BERT_BASE_PATH=./pretrained_model/chinese_L-12_H-768_A-12/
TASK_NAME="XNLI"
#DATA_PATH=./data/XNLI-1.0-demo/
DATA_PATH=./data/XNLI-MT-1.0-dwk/
INIT_CKPT_PATH=./output/step_50
SAVE_INFERENCE_PATH=./output/infer_step_50 ## 这个目录下会有个__model__文件，给在线infer用的，注意paddle的版本要用1.3.1以上的，1.3.0生成的这个目录有bug
python=../../python-2.7.14-paddle-1.3.1/bin/python


export FLAGS_enable_parallel_graph=1
export FLAGS_sync_nccl_allreduce=1
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export CPU_NUM=3 ## 设置跑的cpu核数

TASK_NAME='XNLI'
CKPT_PATH=./output/

function finetune_xnli()
{

### 如果有56核的cpu，会占114g。。；而如果是12核的cpu，只会占25g内存
DATA_PATH=./data/XNLI-MT-1.0-dwk/
$python -u run_classifier.py --task_name ${TASK_NAME} \
           --use_cuda false \
           --do_train true \
           --do_val true \
           --do_test true \
           --batch_size 1 \
           --in_tokens false \
           --init_pretraining_params ${BERT_BASE_PATH}/params \
           --data_dir ${DATA_PATH} \
           --vocab_path ${BERT_BASE_PATH}/vocab.txt \
           --checkpoints ${CKPT_PATH} \
           --save_steps 50 \
           --weight_decay  0.01 \
           --warmup_proportion 0.0 \
           --validation_steps 2500 \
           --epoch 1 \
           --max_seq_len 8 \
           --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
           --learning_rate 1e-4 \
           --skip_steps 1 \
           --random_seed 1
}

function save_inference_model() 
{

### 如果是56核cpu，会占22g内存..
DATA_PATH=./data/XNLI-1.0-demo/
$python -u predict_classifier.py --task_name ${TASK_NAME} \
           --use_cuda false \
           --batch_size 1 \
           --data_dir ${DATA_PATH} \
           --vocab_path ${BERT_BASE_PATH}/vocab.txt \
           --do_lower_case true \
           --init_checkpoint ${INIT_CKPT_PATH} \
           --max_seq_len 8 \
           --bert_config_path ${BERT_BASE_PATH}/bert_config.json \
           --do_predict true \
           --save_inference_model_path ${SAVE_INFERENCE_PATH}

}

function main()
{
    finetune_xnli
    [[ $? -ne 0 ]] && exit 1
    save_inference_model
    [[ $? -ne 0 ]] && exit 1
    return 0
}

main 2>&1 

TASK_NAME="xnli"
DATA_PATH=../data/XNLI-1.0/
BERT_BASE_PATH=../pretrained_model/chinese_L-12_H-768_A-12/
python=../../../python-2.7.14-paddle-1.3.1/bin/python
$python gen_demo_data.py \
           --task_name ${TASK_NAME} \
           --data_path ${DATA_PATH} \
           --vocab_path "${BERT_BASE_PATH}/vocab.txt" \
           --batch_size 1 \
           > bert_data.demo

#           --in_tokens \

INFERENCE_MODEL_PATH=./output/infer_step_50
DATA_PATH=./bert_data.demo
REPEAT_TIMES=1
./bin/bert_demo --logtostderr \
        --model_dir $INFERENCE_MODEL_PATH \
        --data $DATA_PATH \
        --repeat $REPEAT_TIMES \
        --output_prediction