#!/bin/bash BERT_BASE_PATH="./data/pretrained_models/uncased_L-12_H-768_A-12/" TASK_NAME='MNLI' DATA_PATH="./data/glue_data/MNLI/" CKPT_PATH="./data/saved_model/mnli_models"
GPU_TO_USE=0,1,2,3
# start fine-tuning python -m paddle.distributed.launch --selected_gpus=$GPU_TO_USE --log_dir ./cls_log run_classifier.py \ --task_name ${TASK_NAME} \ --use_cuda true \ --use_data_parallel true \ --do_train true \ --do_test true \ --batch_size 64 \ --in_tokens false \ --init_pretraining_params ${BERT_BASE_PATH}/dygraph_params/ \ --data_dir ${DATA_PATH} \ --vocab_path ${BERT_BASE_PATH}/vocab.txt \ --checkpoints ${CKPT_PATH} \ --save_steps 1000 \ --weight_decay 0.01 \ --warmup_proportion 0.1 \ --validation_steps 100 \ --epoch 3 \ --max_seq_len 128 \ --bert_config_path ${BERT_BASE_PATH}/bert_config.json \ --learning_rate 5e-5 \ --skip_steps 10 \ --shuffle true
