掛け算のニューラルネットワークとベイズ推定

#ライブラリのimport
import numpy as np
from keras.layers import Input, Dense,Add
from keras.models import Model
from keras.callbacks import EarlyStopping

'''
データの生成
'''
N = 10000
N_train = int(N * 0.9)
N_validation = N - N_train

#1~100の数を設定
X=np.random.randint(1,100,N*2)
#配列の次元を変更
X=np.reshape(X,(N,2))
#正解データを作る
Y=X[:,0]*X[:,1]

#train, validationを分割し、正規化
X_train=X[:N_train]/100
Y_train=Y[:N_train]/10000
X_validation=X[N_train:]/100
Y_validation=Y[N_train:]/10000

#対数変換した訓練データ兼正解データ
lnX_train=np.log(X_train)

def log_nn_model():
    log_inputs = Input(shape=(1,))
    log_x = Dense(4, activation='relu')(log_inputs)
    log_x = Dense(4, activation='relu')(log_x)
    log_x = Dense(4, activation='relu')(log_x)
    log_outputs = Dense(1, activation='linear')(log_x)
    model = Model(inputs=log_inputs,outputs=log_outputs)
    
    model.compile(loss='mean_squared_error',
                  optimizer='adam')
    model.summary()
    return model

def exp_nn_model():
    lnx1_input=Input(shape=(1,))
    lnx2_input=Input(shape=(1,))
    
    #2つの入力を足し合わせる
    added = Add()([lnx1_input, lnx2_input])

    exp_x = Dense(4, activation='relu')(added)
    exp_x = Dense(4, activation='relu')(exp_x)
    exp_x = Dense(4, activation='relu')(exp_x)
    output = Dense(1, activation='relu')(exp_x)

    model = Model(inputs=[lnx1_input,lnx2_input],outputs=output)
    
    model.compile(loss='mean_squared_error',
                  optimizer='adam')
    model.summary()
    return model

def train():
    early_stopping = EarlyStopping(monitor='val_loss', patience=10, verbose=1)
    epochs = 500
    batch_size = 128
    #対数変換の学習
    log_model=log_nn_model()
    #log_model.load_weights('log_nn_model_weights.h5')
    log_model.fit(X_train[:,0], lnX_train[:,0],
                  batch_size=batch_size, 
                  epochs=epochs,
                  validation_split=0.2,
                  shuffle=True,
                  callbacks=[early_stopping])
    log_model.save_weights('log_nn_model_weights.h5')
    print('save weights')
    
    #指数変換の学習
    exp_model=exp_nn_model()
    #exp_model.load_weights('exp_nn_model_weights.h5')
    exp_model.fit([lnX_train[:,0], lnX_train[:,1]],
                  Y_train,
                  batch_size=batch_size, 
                  epochs=epochs,
                  validation_split=0.2,
                  shuffle=True,
                  callbacks=[early_stopping])
    exp_model.save_weights('exp_nn_model_weights.h5')
    print('save weights')

    # 検証データからランダムに問題を選んで答え合わせ
    for i in range(20):
        index = np.random.randint(N_validation)
        question = X_validation[index]
        answer = Y_validation[index]
        predict_log_x1 = log_model.predict([question[0]])
        predict_log_x2 = log_model.predict([question[1]])
        prediction = exp_model.predict([predict_log_x1 ,predict_log_x2])

        print('-' * 10)
        print('Q:  '+str(question[0]*100)+'x'+str(question[1]*100))
        print('A:  '+str(answer*10000))
        print('P:  '+str(prediction[0][0]*10000))
    print('-' * 10)