零样本文本分类应用：基于UTC的医疗意图多分类，打通数据标注-模型训练-模型调优-预测部署全流程。

pip install label-studio==1.7.2

label-studio start

#下载医疗意图分类数据集：
!wget https://bj.bcebos.com/paddlenlp/datasets/utc-medical.tar.gz
!tar -xvf utc-medical.tar.gz
!mv utc-medical data
!rm utc-medical.tar.gz

[{"id":26092,"annotations":[{"id":59,"completed_by":1,"result":[{"value":{"choices":["注意事项"]},"id":"7iya31L9oc","from_name":"sentiment","to_name":"text","type":"choices","origin":"manual"}],"was_cancelled":false,"ground_truth":false,"created_at":"2023-01-09T07:13:18.982993Z","updated_at":"2023-01-09T07:13:18.983032Z","lead_time":4.022,"prediction":{},"result_count":0,"task":26092,"parent_prediction":null,"parent_annotation":null}],"file_upload":"838fb89a-10-shot.txt","drafts":[],"predictions":[],"data":{"text":"烧氧割要注意那些问题"},"meta":{},"created_at":"2023-01-09T06:48:10.725717Z","updated_at":"2023-01-09T07:13:19.022666Z","inner_id":35,"total_annotations":1,"cancelled_annotations":0,"total_predictions":0,"comment_count":0,"unresolved_comment_count":0,"last_comment_updated_at":null,"project":5,"updated_by":1,"comment_authors":[]},
{"id":26091,"annotations":[{"id":4,"completed_by":1,"result":[{"value":{"choices":["病因分析"]}]

# 生成训练/验证集文件：
!python label_studio.py \
    --label_studio_file ./data/utc-medical/label_studio.json \
    --save_dir ./data \
    --splits 0.8 0.1 0.1 \
    --options ./data/utc-medical/label.txt

[2023-04-14 11:28:46,056] [    INFO] - Save 45 examples to ./data/train.txt.

[2023-04-14 11:28:46,057] [    INFO] - Save 6 examples to ./data/dev.txt.

[2023-04-14 11:28:46,057] [    INFO] - Save 6 examples to ./data/test.txt.

[2023-04-14 11:28:46,057] [    INFO] - Finished! It takes 0.00 seconds



{"text_a": "老年痴呆的症状有哪些", "text_b": "", "question": "", "choices": ["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议", "疾病表述", "后果表述", "注意事项", "功效作用", "医疗费用", "其他"], "labels": [5]}

##代码结构
├── deploy/simple_serving/ # 模型部署脚本
├── utils.py               # 数据处理工具
├── run_train.py           # 模型微调脚本
├── run_eval.py            # 模型评估脚本
├── label_studio.py        # 数据格式转换脚本
├── label_studio_text.md   # 数据标注说明文档
└── README.md

#安装最新版本paddlenlp
!pip install --upgrade paddlenlp

# 单卡启动：
!python run_train.py  \
    --device gpu \
    --logging_steps 10 \
    --save_steps 10 \
    --eval_steps 10 \
    --seed 1000 \
    --model_name_or_path utc-base \
    --output_dir ./checkpoint/model_best \
    --dataset_path ./data/ \
    --max_seq_length 512  \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --num_train_epochs 20 \
    --learning_rate 1e-5 \
    --do_train \
    --do_eval \
    --do_export \
    --export_model_dir ./checkpoint/model_best \
    --overwrite_output_dir \
    --disable_tqdm True \
    --metric_for_best_model macro_f1 \
    --load_best_model_at_end  True \
    --save_total_limit 1 \
    --save_plm

eval_loss: 0.3148668706417084, eval_micro_f1: 0.9848484848484849, eval_macro_f1: 0.9504132231404958, eval_runtime: 0.0757, eval_samples_per_second: 79.286, eval_steps_per_second: 39.643, epoch: 19.6957

