请问自定义损失函数的输入如果是多维的应该怎么写呢？

来源：3-5 实战自定义损失函数与DenseLayer回顾

BloodHound_swh

2019-06-16

def myloss(matrix_true, matrix_pre):
    loss = tf.reduce_sum(matrix_true - matrix_pre)
    return loss

例如上面这一段，matrix_true和matrix_true都是[100, 3]的矩阵，但是在model.fit的时候

history = model.fit(x_train, y_train, epochs=30)

报错

InvalidArgumentError: Index out of range using input dim 2; input has only 2 dims [Op:StridedSlice] name: strided_slice/

另外，自定义损失函数里面的参数顺序是怎样的呢？

为了方便起见，将矩阵改成3*3的，下面是完整的代码

import tensorflow as tf
import numpy as np
from tensorflow import keras
import copy

print(tf.__version__)

x = tf.Variable([[3., 0., 5.], [0., 5., 4.], [3., 3., 0.]])
y = tf.Variable([[3., 0., 6.], [0., 4., 4.], [3., 2., 0.]])


def myloss(matrix_true, matrix_pre):
    loss = tf.reduce_sum(matrix_true - matrix_pre)
    return loss

model = keras.models.Sequential([
    keras.layers.Dense(3, activation='relu', input_shape=x.shape[1:]),
    keras.layers.Dense(3, activation='relu'),
    keras.layers.Dense(3, activation='relu'),
])

op = keras.optimizers.SGD(0.01)

model.summary()
model.compile(loss=myloss, optimizer=op)

history = model.fit(x, y, epochs=10)

不知道是不是我定义的变量的问题，即便是改成预测值是一维的，也无法成功运行

#%%
import tensorflow as tf
import numpy as np
from tensorflow import keras

print(tf.__version__)


#%%
x = tf.Variable([[3., 0., 5.], [0., 5., 4.], [3., 3., 0.]])
# y = tf.Variable([[3., 0., 6.], [0., 4., 4.], [3., 2., 0.]])
z = tf.Variable([[2.], [3.], [4.]])


def myloss(y_true, y_pred):
    loss = tf.reduce_sum(y_true - y_pred)
    return loss

model = keras.models.Sequential([
    keras.layers.Dense(3, activation='relu', input_shape=x.shape[1:]),
    keras.layers.Dense(3, activation='relu'),
    keras.layers.Dense(1),
])

op = keras.optimizers.SGD(0.01)

model.summary()
model.compile(loss=myloss, optimizer=op)

#%%
history = model.fit(x, z, epochs=10)

写回答

2回答