BERT NER下载 - BERT NER源代码下载

BERT NER

Python

1.0.0

下载

伯特·纳尔（Bert ner）

使用Google Bert做Conll-2003 NER！

使用Python和使用C ++推理的火车模型

Albert-TF2.0

Bert-ner-TensorFlow-2.0

伯特很面

要求

python3
pip3 install -r requirements.txt

跑步

python run_ner.py --data_dir=data/ --bert_model=bert-base-cased --task_name=ner --output_dir=out_base --max_seq_length=128 --do_train --num_train_epochs 5 --do_eval --warmup_proportion=0.1

结果

伯特基

验证数据

             precision    recall  f1-score   support

        PER     0.9677    0.9745    0.9711      1842
        LOC     0.9654    0.9711    0.9682      1837
       MISC     0.8851    0.9111    0.8979       922
        ORG     0.9299    0.9292    0.9295      1341

avg / total     0.9456    0.9534    0.9495      5942

测试数据

             precision    recall  f1-score   support

        PER     0.9635    0.9629    0.9632      1617
        ORG     0.8883    0.9097    0.8989      1661
        LOC     0.9272    0.9317    0.9294      1668
       MISC     0.7689    0.8248    0.7959       702

avg / total     0.9065    0.9209    0.9135      5648

验证模型从这里下载

Bert-large

验证数据

             precision    recall  f1-score   support

        ORG     0.9288    0.9441    0.9364      1341
        LOC     0.9754    0.9728    0.9741      1837
       MISC     0.8976    0.9219    0.9096       922
        PER     0.9762    0.9799    0.9781      1842

avg / total     0.9531    0.9606    0.9568      5942

测试数据

             precision    recall  f1-score   support

        LOC     0.9366    0.9293    0.9329      1668
        ORG     0.8881    0.9175    0.9026      1661
        PER     0.9695    0.9623    0.9659      1617
       MISC     0.7787    0.8319    0.8044       702

avg / total     0.9121    0.9232    0.9174      5648

验证模型从这里下载

推理

 from bert import Ner

model = Ner ( "out_base/" )

output = model . predict ( "Steve went to Paris" )

print ( output )
'''
    [
        {
            "confidence": 0.9981840252876282,
            "tag": "B-PER",
            "word": "Steve"
        },
        {
            "confidence": 0.9998939037322998,
            "tag": "O",
            "word": "went"
        },
        {
            "confidence": 0.999891996383667,
            "tag": "O",
            "word": "to"
        },
        {
            "confidence": 0.9991968274116516,
            "tag": "B-LOC",
            "word": "Paris"
        }
    ]
'''

推断C ++

从这里预估计和转换的bert-base模型下载

从这里下载libtorch

安装cmake ，使用cmake版本3.10.2测试
BERT-NER中下载的型号和libtorch

编译C ++应用

  cd cpp-app/
  cmake -DCMAKE_PREFIX_PATH=../libtorch

CMAKE输出图像

make

制作输出图像

运行应用程序
```
   ./app ../base
```

NB：Bert-Base C ++模型分为两个部分。

BERT特征提取器和NER分类器。
之所以这样做，是因为jit trace不支持input for循环或model的forword函数中的if 。

部署REST-API

Bert NER模型部署为REST API

python api.py

API将在0.0.0.0:8000端点predict直播

卷曲请求

curl -X POST http://0.0.0.0:8000/predict -H 'Content-Type: application/json' -d '{ "text": "Steve went to Paris" }'

输出

{
    "result" : [
        {
            "confidence" : 0.9981840252876282 ,
            "tag" : " B-PER " ,
            "word" : " Steve "
        },
        {
            "confidence" : 0.9998939037322998 ,
            "tag" : " O " ,
            "word" : " went "
        },
        {
            "confidence" : 0.999891996383667 ,
            "tag" : " O " ,
            "word" : " to "
        },
        {
            "confidence" : 0.9991968274116516 ,
            "tag" : " B-LOC " ,
            "word" : " Paris "
        }
    ]
}