在keras中同时训练神经网络，并让它们在训练时共同分担损失？

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense

# 输入层
input_layer = Input(shape=(10,))

# 共享层
shared_layer = Dense(64, activation='relu')(input_layer)

# 子任务1的输出层
output1 = Dense(1, activation='sigmoid', name='task1')(shared_layer)

# 子任务2的输出层
output2 = Dense(1, activation='linear', name='task2')(shared_layer)

# 构建模型
model = Model(inputs=input_layer, outputs=[output1, output2])

# 编译模型，定义损失函数
model.compile(optimizer='adam',
              loss={'task1': 'binary_crossentropy', 'task2': 'mse'})

# 假设有一些训练数据
import numpy as np
X_train = np.random.random((100, 10))
y_train_task1 = np.random.randint(2, size=(100, 1))
y_train_task2 = np.random.random((100, 1))

# 训练模型
model.fit(X_train, {'task1': y_train_task1, 'task2': y_train_task2}, epochs=10)