인공지능팩토리 김태영 대표이사 인공지능 모델을 데이터셋에 맞게 대량으로 찍어내는 방법 (only 파이썬)

문제정의 ● 3천대 서버의 인공지능 기반 이상징후 탐지 모델을 만들어라! ● 데이터의 입력 형식과 라벨 형식은 동일 ● 데이터 분포는 3천대 서버마다 다름 (유사한 것도 있으나)

해결방안 ● 3천대 서버의 모든 데이터셋을 학습시킨 글로벌 모델을 생성 ● 각 서버마다 개별 이상징후 탐지 모델을 생성 즉 3천대의 인공지능 모델 생성

해결방안 ● 3천대 서버의 모든 데이터셋을 학습시킨 글로벌 모델을 생성 ○ 장점: 하나의 모델로 모든 서버에 대한 추론 가능 ○ 단점: 서버마다 패턴 및 데이터 분포가 다른 경우 높은 성능을 가진 글로벌 모델 생성이 쉽지 않음 ● 각 서버마다 개별 이상징후 탐지 모델을 생성 즉 3천대의 인공지능 모델 생성

해결방안 ● 3천대 서버의 모든 데이터셋을 학습시킨 글로벌 모델을 생성 ○ 장점: 하나의 모델로 모든 서버에 대한 추론 가능 ○ 단점: 서버마다 패턴 및 데이터 분포가 다른 경우 높은 성능을 가진 글로벌 모델 생성이 쉽지 않음 ● 각 서버마다 개별 이상징후 탐지 모델을 생성 즉 3천대의 인공지능 모델 생성 ○ 장점: 각 서버에 맞는 개인화된 모델을 생성할 수 있음 ○ 단점: 개별 서버마다 데이터셋 구축 및 모델 개발이 필요함

해결방안 ● 3천대 서버의 모든 데이터셋을 학습시킨 글로벌 모델을 생성 ○ 장점: 하나의 모델로 모든 서버에 대한 추론 가능 ○ 단점: 서버마다 패턴 및 데이터 분포가 다른 경우 높은 성능을 가진 글로벌 모델 생성이 쉽지 않음 ● 각 서버마다 개별 이상징후 탐지 모델을 생성 즉 3천대의 인공지능 모델 생성 ○ 장점: 각 서버에 맞는 개인화된 모델을 생성할 수 있음 ○ 단점: 개별 서버마다 데이터셋 구축 및 모델 개발이 필요함 >> mlops로 해보자

개념 traingen worker wo testgen worker wo predictgen worker wo train worker wo test worker wo predict worker wo train workflow test workflow predict workflow Today Talk!

데이터셋 구성

데이터셋 구성 MNIST Fashion-MNIST

데이터셋 구성 MNIST Fashion-MNIST mixNIST

데이터셋 asset ● asset ○ train ○ test ○ real ○ arch ○ model ○ score

ID 체계 ● id ○ task ○ train ○ test ○ arch ○ model

작업요청서(workorder) 와 작업자(worker) ● wo 구성 ○ wait ○ running ○ success ○ fail

작업요청서(workorder) 와 작업자(worker) ● wo - worker ○ traingen traingen_worker.py ○ testgen testgen_worker.py ○ predictgen predictgen_worker.py ○ train train_worker.py ○ test test_worker.py ○ predict predict_worker.py

● traingen 데이터셋 작업 TRD000001_x.npy TRD000001_y.npy datagen_train.py traingen_worker.py traingen.py traingen_TSK000001_ TRD000001.wo ? (arg)

1세대 - 모델 하나 만들어보자 ● 모델 개발 프로세스 train dataset test dataset architecture train architecture test model score architecture predict in out

1세대 - 모델 하나 만들어보자 ● 모델 개발 프로세스 train dataset test dataset architecture train architecture test model score architecture predict in out

1세대 - 모델 하나 만들어보자 ● 모델 개발 프로세스 train dataset test dataset architecture train architecture test model score architecture predict in out

1세대 - 모델 하나 만들어보자 ● 모델 개발 프로세스 TRD000001_x.npy TRD000001_y.npy TST000001_x.npy TST000001_y.npy ARC000001_train.py ARC000001_test.py MDL000001.h5 SCR000001.txt ARC000001_predict.py PRD000001_x.npy PRD000001_yhat.npy

