[PYTHON/KERAS] 커스텀 히스토리 콜백 함수 사용하기

■ 커스텀 히스토리 콜백 함수를 사용하는 방법을 보여준다.

 

▶ 예제 코드 (PY)

import keras
import keras.callbacks as callbacks
import keras.datasets.mnist as mnist
import keras.models as models
import keras.utils as utils
import keras.layers as layers
import matplotlib.pyplot as pp
import numpy as np

np.random.seed(3)

# 사용자 정의 히스토리 클래스를 정의한다.
class CustomHistory(callbacks.Callback):
    def init(self):
        self.trainLossList = []
        self.validationLossList = []
        self.trainAccuracyList = []
        self.validationAccuracyList = []

    def on_epoch_end(self, batch, logTuple = {}):
        self.trainLossList.append(logTuple.get("loss"))
        self.validationLossList.append(logTuple.get("val_loss"))
        self.trainAccuracyList.append(logTuple.get("acc"))
        self.validationAccuracyList.append(logTuple.get("val_acc"))

print("데이터 로드를 시작합니다.")

(trainInputNDArray, trainCottectOutputNDArray), (testInputNDArray, testCorrectOutputNDArray) = mnist.load_data()

# trainInputNDArray         : (60000, 28, 28)
# trainCottectOutputNDArray : (60000,)
# testInputNDArray          : (10000, 28, 28)
# testCorrectOutputNDArray  : (10000,)

# 훈련/검증 데이터를 분리한다.
validationInputNDArray         = trainInputNDArray[50000:]
validationCorrectOutputNDArray = trainCottectOutputNDArray[50000:]
trainInputNDArray              = trainInputNDArray[:50000]
trainCottectOutputNDArray      = trainCottectOutputNDArray[:50000]

# 훈련/검증/테스트 데이터
trainInputNDArray      = trainInputNDArray.reshape(50000, 784).astype("float32") / 255.0
validationInputNDArray = validationInputNDArray.reshape(10000, 784).astype("float32") / 255.0
testInputNDArray       = testInputNDArray.reshape(10000, 784).astype("float32") / 255.0

# trainInputNDArray      : (50000, 784)
# validationInputNDArray : (10000, 784)
# testInputNDArray       : (10000, 784)

# 훈련/검증 데이터를 섞는다.
#trainRandomIndexNDArray      = np.random.choice(50000, 700)
#validationRandomIndexNDArray = np.random.choice(10000, 300)
#
#trainInputNDArray              = trainInputNDArray[trainRandomIndexNDArray]
#trainCottectOutputNDArray      = trainCottectOutputNDArray[trainRandomIndexNDArray]
#validationInputNDArray         = validationInputNDArray[validationRandomIndexNDArray]
#validationCorrectOutputNDArray = validationCorrectOutputNDArray[validationRandomIndexNDArray]
#
# trainInputNDArray              : (700, 784)
# trainCottectOutputNDArray      : (700,)
# validationInputNDArray         : (300, 784)
# validationCorrectOutputNDArray : (300,)

# 정답 데이터에 대해 원핫 인코딩 처리한다.
trainCottectOutputNDArray      = utils.np_utils.to_categorical(trainCottectOutputNDArray)
validationCorrectOutputNDArray = utils.np_utils.to_categorical(validationCorrectOutputNDArray)
testCorrectOutputNDArray       = utils.np_utils.to_categorical(testCorrectOutputNDArray)

# trainCottectOutputNDArray      : (50000, 10)
# validationCorrectOutputNDArray : (10000, 10)
# testCorrectOutputNDArray       : (10000, 10)

print("데이터 로드를 종료합니다.")

print("모델 정의를 시작합니다.")

model = models.Sequential()

model.add(layers.Dense(units = 64, input_dim = 784, activation = "relu"))
model.add(layers.Dense(units = 10, activation = "softmax"))

model.compile(loss = "categorical_crossentropy", optimizer = "sgd", metrics = ["accuracy"])

print("모델 정의를 종료합니다.")

print("모델 학습을 시작합니다.")

customHistory = CustomHistory()

customHistory.init()

model.fit(trainInputNDArray, trainCottectOutputNDArray, epochs = 1000, batch_size = 100,\
    validation_data = (validationInputNDArray, validationCorrectOutputNDArray), callbacks = [customHistory])

print("모델 학습을 종료합니다.")

print("학습 결과를 조회합니다.")

figure, lossAxeSubplot = pp.subplots()

accuracyAxeSubplot = lossAxeSubplot.twinx()

lossAxeSubplot.plot(customHistory.trainLossList     , "y", label = "train loss")
lossAxeSubplot.plot(customHistory.validationLossList, "r", label = "val loss"  )

accuracyAxeSubplot.plot(customHistory.trainAccuracyList     , "b", label = "train acc")
accuracyAxeSubplot.plot(customHistory.validationAccuracyList, "g", label = "val acc"  )

lossAxeSubplot.set_xlabel("epoch")
lossAxeSubplot.set_ylabel("loss")

lossAxeSubplot.legend(loc = "upper left")

accuracyAxeSubplot.set_ylabel("accuracy")

accuracyAxeSubplot.legend(loc = "lower left")

pp.show()