NLP训练&落地的标准流程

代码示例

数据下载

if [ ! -f ./nlu_train.csv ]; then
  wget https://atp-modelzoo-sh.oss-cn-shanghai.aliyuncs.com/release/tutorials/knowledge_nlu/nlu_train.csv
fi
if [ ! -f ./nlu_dev.csv ]; then
  wget https://atp-modelzoo-sh.oss-cn-shanghai.aliyuncs.com/release/tutorials/knowledge_nlu/nlu_dev.csv
fi

标准Finetune测试脚本

easynlp \
    --mode=train \
    --worker_count=1 \
    --worker_gpu=1 \
    --tables=nlu_train.csv,nlu_dev.csv \
    --input_schema=label:str:1,text1:str:1,text2:str:1 \
    --first_sequence=text1 \
    --second_sequence=text2 \
    --label_name=label \
    --label_enumerate_values=0,1 \
    --checkpoint_dir=./base_model \
    --learning_rate=3e-5  \
    --epoch_num=3  \
    --random_seed=42 \
    --save_checkpoint_steps=200 \
    --sequence_length=128 \
    --micro_batch_size=32 \
    --app_name=text_match \
    --user_defined_parameters="pretrain_model_name_or_path=bert-base-chinese"

知识增强预训练测试脚本

easynlp \
    --mode=train \
    --worker_count=1 \
    --worker_gpu=1 \
    --tables=nlu_train.csv,nlu_dev.csv \
    --input_schema=label:str:1,text1:str:1,text2:str:1 \
    --first_sequence=text1 \
    --second_sequence=text2 \
    --label_name=label \
    --label_enumerate_values=0,1 \
    --checkpoint_dir=./dkplm_model \
    --learning_rate=3e-5  \
    --epoch_num=3  \
    --random_seed=42 \
    --save_checkpoint_steps=200 \
    --sequence_length=128 \
    --micro_batch_size=32 \
    --app_name=text_match \
    --user_defined_parameters="pretrain_model_name_or_path=alibaba-pai/pai-dkplm-medical-base-zh"

知识蒸馏测试脚本：

# forward teacher logits
easynlp \
    --mode=predict \
    --worker_count=1 \
    --worker_gpu=1 \
    --tables=nlu_train.csv \
    --outputs=logits_pred.csv \
    --input_schema=label:str:1,text1:str:1,text2:str:1 \
    --output_schema=logits \
    --first_sequence=text1 \
    --second_sequence=text2 \
    --checkpoint_path=./dkplm_model \
    --micro_batch_size=32 \
    --sequence_length=128 \
    --app_name=text_match
easynlp \
    --app_name=data_augmentation \
    --worker_count=1 \
    --worker_gpu=1 \
    --mode=predict \
    --tables=nlu_train.csv \
    --input_schema=label:str:1,text1:str:1,text2:str:1 \
    --first_sequence=text1 \
    --second_sequence=text2 \
    --label_name    =label \
    --outputs=nlu_aug.tsv \
    --output_schema=augmented_data \
    --checkpoint_dir=_ \
    --sequence_length=128 \
    --micro_batch_size=32 \
    --user_defined_parameters="
        pretrain_model_name_or_path=alibaba-pai/pai-dkplm-medical-base-zh
        type=mlm_da
        expansion_rate=10
        mask_proportion=0.1
        remove_blanks=True
    "
# train student

小样本学习测试脚本

详见：https://www.yuque.com/easyx/easynlp/ochmnf