberta \
        --output_dir=${fm_model_path} \
        --model_name_or_path=xlm-roberta-base \
        --do_eval \
        --eval_data_file=${hyp_path} \
        --overwrite_cache \
        --mlm
fi

if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
    echo "stage 3: Compute reference sentence-level probability"
    python compute_ppl.py \
        --model_type=xlm-roberta \
        --output_dir=${fm_model_path} \
        --model_name_or_path=xlm-roberta-base \
        --do_eval \
        --eval_data_file=${ref_path} \
        --overwrite_cache \
        --mlm
fi

if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
    echo "stage 4: Compute FM Score"
    python calc_fm.py \
        --hyp_file=${hyp_fm_output_path} \
        --ref_file=${ref_fm_output_path} \
        --num_test=${num_test_cases} \
        --save_path=${result_path}/${TASK}_${SYSID}_${TGT_LANG}_fm.score
fi

Stage 5 - Stage 5 - Compute final AM-FM Score