[2023-04-13 17:02:45,941] [    INFO] -   epoch                    =    19.6957
[2023-04-13 17:02:45,941] [    INFO] -   train_loss               =     0.9758
[2023-04-13 17:02:45,942] [    INFO] -   train_runtime            = 0:00:45.91
[2023-04-13 17:02:45,942] [    INFO] -   train_samples_per_second =     19.602
[2023-04-13 17:02:45,942] [    INFO] -   train_steps_per_second   =      0.871

安装一下fast tokenizer
pip install --upgrade fast_tokenizer

metric_value = metrics[metric_to_check]
KeyError: 'eval_macro_f1'

# !python -u -m paddle.distributed.launch --gpus "0,1,2,3" run_train.py \
#     --device gpu \
#     --logging_steps 10 \
#     --save_steps 10 \
#     --eval_steps 10 \
#     --seed 1000 \
#     --model_name_or_path utc-base \
#     --output_dir ./checkpoint/model_best \
#     --dataset_path ./data/ \
#     --max_seq_length 512  \
#     --per_device_train_batch_size 2 \
#     --per_device_eval_batch_size 2 \
#     --gradient_accumulation_steps 8 \
#     --num_train_epochs 20 \
#     --learning_rate 1e-5 \
#     --do_train \
#     --do_eval \
#     --do_export \
#     --export_model_dir ./checkpoint/model_best \
#     --overwrite_output_dir \
#     --disable_tqdm True \
#     --metric_for_best_model macro_f1 \
#     --load_best_model_at_end  True \
#     --save_total_limit 1 \
#     --save_plm

!python run_eval.py \
    --model_path ./checkpoint/model_best \
    --test_path ./data/test.txt \
    --per_device_eval_batch_size 16 \
    --max_seq_len 512 \
    --output_dir ./checkpoint_test

[2023-04-13 17:06:59,413] [    INFO] -   test_loss               =     1.6392
[2023-04-13 17:06:59,413] [    INFO] -   test_macro_f1           =     0.8167
[2023-04-13 17:06:59,413] [    INFO] -   test_micro_f1           =     0.9394
[2023-04-13 17:06:59,413] [    INFO] -   test_runtime            = 0:00:00.87
[2023-04-13 17:06:59,413] [    INFO] -   test_samples_per_second =      6.835
[2023-04-13 17:06:59,413] [    INFO] -   test_steps_per_second   =      1.139

!pip install onnxruntime-gpu onnx onnxconverter-common
!pip install paddle2onnx
#如果出现这个报错 local variable 'paddle2onnx' referenced before assignment ，请安装上述库onnx 的包需要安装
#中途出现一些警告可以忽视

from pprint import pprint
from paddlenlp import Taskflow
schema = ["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议", "疾病表述", "后果表述", "注意事项", "功效作用", "医疗费用", "其他"]
# my_cls = Taskflow("zero_shot_text_classification", model="utc-base", schema=schema, task_path='/home/aistudio/checkpoint/model_best/plm', precision="fp16")
my_cls = Taskflow("zero_shot_text_classification", model="utc-base", schema=schema, task_path='/home/aistudio/checkpoint/model_best/plm')
#支持FP16半精度推理加速,需要安装onnx
pprint(my_cls(["老年斑为什么都长在面部和手背上","老成都市哪家内痔医院比较好怎么样最好？","中性粒细胞比率偏低"]))

[2023-04-14 11:45:11,057] [    INFO] - We are using <class 'paddlenlp.transformers.ernie.tokenizer.ErnieTokenizer'> to load 'utc-base'.

