Steps:

1. git clone https://github.com/google-research/bert
2. prepare data, download pre-trained models
3. modify code in run_classifier.py
   1. add a new processor

      

      ?
      ?

   2. add the processor in main function

      ?
      ?

      

?
?

Train and predict

1. train

   python run_classifier.py

   --task_name=multiclass

   --do_train=true

   --do_eval=true

   --data_dir=/home/wxl/bertProject/bertTextClassification/data

   --vocab_file=/home/wxl/bertProject/chinese_L-12_H-768_A-12/vocab.txt

   --bert_config_file=/home/wxl/bertProject/chinese_L-12_H-768_A-12/bert_config.json

   --init_checkpoint=/home/wxl/bertProject/chinese_L-12_H-768_A-12/bert_model.ckpt

   --max_seq_length=128

   --train_batch_size=16

   --learning_rate=2e-5

   --num_train_epochs=100.0

   --output_dir=/home/wxl/bertProject/bertTextClassification/outputThree/

   ?
   ?

   you would get the following result if success:

   

   ?
   ?

   ?
   ?

   ?
   ?

2. predict

   python run_classifier.py

   --task_name=multiclass

   --do_predict=true

   --data_dir=/home/wxl/bertProject/bertTextClassification/data

   --vocab_file=/home/wxl/bertProject/chinese_L-12_H-768_A-12/vocab.txt

   --bert_config_file=/home/wxl/bertProject/chinese_L-12_H-768_A-12/bert_config.json

   --init_checkpoint=/home/wxl/bertProject/bertTextClassification/outputThreeV1

   --max_seq_length=128

   --output_dir=/home/wxl/bertProject/bertTextClassification/mulitiPredictThreeV1/

   ?
   ?

?
?

原文地址：https://www.cnblogs.com/wuxiangli/p/10432160.html