1세대 - 모델 하나 만들어보자 ● Architecture (asset) ○ ARC000001_train.py ○ ARC000001_test.py ○ ARC000001_predict.py

● Architecture (asset) ○ ARC000001_train.py 1세대 - 모델 하나 만들어보자 x_train = np.load(train_x_fn) y_train = np.load(train_y_fn) x_train = x_train.reshape(60000, 784).astype('float32') / 255.0 y_train = utils.to_categorical(y_train) model = models.Sequential() model.add(layers.Dense(64, input_dim=28*28, activation='relu')) model.add(layers.Dense(32, activation='relu')) model.add(layers.Dense(10, activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) callbacks = [ keras.callbacks.ModelCheckpoint( model_fn, save_best_only=True, monitor="val_loss" ) ] hist = model.fit(x_train, y_train, validation_split=0.2, epochs=2, batch_size=32, callbacks=callbacks)

● Architecture (asset) ○ ARC000001_test.py 1세대 - 모델 하나 만들어보자 x_test = np.load(test_x_fn) y_test = np.load(test_y_fn) x_test = x_test.reshape(10000, 784).astype('float32') / 255.0 y_test = utils.to_categorical(y_test) model = keras.models.load_model(model_fn) test_loss, test_acc = model.evaluate(x_test, y_test) print("Test accuracy", test_acc) print("Test loss", test_loss) f = open(score_fn, 'wt') f.write(str(test_acc)) f.close() print("output: {0} {1}".format(score_fn, test_acc))

● Architecture (asset) ○ ARC000001_predict.py 1세대 - 모델 하나 만들어보자 x_predict = np.load(predict_fn) print('x_predict shape : ' + str(x_predict.shape)) x_predict = x_predict.reshape(100, 784).astype('float32') / 255.0 # model model = keras.models.load_model(model_fn) y_predict = model.predict(x_predict) # report y_predict_class = np.argmax(y_predict, axis=1) np.savetxt(out_fn, y_predict) print("output: {0} {1}".format(out_fn, y_predict_class))

작업요청서(workorder) 와 작업자(worker) ● wo - worker ○ traingen traingen_worker.py ○ testgen testgen_worker.py ○ predictgen predictgen_worker.py ○ train train_worker.py ○ test test_worker.py ○ predict predict_worker.py

학습/시험/추론 작업 TSK000001_model.csv train_record.py train_worker.py train.py train_TSK000001_ARC 000001_TRD000001_ MDL000001 ● train TRD000001_x.npy TRD000001_y.npy ARC000001_train.py MDL000001.h5

학습/시험/추론 작업 TSK000001_model.csv test_record.py test_worker.py test.py test_TSK000001_ARC00 0001_MDL000001_TST0 00001_SCR000001.wo ● test TST000001_x.npy TST000001_y.npy ARC000001_test.py SCR000001.txt MDL000001.h5

학습/시험/추론 작업 TSK000001_predict.csv predict_record.py predict_worker.py predict.py predcit_TSK000001_ARC 000001_MDL000001_PR D000001.wo ● predict PRD000001_x.npy ARC000001_predict.py PRD000001_yhat.npy MDL000001.h5

시나리오를 짜보자 traingen worker wo testgen worker wo predictgen worker wo train worker wo test worker wo predict worker wo train workflow test workflow predict workflow

데이터셋 구성 MNIST Fashion-MNIST mixNIST ● 100개의 mixNIST 준비

시나리오 ● 100개의 태스크 ● 100개의 시험셋 생성 >> testgen_wo ○ testgen_TSK000001_TST000001.wo ○ ... ○ testgen_TSK000100_TST000100.wo ● 모델 1개로 100개의 태스크 평가 >> test_wo ○ test_TSK000001_ARC000001_MDL000001_TST000001_SCR000001.wo ○ … ○ test_TSK000100_ARC000001_MDL000001_TST000100_SCR000100.wo

시나리오 ● 1단계 - 100개 중 10개 태스크의 MDL000001에 대한 스코어 확인 >> 60%이상(3/10) ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 ○ TSK000004_score 0.580299973487854 ○ TSK000005_score 0.29809999465942383 ○ TSK000006_score 0.3885999917984009 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 ○ TSK000009_score 0.5809999704360962 ○ TSK000010_score 0.5033000111579895

