eval.sh

export PYTHONPATH="./:${PYTHONPATH}"
export CUDA_VISIBLE_DEVICES=5

PRETRAINED_DIR="./models/pretrained_dorm"
DATE_DIR="data"


index=1
learning_rate=75e-6
pinyin_weight=0.2
add_pinyin_mask="True"
add_pinyin_loss="True"
use_kl="True"
kl_weight=1.5
second_loss_weight=0.97
gradient_accumulation_steps=1
random_seed=42

OUTPUT_DIR="./models/sota_dorm"

for year in {13..15}
do
    python -m torch.distributed.launch --master_port=24679 --nproc_per_node=1 src/dorm_finetune.py \
        --model_type dorm \
        --model_name_or_path $PRETRAINED_DIR \
        --image_model_type 0 \
        --output_dir $OUTPUT_DIR  \
        --do_eval --do_predict \
        --data_dir $DATE_DIR \
        --train_file trainall.times2_pinyin2.pkl \
        --dev_file "test.sighan${year}_pinyin2.pkl" \
        --dev_label_file "test.sighan${year}.lbl.tsv" \
        --predict_file "test.sighan${year}_pinyin2.pkl" \
        --predict_label_file "test.sighan${year}.lbl.tsv" \
        --order_metric sent-correct-f1  \
        --max_seq_length 512 \
        --metric_reverse  \
        --num_save_ckpts 3 \
        --remove_unused_ckpts  \
        --per_gpu_train_batch_size 16 \
        --gradient_accumulation_steps $gradient_accumulation_steps \
        --per_gpu_eval_batch_size 1  \
        --learning_rate $learning_rate \
        --num_train_epochs 3  \
        --seed $random_seed \
        --warmup_steps 1000  \
        --overwrite_output_dir \
        --add_pinyin_mask \
        --add_pinyin_loss \
        --pinyin_weight $pinyin_weight \
        --use_kl \
        --kl_weight $kl_weight \
        --second_loss_weight $second_loss_weight \
        --save_steps 2000 \
        --comment "${year}-dorm-copy-test-"

done