train_cmrc2018.sh

GS_BUCKET=gs://your-bucket
TPU_NAME=your-tpu-name
TPU_ZONE=your-tpu-zone
MODEL_OUTPUT_DIR=$GS_BUCKET/path-to-output-dir
python -u run_cmrc2018.py \
  --vocab_file=$GS_BUCKET/bert/chinese_L-12_H-768_A-12/vocab.txt \
  --bert_config_file=$GS_BUCKET/bert/chinese_L-12_H-768_A-12/bert_config.json \
  --init_checkpoint=$GS_BUCKET/bert/chinese_L-12_H-768_A-12/bert_model.ckpt \
  --do_train=True \
  --train_file=./cmrc2018/cmrc2018_train.json \
  --do_predict=True \
  --predict_file=./cmrc2018/cmrc2018_dev.json \
  --train_batch_size=64 \
  --predict_batch_size=32 \
  --num_train_epochs=2 \
  --max_seq_length=512 \
  --doc_stride=128 \
  --learning_rate=3e-5 \
  --do_lower_case=True \
  --output_dir=$MODEL_OUTPUT_DIR \
  --use_tpu=True \
  --tpu_name=$TPU_NAME \
  --tpu_zone=$TPU_ZONE