Kerasで複数出力する回帰モデルを構築する方法【Functional API】

# 使用するライブラリをインポート
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

import io
from google.colab import files

from sklearn.model_selection import train_test_split
from keras.layers import Dense, Activation
from keras.optimizers import Adam
from keras.layers.core import Dropout
from keras.callbacks import EarlyStopping, LearningRateScheduler
from keras.models import Model
from keras.layers import Input, Dense

from sklearn.metrics import mean_absolute_error
from sklearn.metrics import mean_squared_error
from math import sqrt
from sklearn.metrics import r2_score


# Google Colaboratoryでファイルをアップロード
uploaded = files.upload()


# データフレームとして、変数dfに4列×10000行のデータを格納
df = pd.read_csv(io.BytesIO(uploaded['Multiple-Outputs.csv']))


# データの分割
(train, test) = train_test_split(df, test_size=0.2, shuffle=True)

# x_train, x_testではcsvファイルの1列目から2列目を，y_train, y_testでは3, 4列目のデータを活用
x_train = train.iloc[:, [0, 1]]
y_train = train.iloc[:, [2,3]]

x_test = test.iloc[:, [0, 1]]
y_test = test.iloc[:, [2,3]]


# データの正規化
x_train = (x_train - x_train.min()) / (x_train.max() - x_train.min())
x_test = (x_test - x_test.min()) / (x_test.max() - x_test.min())


# モデルの構築
# inputの数:c2, s
inputs = Input(shape=(2,))

# 学習回数
epochs = 100
batch_size = 64

# ノード数
node = 64

x1 = Dense(node, activation='relu')(inputs)
x2 = Dense(node, activation='relu')(x1)
x3 = Dense(node, activation='relu')(x2)


x4 = Dense(node, activation='relu')(x3) # アウトプット直前の層
output1 = Dense(1, name='output1')(x4)

x5 = Dense(node, activation='relu')(x3) # アウトプット直前の層
output2 = Dense(1, name='output2')(x5)

model = Model(inputs=inputs, outputs=[output1, output2])

optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999)
model.compile(
    optimizer=optimizer,
    loss=['mse', 'mse']
)


# 学習オプション
history = model.fit(x_train, 
                    {"output1":y_train.iloc[:, [0]], "output2":y_train.iloc[:, [1]]}, 
                    epochs=epochs, 
                    batch_size=batch_size, 
                    validation_data=(x_test, {"output1":y_test.iloc[:, [0]], "output2":y_test.iloc[:, [1]]})
)


# 構築したモデルで予測
pred1, pred2 = model.predict(x_test)


#精度評価指標
# Output1 ====================================================================
print("=======  Output1  =======")
# MAE
test_MAE = mean_absolute_error(y_test.iloc[:, 0], pred1)
print('MAE：', test_MAE)

# RMSE
test_RMSE = sqrt(mean_squared_error(y_test.iloc[:, 0], pred1))
print('RMSE：', test_RMSE)

# 決定係数
R2 = r2_score(y_test.iloc[:, 0], pred1)
print('R2：', R2)

print("")

# Output2 ====================================================================
print("=======  Output2  =======")
# MAE
test_MAE = mean_absolute_error(y_test.iloc[:, 1], pred2)
print('MAE：', test_MAE)

# RMSE
test_RMSE = sqrt(mean_squared_error(y_test.iloc[:, 1], pred2))
print('RMSE：', test_RMSE)

# 決定係数
R2 = r2_score(y_test.iloc[:, 1], pred2)
print('R2：', R2)



# Figureを追加
fig = plt.figure(figsize = (15, 5))

# Output1 ====================================================================
# 1×2 の1枚目に描画する
loss_graph1 = fig.add_subplot(1, 2, 1)

# x,yに要素追加
loss1 = history.history['output1_loss']
val_loss1 = history.history['val_output1_loss']
epochs1 = len(loss1)

# グラフ設定
loss_graph1.plot(range(epochs), loss1, marker='.', label='loss')
loss_graph1.plot(range(epochs), val_loss1, marker='.', label='val_loss')
loss_graph1.set_title("Loss_Output1", size = 18)
loss_graph1.legend(loc='best')
loss_graph1.grid()
loss_graph1.set_xlabel("epochs", size = 14, color = "r")
loss_graph1.set_ylabel("loss", size = 14, color = "r")

#==============================================================================
# Output2 ====================================================================
# 1×2 の2枚目に描画する
loss_graph2 = fig.add_subplot(1, 2, 2)

# x,yに要素追加
loss2 = history.history['output2_loss']
val_loss2 = history.history['val_output2_loss']
epochs2 = len(loss2)

# グラフ設定
loss_graph2.plot(range(epochs), loss2, marker='.', label='loss')
loss_graph2.plot(range(epochs), val_loss2, marker='.', label='val_loss')
loss_graph2.set_title("Loss_Output2", size = 18)
loss_graph2.legend(loc='best')
loss_graph2.grid()
loss_graph2.set_xlabel("epochs", size = 14, color = "r")
loss_graph2.set_ylabel("loss", size = 14, color = "r")
#==============================================================================



# Observed-Predicted Plot
# Figureを追加
fig = plt.figure(figsize = (15, 5))

# Output1 ====================================================================
# 1×2 の1枚目に描画する
yy1 = fig.add_subplot(1, 2, 1)

yy1.scatter(y_test.iloc[:, 0], pred1, c='blue', alpha=0.8)

# グラフ設定
yy1.set_title("observed_predicted1", size = 18)
yy1.grid()
yy1.set_xlabel("observed_data", size = 14, color = "r")
yy1.set_ylabel("predicted_data", size = 14, color = "r")

# Output2 ====================================================================
# 1×2 の2枚目に描画する
yy2 = fig.add_subplot(1, 2, 2)

yy2.scatter(y_test.iloc[:, 1], pred2, c='blue', alpha=0.8)

# グラフ設定
yy2.set_title("observed_predicted2", size = 18)
yy2.grid()
yy2.set_xlabel("observed_data", size = 14, color = "r")
yy2.set_ylabel("predicted_data", size = 14, color = "r")