自定义损失函数，包括层的输出

Question

我有一个数据集，每个案例包含 1 个 y_true 值。我想构建一个DNN，它输出3个系数，稍后将用于创建y_pred

y_pred = 4*coeff_1 + 5*coeff_2 + 6 *coeff_3

我正在使用keras，当我尝试定义这样的自定义函数时

from keras.callbacks import ModelCheckpoint
from keras.layers import advanced_activations
from keras.models import Sequential
from keras.layers import Dense, Activation, Flatten
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_absolute_error 

import keras.backend as K

def custom_objective(layer):
    return K.sum(layer.output)

NN_model = Sequential()

# The Input Layer :
NN_model.add(Dense(X_train.shape[1], kernel_initializer='normal',input_dim = X_train.shape[1], activation='relu'))

# The Hidden Layers :
NN_model.add(Dense(20, kernel_initializer='normal',activation='elu'))
NN_model.add(Dense(20, kernel_initializer='normal',activation='elu'))

output_layer = Dense(1, kernel_initializer='normal',activation='linear')

# The Output Layer :
NN_model.add(output_layer)

# Compile the network :
NN_model.compile(loss=custom_objective(output_layer), optimizer='Adamax', metrics=['mean_absolute_error'])
NN_model.summary()

NN_model.fit(X_train, y_train, epochs=10,verbose = 1)

print('NN train = ', mean_absolute_error(y_train , NN_model.predict(X_train)))    

predictions = NN_model.predict(X_test)

MAE = mean_absolute_error(y_test , predictions)

print('NN MAE = ', MAE)

我都明白了

TypeError：不允许将tf.Tensor 用作Python bool。利用 if t is not None: 而不是 if t: 来测试是否定义了张量， 并使用 tf.cond 等 TensorFlow 操作来执行子图 以张量的值为条件。

所以我的问题是

我如何定义一个DNN，每个数据需要1个y_true，输出3个值，它将线性组合成一个y_pred，用于获取损失函数和训练网络

感谢您的宝贵时间

Answer 1

按照这些思路怎么样？

from keras.models import Model
from keras.layers import Dense, Input, Add, Lambda


def model(inp_size):
    inp = Input(shape=(inp_size, 1))

    x1 = Dense(20, activation='elu')(inp)
    x1 = Dense(20, activation='elu')(x1)
    x1 = Dense(1, activation = 'linear')(x1)

    x2 = Dense(20, activation='elu')(inp)
    x2 = Dense(20, activation='elu')(x2)
    x2 = Dense(1, activation = 'linear')(x2)

    x3 = Dense(20, activation='elu')(inp)
    x3 = Dense(20, activation='elu')(x3)
    x3 = Dense(1, activation = 'linear')(x3)

    x1 = Lambda(lambda x: x * 4.0)(x1)
    x2 = Lambda(lambda x: x * 5.0)(x2)
    x3 = Lambda(lambda x: x * 6.0)(x3)
    out = Add()([x1, x2, x3])

    return Model(inputs = inp, outputs = out)