from keras import Input, Model
from keras.layers import Dense, Concatenate
import numpy as np
from keras.utils import plot_model
from numpy import random as rd

samples_n = 3000
samples_dim_01 = 2
samples_dim_02 = 2
# 樣本數(shù)據(jù)
x1 = rd.rand(samples_n, samples_dim_01)
x2 = rd.rand(samples_n, samples_dim_02)
y_1 = []
y_2 = []
y_3 = []
for x11, x22 in zip(x1, x2):
  y_1.append(np.sum(x11) + np.sum(x22))
  y_2.append(np.max([np.max(x11), np.max(x22)]))
  y_3.append(np.min([np.min(x11), np.min(x22)]))
y_1 = np.array(y_1)
y_1 = np.expand_dims(y_1, axis=1)
y_2 = np.array(y_2)
y_2 = np.expand_dims(y_2, axis=1)
y_3 = np.array(y_3)
y_3 = np.expand_dims(y_3, axis=1)

# 輸入層
inputs_01 = Input((samples_dim_01,), name='input_1')
inputs_02 = Input((samples_dim_02,), name='input_2')
# 全連接層
dense_01 = Dense(units=3, name="dense_01", activation='softmax')(inputs_01)
dense_011 = Dense(units=3, name="dense_011", activation='softmax')(dense_01)
dense_02 = Dense(units=6, name="dense_02", activation='softmax')(inputs_02)
# 加入合并層
merge = Concatenate()([dense_011, dense_02])
# 分成兩類輸出 --- 輸出01
output_01 = Dense(units=6, activation="relu", name='output01')(merge)
output_011 = Dense(units=1, activation=None, name='output011')(output_01)
# 分成兩類輸出 --- 輸出02
output_02 = Dense(units=1, activation=None, name='output02')(merge)
# 分成兩類輸出 --- 輸出03
output_03 = Dense(units=1, activation=None, name='output03')(merge)
# 構(gòu)造一個新模型
model = Model(inputs=[inputs_01, inputs_02], outputs=[output_011,
                           output_02,
                           output_03
                           ])
# 顯示模型情況
plot_model(model, show_shapes=True)
print(model.summary())
# # 編譯
# model.compile(optimizer="adam", loss='mean_squared_error', loss_weights=[1,
#                                     0.8,
#                                     0.8
#                                     ])
# # 訓練
# model.fit([x1, x2], [y_1,
#           y_2,
#           y_3
#           ], epochs=50, batch_size=32, validation_split=0.1)

# 以下的方法可靈活設(shè)置
model.compile(optimizer='adam',
       loss={'output011': 'mean_squared_error',
          'output02': 'mean_squared_error',
          'output03': 'mean_squared_error'},
       loss_weights={'output011': 1,
              'output02': 0.8,
              'output03': 0.8})
model.fit({'input_1': x1,
      'input_2': x2},
     {'output011': y_1,
      'output02': y_2,
      'output03': y_3},
     epochs=50, batch_size=32, validation_split=0.1)

# 預測
test_x1 = rd.rand(1, 2)
test_x2 = rd.rand(1, 2)
test_y = model.predict(x=[test_x1, test_x2])
# 測試
print("測試結(jié)果：")
print("test_x1:", test_x1, "test_x2:", test_x2, "y:", test_y, np.sum(test_x1) + np.sum(test_x2))

# create model
model = Model(inputs=x_inp, outputs=[main_pred, aux_pred])
# complie model
model.compile(
  optimizer=optimizers.Adam(lr=learning_rate),
  loss={"main": weighted_binary_crossentropy(weights), "auxiliary":weighted_binary_crossentropy(weights)},
  loss_weights={"main": 0.5, "auxiliary": 0.5},
  metrics=[metrics.binary_accuracy],
)
# Train model
model.fit_generator(
  train_gen, epochs=num_epochs, verbose=0, shuffle=True
)

class Batch_generator(Sequence):
 """
 用于產(chǎn)生batch_1, batch_2(記住是numpy.array格式轉(zhuǎn)換)
 """
 y_batch = {'main':batch_1,'auxiliary':batch_2}
 return X_batch, y_batch

# or in another way
def batch_generator():
 """
 用于產(chǎn)生batch_1, batch_2(記住是numpy.array格式轉(zhuǎn)換)
 """
 yield X_batch, {'main': batch_1,'auxiliary':batch_2}