EasyNLP带你实现中英文机器阅读理解( 五 ) _模型

当前已经对-base/large两种模型进行了集成，在中的模型调用名称分别为：hfl/-base-zh 和 hfl/-large-zh，两种模型的结构参数如下表所示。用户可直接在参数中指定名称进行调用。
【EasyNLP带你实现中英文机器阅读理解】中文结构参数
hfl/-base-zh
12-layer, 768-, 12-heads, 102M
hfl/-large-zh
24-layer, 1024-, 16-heads, 324M
我们也使用这两种模型分别在中文机器阅读理解数据上进行了，效果如下：
中文EMF1
hfl/-base-zh
61.35
78.62
hfl/-large-zh
59.15
79.33
上述所有训练/评估/预测代码，都已经被集成在////main.py 中。用户可以通过带参数运行main.py，一步执行上述所有训练/评估/预测操作。
# 模型训练python main.py \--mode train \--app_name=machine_reading_comprehension \--worker_gpu=1 \--tables=train.tsv,dev.tsv \--input_schema=qas_id:str:1,context_text:str:1,question_text:str:1,answer_text:str:1,start_position_character:str:1,title:str:1 \--first_sequence=question_text \--second_sequence=context_text \--sequence_length=384 \--checkpoint_dir=./model_dir \--learning_rate=3.5e-5 \--epoch_num=3 \--random_seed=42 \--save_checkpoint_steps=500 \--train_batch_size=16 \--user_defined_parameters='pretrain_model_name_or_path=hfl/macbert-base-zhlanguage=zhanswer_name=answer_textqas_id=qas_idstart_position_name=start_position_characterdoc_stride=128max_query_length=64'
# 模型预测python main.py \--mode predict \--app_name=machine_reading_comprehension \--worker_gpu=1 \--tables=dev.tsv \--outputs=dev.pred.csv \--input_schema=qas_id:str:1,context_text:str:1,question_text:str:1,answer_text:str:1,start_position_character:str:1,title:str:1 \--output_schema=unique_id,best_answer,query,context \--first_sequence=question_text \--second_sequence=context_text \--sequence_length=384 \--checkpoint_dir=./model_dir \--micro_batch_size=256 \--user_defined_parameters='pretrain_model_name_or_path=hfl/macbert-base-zhlanguage=zhqas_id=qas_idanswer_name=answer_textstart_position_name=start_position_charactermax_query_length=64max_answer_length=30doc_stride=128n_best_size=10output_answer_file=dev.ans.csv'
除了main.py一步执行之外，我们同样可以使用命令行执行的方式进行快速训练/预测，命令如下：
# 模型训练easynlp \--mode train \--app_name=machine_reading_comprehension \--worker_gpu=1 \--tables=train.tsv,dev.tsv \--input_schema=qas_id:str:1,context_text:str:1,question_text:str:1,answer_text:str:1,start_position_character:str:1,title:str:1 \--first_sequence=question_text \--second_sequence=context_text \--sequence_length=384 \--checkpoint_dir=./model_dir \--learning_rate=3.5e-5 \--epoch_num=5 \--random_seed=42 \--save_checkpoint_steps=600 \--train_batch_size=16 \--user_defined_parameters='pretrain_model_name_or_path=hfl/macbert-base-zh language=zhanswer_name=answer_textqas_id=qas_idstart_position_name=start_position_characterdoc_stride=128max_query_length=64'
# 模型预测easynlp \--mode predict \--app_name=machine_reading_comprehension \--worker_gpu=1 \--tables=dev.tsv \--outputs=dev.pred.csv \--input_schema=qas_id:str:1,context_text:str:1,question_text:str:1,answer_text:str:1,start_position_character:str:1,title:str:1 \--output_schema=unique_id,best_answer,query,context \--first_sequence=question_text \--second_sequence=context_text \--sequence_length=384 \--checkpoint_dir=./model_dir \--micro_batch_size=256 \--user_defined_parameters='pretrain_model_name_or_path=hfl/macbert-base-zh language=zhqas_id=qas_idanswer_name=answer_textstart_position_name=start_position_charactermax_query_length=64max_answer_length=30doc_stride=128n_best_size=10output_answer_file=dev.ans.csv'

文章插图
此外，我们还在////文件夹下封装好了多种可直接执行的bash脚本，用户同样可以通过使用bash文件命令行执行的方式来一步完成模型的训练/评估/预测。bash文件需要传入两个参数，第一个参数为运行程序的GPU编号，一般为0；第二个参数代表模型的训练/评估/预测。