#!/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" # start fine-tuning
python3.7 -m paddle.distributed.launch --started_port 8899 --selected_gpus=1,2,3 bert_classifier.py\
--use_cuda true \ --do_train true \ --do_test true \ --batch_size 64 \ --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