[2023-04-14 11:45:11,059] [    INFO] - Already cached /home/aistudio/.paddlenlp/models/utc-base/utc_base_vocab.txt

[2023-04-14 11:45:11,083] [    INFO] - tokenizer config file saved in /home/aistudio/.paddlenlp/models/utc-base/tokenizer_config.json

[2023-04-14 11:45:11,085] [    INFO] - Special tokens file saved in /home/aistudio/.paddlenlp/models/utc-base/special_tokens_map.json

[2023-04-14 11:45:11,088] [    INFO] - Assigning ['[O-MASK]'] to the additional_special_tokens key of the tokenizer

[{'predictions': [{'label': '病因分析', 'score': 0.7360146263899581}],

  'text_a': '老年斑为什么都长在面部和手背上'},

 {'predictions': [{'label': '就医建议', 'score': 0.9940570944549809}],

  'text_a': '老成都市哪家内痔医院比较好怎么样最好？'},

 {'predictions': [{'label': '指标解读', 'score': 0.6683004187689248}],

  'text_a': '中性粒细胞比率偏低'}]

from pprint import pprint
from paddlenlp import Taskflow
schema = ["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议", "疾病表述", "后果表述", "注意事项", "功效作用", "医疗费用", "其他"]
my_cls = Taskflow("zero_shot_text_classification", model="utc-base", schema=schema)
pprint(my_cls(["老年斑为什么都长在面部和手背上","老成都市哪家内痔医院比较好怎么样最好？","中性粒细胞比率偏低"]))

[2023-04-14 11:45:20,869] [    INFO] - We are using <class 'paddlenlp.transformers.ernie.tokenizer.ErnieTokenizer'> to load 'utc-base'.

[2023-04-14 11:45:20,872] [    INFO] - Already cached /home/aistudio/.paddlenlp/models/utc-base/utc_base_vocab.txt

[2023-04-14 11:45:20,897] [    INFO] - tokenizer config file saved in /home/aistudio/.paddlenlp/models/utc-base/tokenizer_config.json

[2023-04-14 11:45:20,900] [    INFO] - Special tokens file saved in /home/aistudio/.paddlenlp/models/utc-base/special_tokens_map.json

[2023-04-14 11:45:20,903] [    INFO] - Assigning ['[O-MASK]'] to the additional_special_tokens key of the tokenizer

[{'predictions': [], 'text_a': '老年斑为什么都长在面部和手背上'},

 {'predictions': [{'label': '就医建议', 'score': 0.9283481315032535},

                  {'label': '其他', 'score': 0.5083715719139965}],

  'text_a': '老成都市哪家内痔医院比较好怎么样最好？'},

 {'predictions': [{'label': '其他', 'score': 0.9437889944553786}],

  'text_a': '中性粒细胞比率偏低'}]

# 安装 fast_tokenizer 以及 GPU 版本 fastdeploy
pip install fast-tokenizer-python fastdeploy-gpu-python -f https://www.paddlepaddle.org.cn/whl/fastdeploy.html

# CPU 推理
python /home/aistudio/deploy/python/infer.py--model_dir /home/aistudio/checkpoint/model_best --device cpu
# GPU 推理
python /home/aistudio/deploy/python/infer.py --model_dir /home/aistudio/checkpoint/model_best --device gpu

# !pip install --user fast-tokenizer-python fastdeploy-gpu-python -f https://www.paddlepaddle.org.cn/whl/fastdeploy.html
#比较大1.4G 去终端安装

  File "/home/aistudio/.data/webide/pip/lib/python3.7/site-packages/fastdeploy/c_lib_wrap.py", line 166, in <module>
    raise RuntimeError("FastDeploy initalized failed!")
RuntimeError: FastDeploy initalized failed!

    predictor = Predictor(args, schema=["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议", "疾病表述", "后果表述", "注意事项", "功效作用", "医疗费用", "其他"])
    results = predictor.predict(["月经期间刮痧拔罐会引起身体什么","老年斑为什么都长在面部和手背上","成都市哪家内痔医院比较好怎么样最好？","中性粒细胞比率偏低"])

pip install paddlenlp >= 2.5.0

paddlenlp server server:app --workers 1 --host 0.0.0.0 --port 8190

python client.py

* Server 自定义参数

* schema替换

# Default schema
schema = ["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议", "疾病表述", "后果表述", "注意事项", "功效作用", "医疗费用", "其他"]

* 设置模型路径

# Default task_path
utc = Taskflow("zero_shot_text_classification", model="utc-base", task_path="../../checkpoint/model_best/plm", schema=schema)

