albert_xxlargev1_sqd2_512.sh

#!/bin/bash

export OMP_NUM_THREADS=12
EXAMPLES=/media/dn/work2/nlpu/transformers/examples
SQUAD=${EXAMPLES}/scripts/squad2.0
MODEL_PATH=${EXAMPLES}/scripts/runs/albert_xxlargev1_squad2_512_bs48_3ep_fp16

python ${EXAMPLES}/question-answering/run_squad.py \
  --model_type albert \
  --model_name_or_path albert-xxlarge-v1 \
  --do_train \
  --do_eval \
  --train_file ${SQUAD}/train-v2.0.json \
  --predict_file ${SQUAD}/dev-v2.0.json \
  --version_2_with_negative \
  --do_lower_case \
  --num_train_epochs 3 \
  --max_steps 8144 \
  --warmup_steps 814 \
  --learning_rate 3e-5 \
  --max_seq_length 512 \
  --doc_stride 128 \
  --per_gpu_train_batch_size 6 \
  --gradient_accumulation_steps 8 \
  --per_gpu_eval_batch_size 48 \
  --fp16 \
  --fp16_opt_level O1 \
  --threads 12 \
  --logging_steps 50 \
  --save_steps 3000 \
  --overwrite_output_dir \
  --output_dir ${MODEL_PATH}
$@