run_glue_gazesup_bert_high_resource.sh

#!/bin/bash

#set learning rate
# task_to_lr = {'rte': 2e-5,
#         'mrpc': 3e-5,
#         'stsb': 4e-5,
#         'sst2': 2e-5,
#         'cola': 2e-5,
#         'qqp': 2e-5,
#         'mnli': 2e-5,
#         'qnli': 2e-5}


export TASK_NAME=cola
export LR=2e-05
export AUG_WEIGHT=0.5


CUDA_VISIBLE_DEVICES=6 python train_glue_gazesup_bert_high_resource.py \
    --model_name_or_path bert-base-cased \
    --task_name $TASK_NAME \
    --output_dir result/gazesup_bert/$TASK_NAME/All/$AUG_WEIGHT/ \
    --num_train_epochs 10 \
    --learning_rate $LR \
    --per_device_train_batch_size 32 \
    --max_seq_length 128 \
    --augweight $AUG_WEIGHT \
    --evaluation_strategy epoch \
    --save_strategy epoch \
    --metric_for_best_model matthews_correlation \
    --train_as_val False \
    --load_best_model_at_end \
    --overwrite_output_dir \
    --do_train \
    --do_eval \
    --do_predict \
    --fp16 \
    "$@"