service.py

import bentoml
from bentoml.frameworks.pytorch import PytorchModelArtifact
from bentoml.service.artifacts.common import PickleArtifact, JSONArtifact
from bentoml.adapters import JsonInput

from easydict import EasyDict
import time
from datetime import datetime
import random
import pandas as pd

import inference


@bentoml.env(infer_pip_packages=True)
@bentoml.artifacts([PytorchModelArtifact('model'), PickleArtifact('test'), JSONArtifact('config'), \
    PickleArtifact('assessmentItemID_classes'), PickleArtifact('testId_classes'), PickleArtifact('KnowledgeTag_classes'), \
    PickleArtifact('paperID_classes'), PickleArtifact('head_classes'), PickleArtifact('mid_classes'), PickleArtifact('tail_classes')]) 
class PytorchDKT(bentoml.BentoService):
     
    #이번에는 input을 파일 자체로 입력받겠습니다.
    #마찬가지로 주소는 HOST:PORT/predict 입니다.
    @bentoml.api(input=JsonInput(), batch=False)
    def predict(self, data):
        # get config
        args = EasyDict(self.artifacts.config)
        
        # add new data into training dataset
        additional_df = pd.read_csv(args.data_dir + 'additional_data.csv')
        new_userID = list(additional_df.userID)[-1]+1 if len(additional_df)!=0 else 0
        new_userdata = []
        for d in data:
            row = [new_userID, d['assess_id'], d['test_id'], d['answer'], d['timestamp'], d['tag']]
            new_userdata.append(row)
        additional_df = pd.concat([additional_df, pd.DataFrame(new_userdata, columns=additional_df.columns)], axis=0)
        additional_df.to_csv(args.data_dir + 'additional_data.csv', index=False)

        # get label encoder list
        le = {}
        le['assessmentItemID'] = self.artifacts.assessmentItemID_classes
        le['testId'] = self.artifacts.testId_classes
        le['KnowledgeTag'] = self.artifacts.KnowledgeTag_classes
        le['paperID'] = self.artifacts.paperID_classes
        le['head'] = self.artifacts.head_classes
        le['mid'] = self.artifacts.mid_classes
        le['tail'] = self.artifacts.tail_classes

        # transform data into input structure
        user_data = []
        for d in data:
            if 'answer' in d:
                row = [d['assess_id'], d['test_id'],d['tag'], d['timestamp'], d['answer']]
                user_data.append(row)

        # set data data`s timestamp randomly (10~15 sec)
        last_timestamp = user_data[-1][-2]
        last_sec = time.mktime(datetime.strptime(last_timestamp, '%Y-%m-%d %H:%M:%S').timetuple()) 
        sec_list = []
        for _ in range(len(self.artifacts.test)):
            sec_list.append(datetime.fromtimestamp(last_sec+random.randint(5,15)).strftime('%Y-%m-%d %H:%M:%S'))
        self.artifacts.test['Timestamp'] = sec_list

        
        print('* data :  \n', data)
        print('-'*100)
        print('* user data : \n', user_data)
        print('-'*100)
        print('* test data : \n', self.artifacts.test)
        print('-'*100)
        print('* config : \n', self.artifacts.config)
        print('-'*100)
        print('* args : \n', args)
        print('-'*100)
        print('* model : \n',  self.artifacts.model)
        print('-'*100)
        print('* assessmentItemID_classes : \n',  self.artifacts.assessmentItemID_classes)
        print('-'*100)
        
        score = inference.inference(user_data, self.artifacts.test, self.artifacts.model, le, args)

        print('* score : \n', score)

        return score