* 多卡服务化预测PaddleNLP SimpleServing 支持多卡负载均衡预测，主要在服务化注册的时候，注册两个Taskflow的task即可，下面是示例代码

utc1 = Taskflow("zero_shot_text_classification", model="utc-base", task_path="../../checkpoint/model_best", schema=schema)
utc2 = Taskflow("zero_shot_text_classification", model="utc-base", task_path="../../checkpoint/model_best", schema=schema)
service.register_taskflow("taskflow/utc", [utc1, utc2])

from paddlenlp import Taskflow
schema = ["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议", "疾病表述", "后果表述", "注意事项", "功效作用", "医疗费用", "其他"]
utc = Taskflow("zero_shot_text_classification",
                   schema=schema,
                   model="utc-base",
                   max_seq_len=512,
                   batch_size=1,
                   pred_threshold=0.5,
                   precision="fp32")

* Client 自定义参数

# Changed to input texts you wanted
texts = ["中性粒细胞比率偏低"]

%cd /home/aistudio/deploy/simple_serving
!paddlenlp server server:app --workers 1 --host 0.0.0.0 --port 8190
#Error loading ASGI app. Could not import module "server".
#去终端执行即可

/home/aistudio/deploy/simple_serving

[2023-04-13 18:26:51,839] [    INFO] - starting to PaddleNLP SimpleServer...

[2023-04-13 18:26:51,840] [    INFO] - The PaddleNLP SimpleServer is starting, backend component uvicorn arguments as follows:

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [host]=0.0.0.0

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [port]=8190

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [log_level]=None

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [workers]=1

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [limit_concurrency]=None

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [limit_max_requests]=None

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [timeout_keep_alive]=15

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [app_dir]=/home/aistudio/deploy/simple_serving

[2023-04-13 18:26:51,840] [    INFO] -    the starting argument [reload]=False

[2023-04-13 18:26:52,037] [    INFO] - We are using <class 'paddlenlp.transformers.ernie.tokenizer.ErnieTokenizer'> to load 'utc-base'.

[2023-04-13 18:26:52,038] [    INFO] - Already cached /home/aistudio/.paddlenlp/models/utc-base/utc_base_vocab.txt

[2023-04-13 18:26:52,067] [    INFO] - tokenizer config file saved in /home/aistudio/.paddlenlp/models/utc-base/tokenizer_config.json

[2023-04-13 18:26:52,067] [    INFO] - Special tokens file saved in /home/aistudio/.paddlenlp/models/utc-base/special_tokens_map.json

[2023-04-13 18:26:52,069] [    INFO] - Assigning ['[O-MASK]'] to the additional_special_tokens key of the tokenizer

[2023-04-13 18:26:55,628] [    INFO] -    Taskflow  request [path]=/taskflow/utc is genereated.

INFO:     Started server process [1718]

INFO:     Waiting for application startup.

INFO:     Application startup complete.

INFO:     Uvicorn running on http://0.0.0.0:8190 (Press CTRL+C to quit)

^C

INFO:     Shutting down

INFO:     Waiting for application shutdown.

INFO:     Application shutdown complete.

INFO:     Finished server process [1718]

# Save at server.py
from paddlenlp import SimpleServer, Taskflow

schema = ["病情诊断", "治疗方案", "病因分析", "指标解读", "就医建议"]
utc = Taskflow("zero_shot_text_classification",
               model="utc-base",
               schema=schema,
               task_path="/home/aistudio/checkpoint/model_best/plm",
               precision="fp32")
app = SimpleServer()
app.register_taskflow("taskflow/utc", utc)

# %cd /home/aistudio/deploy/simple_serving
!python client.py

        ["一个留着长发和胡须的怪人，在地铁里穿着一件颜色鲜艳的衬衫。", "这件衬衫是新的。"],

        ["一个穿着绿色衬衫的妈妈和一个穿全黑衣服的男人在跳舞。", "两人都穿着白色裤子。"]]))