train_bert.sh

#!/bin/bash
# Before running this script, please make sure you ran pip install . in code_dir
read -p "Enter the training data data location(must be a path): " glue_dir
read -p "Enter the code dir(must be a path): " code_dir

# Train BERT with two different random seed
dataset_names=("RTE" "MRPC" "QNLI" "SST-2" "STS-B" "CoLA")

for dataset in "${dataset_names[@]}"
do
    python $code_dir/examples/run_glue.py \
    --model_name_or_path /mnt/d/glue_results/$dataset \
    --task_name $dataset \
    --do_eval \
    --data_dir $glue_dir/$dataset \
    --max_seq_length 128 \
    --per_gpu_eval_batch_size=8 \
    --per_gpu_train_batch_size=8 \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir /mnt/d/glue_results/$dataset \
    --attention_type standard \
    --dataset_name $dataset \
    --plot_fig 6

    python $code_dir/examples/run_glue.py \
    --model_name_or_path /mnt/d/glue_results/$dataset \
    --task_name $dataset \
    --do_eval \
    --data_dir $glue_dir/$dataset \
    --max_seq_length 128 \
    --per_gpu_eval_batch_size=8 \
    --per_gpu_train_batch_size=8 \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir /mnt/d/glue_results/$dataset \
    --attention_type effective \
    --dataset_name $dataset \
    --plot_fig 6
done