시나리오 ● 2단계 - 실패 태스크 중 한 개를 재학습 시킴 ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 ○ TSK000004_score 0.580299973487854 ○ TSK000005_score 0.29809999465942383 ○ TSK000006_score 0.3885999917984009 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 ○ TSK000009_score 0.5809999704360962 ○ TSK000010_score 0.5033000111579895

시나리오 ● 2단계 - 실패 태스크 중 한 개(TSK000004)를 재학습 시킴 ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 ■ traingen_TSK000003_TRD000003.wo ■ train_TSK000003_ARC000001_TRD000003_MDL000002.wo ● MDL000002.h5 생성

시나리오 ● 3단계 - 신규 모델로 실패 태스크들 평가 >> 60%이상(6/10) ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 ○ TSK000006_score 0.3885999917984009 0.48890000581741333 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 0.26409998536109924 ○ TSK000009_score 0.5809999704360962 0.6718000173568726 ○ TSK000010_score 0.5033000111579895 0.7215999960899353

시나리오 ● 4단계 - 실패 태스크 중 한 개(TSK000004)를 재학습 시킴 ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 ■ traingen_TSK000004_TRD000004.wo ■ train_TSK000004_ARC000001_TRD000004_MDL000003.wo ● MDL000003.h5 생성

시나리오 ● 5단계 - 신규 모델로 실패 태스크들 평가 >> 60%이상(7/10) ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 0.95169997215271 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 0.4551999866962433 ○ TSK000006_score 0.3885999917984009 0.48890000581741333 0.1986999958753585 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 0.26409998536109924 0.394899994134902 ○ TSK000009_score 0.5809999704360962 0.6718000173568726 ○ TSK000010_score 0.5033000111579895 0.7215999960899353

시나리오 ● 6단계 - 실패 태스크 중 한 개(TSK000005)를 재학습 시킴 ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 0.95169997215271 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 0.4551999866962433 ■ traingen_TSK000005_TRD000005.wo ■ train_TSK000005_ARC000001_TRD000005_MDL000004.wo ● MDL000004.h5 생성

시나리오 ● 7단계 - 신규 모델로 실패 태스크들 평가 >> 60%이상(9/10) ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 0.95169997215271 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 0.4551999866962433 0.9332000017166138 ○ TSK000006_score 0.3885999917984009 0.48890000581741333 0.1986999958753585 0.6723999977111816 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 0.26409998536109924 0.394899994134902 0.09929999709129333 ○ TSK000009_score 0.5809999704360962 0.6718000173568726 ○ TSK000010_score 0.5033000111579895 0.7215999960899353

시나리오 ● 8단계 - 실패 태스크 중 한 개(TSK000008)를 재학습 시킴 ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 0.95169997215271 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 0.4551999866962433 0.9332000017166138 ○ TSK000006_score 0.3885999917984009 0.48890000581741333 0.1986999958753585 0.6723999977111816 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 0.26409998536109924 0.394899994134902 0.09929999709129333 ○ TSK000009_score 0.5809999704360962 0.6718000173568726 ○ TSK000010_score 0.5033000111579895 0.7215999960899353 ■ traingen_TSK000008_TRD000008.wo ■ train_TSK000008_ARC000001_TRD000008_MDL000005.wo ● MDL000005.h5 생성

시나리오 ● 9단계 - 신규 모델로 실패 태스크들 평가 >> 60%이상(10/10) ○ TSK000001_score 0.9731000065803528 ○ TSK000002_score 0.7840999960899353 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 ○ TSK000004_score 0.580299973487854 0.46950000524520874 0.95169997215271 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 0.4551999866962433 0.9332000017166138 ○ TSK000006_score 0.3885999917984009 0.48890000581741333 0.1986999958753585 0.6723999977111816 ○ TSK000007_score 0.6876000165939331 ○ TSK000008_score 0.5837000012397766 0.26409998536109924 0.394899994134902 0.09929999709129333 0.916700005531311 ○ TSK000009_score 0.5809999704360962 0.6718000173568726 ○ TSK000010_score 0.5033000111579895 0.7215999960899353

시나리오 ● 정리 ○ 시험셋 생성 10개 100% ○ 훈련셋 생성 5개 50% ○ 훈련 횟수 5번 50% ○ 모델 갯수 5개 50% ○ 아키텍처 갯수 1개 10%

두번째 시나리오 - 새로운 아키텍처! ● ARC000001 ○ model = models.Sequential() ○ model.add(layers.Dense(64, input_dim=28*28, activation='relu')) ○ model.add(layers.Dense(32, activation='relu')) ○ model.add(layers.Dense(10, activation='softmax')) ● ARC000002 ○ model = models.Sequential() ○ model.add(layers.Conv2D(32, (5, 5), padding='valid', input_shape=(28, 28, 1), activation='relu')) ○ model.add(layers.MaxPooling2D(pool_size=(2, 2))) ○ model.add(layers.Conv2D(64, (3, 3), activation="relu")) ○ model.add(layers.MaxPooling2D(pool_size=(2, 2))) ○ model.add(layers.Flatten()) ○ model.add(layers.Dropout(0.5)) ○ model.add(layers.Dense(10, activation="softmax"))

두번째 시나리오 - 새로운 아키텍처! ● ARC000001 ○ model = models.Sequential() ○ model.add(layers.Dense(64, input_dim=28*28, activation='relu')) ○ model.add(layers.Dense(32, activation='relu')) ○ model.add(layers.Dense(10, activation='softmax')) ● ARC000002 ○ model = models.Sequential() ○ model.add(layers.Conv2D(32, (5, 5), padding='valid', input_shape=(28, 28, 1), activation='relu')) ○ model.add(layers.MaxPooling2D(pool_size=(2, 2))) ○ model.add(layers.Conv2D(64, (3, 3), activation="relu")) ○ model.add(layers.MaxPooling2D(pool_size=(2, 2))) ○ model.add(layers.Flatten()) ○ model.add(layers.Dropout(0.5)) ○ model.add(layers.Dense(10, activation="softmax"))

두번째 시나리오 - 새로운 아키텍처! ● train_TSK000001_ARC000001_TRD000001_MDL000001.wo ● train_TSK000003_ARC000001_TRD000003_MDL000002.wo ● train_TSK000004_ARC000001_TRD000004_MDL000003.wo ● train_TSK000005_ARC000001_TRD000005_MDL000004.wo ● train_TSK000008_ARC000001_TRD000008_MDL000005.wo ● train_TSK000001_ARC000002_TRD000001_MDL000006.wo

두번째 시나리오 - 새로운 아키텍처! ● 9단계 - 신규 모델로 실패 태스크들 평가 >> 60%이상(10/10) ○ TSK000001_score 0.9731000065803528 0.9897000193595886 ○ TSK000002_score 0.7840999960899353 0.795799970626831 ○ TSK000003_score 0.2980000078678131 0.9093999862670898 0.31189998984336853 ○ TSK000004_score 0.580299973487854 0.46950000524520874 0.95169997215271 0.5945000052452087 ○ TSK000005_score 0.29809999465942383 0.5546000003814697 0.4551999866962433 0.9332000017166138 0.31310001015663147 ○ TSK000006_score 0.3885999917984009 0.48890000581741333 0.1986999958753585 0.6723999977111816 0.40799999237060547 ○ TSK000007_score 0.6876000165939331 0.7035999894142151 ○ TSK000008_score 0.5837000012397766 0.26409998536109924 0.394899994134902 0.09929999709129333 0.916700005531311 0.5935999751091003 ○ TSK000009_score 0.5809999704360962 0.6718000173568726 0.5943999886512756 ○ TSK000010_score 0.5033000111579895 0.7215999960899353 0.5135999917984009

케라스 튜너

케라스 튜너

오토케라스

오토케라스

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2 predict 4 out 4 predict 5 out 5 predict 6 out 6

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2 arch 2 model 3 score 3

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2 arch 2 model 3 task 2 test 2 score 3

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2 arch 2 model 3 task 2 test 2 score 3 task 3 test 3 score 4

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2 arch 2 model 3 task 2 test 2 score 3 task 3 train 3 model 4 test 3 score 4

task 1 arch 1 train 1 model 1 test 1 score 1 predict 1 out 1 predict 2 out 2 predict 3 out 3 train 2 model 2 score 2 arch 2 model 3 task 2 test 2 score 3 task 3 arch 3 train 3 model 4 test 3 score 4

다양한 시나리오는 workflow에서~ traingen worker wo testgen worker wo predictgen worker wo train worker wo test worker wo predict worker wo train workflow test workflow predict workflow

정리 ● 추론 데이터 입력 시 ● 시험셋 갱신 시 ● 훈련셋 갱신 시 ● 새로운 아키텍처 추가 시 >> ● work order 만 추가하자! ● work order 관리 정책은 work flow로 정의하자