tensorflow - 使用tf.metrics.mean_absolute_error时，获取 'AttributeError: '元组'对象没有属性 'dtype''

Question

我想用一个隐藏层训练一个非常简单的网络，但是我似乎无法训练该网络。我一直在标题中得到错误。但是，当我将损失定义为y - a2时，没有问题(除非结果是所有Nan，而不是我期望的结果)。我想念什么？

import tensorflow as tf
import numpy as np

# import data
X = np.array([[0,0,1], #XOR prob
              [0,1,1],
              [1,0,1],
              [1,1,1],])


# output dataset, same as before
y = np.array([[0,1,1,0]]).T


# ----------------design network architecture
# define variables

X = tf.convert_to_tensor(X, dtype=tf.float32) # convert np X to a tensor
y = tf.convert_to_tensor(y, dtype=tf.float32) # convert np y to a tensor
W1 = tf.Variable(tf.random_normal([3, 4]))
W2 = tf.Variable(tf.random_normal([4, 1]))
a1 = tf.matmul(X, W1)
a2 = tf.matmul(a1, W2)

# define operations

# ---------------define loss and select training algorithm
loss = tf.metrics.mean_absolute_error(labels=y, predictions=a2)
#loss = y - a2
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss)

# ----------------run graph to train and get result
with tf.Session() as sess:

    #initialize variables
    sess.run(tf.initialize_all_variables())

    for i in range(60000):
        sess.run(train)
        if i % 10000 == 0:
            print("Loss: ", sess.run(loss))

    print("Activation: ", sess.run(a2))
    print("Loss: ", sess.run(loss))

Answer 1

@伊曼纽尔·阿科萨(Emmanuel Akosah)

这是运行的代码。我刚刚修改了损失函数以及global_variables_initializer()。

import tensorflow as tf
import numpy as np

# import data
X = np.array([[0,0,1], #XOR prob
              [0,1,1],
              [1,0,1],
              [1,1,1],])


# output dataset, same as before
y = np.array([[0,1,1,0]]).T


# ----------------design network architecture
# define variables

X = tf.convert_to_tensor(X, dtype=tf.float32) # convert np X to a tensor
y = tf.convert_to_tensor(y, dtype=tf.float32) # convert np y to a tensor
W1 = tf.Variable(tf.random_normal([3, 4]))
W2 = tf.Variable(tf.random_normal([4, 1]))
a1 = tf.matmul(X, W1)
a2 = tf.matmul(a1, W2)

# define operations

# ---------------define loss and select training algorithm
#loss = tf.metrics.mean_absolute_error(labels=y, predictions=a2) #commented this line
loss=tf.keras.losses.mean_absolute_error(y_true=y, y_pred=a2) #added this line
#loss = y - a2
optimizer = tf.train.GradientDescentOptimizer(0.1)
train = optimizer.minimize(loss)

# ----------------run graph to train and get result
with tf.Session() as sess:

    #initialize variables
    #sess.run(tf.initialize_all_variables()) #commented this line
    sess.run(tf.global_variables_initializer()) #added this line

    for i in range(60000):
        sess.run(train)
        if i % 10000 == 0:
            print("Loss: ", sess.run(loss))

    print("Activation: ", sess.run(a2))
    print("Loss: ", sess.run(loss))

输出如下

Loss:  [0.04997864 1.2521797  1.6842678  0.8864688 ]
Loss:  [0.44828027 0.92680335 0.34281957 0.2820967 ]
Loss:  [0.44828027 0.92680335 0.34281957 0.2820967 ]
Loss:  [0.44828027 0.92680335 0.34281957 0.2820967 ]
Loss:  [0.44828027 0.92680335 0.34281957 0.2820967 ]
Loss:  [0.44828027 0.92680335 0.34281957 0.2820967 ]
Activation:  [[0.44828027]
 [0.07319663]
 [0.6571804 ]
 [0.2820967 ]]
Loss:  [0.44828027 0.92680335 0.34281957 0.2820967 ]

我认为最好遵循tensorflow网站上的 this regression tutorial更新代码。您可以使用其他损失函数和优化器，以获得更好的结果。

如果您认为此答案有用，请接受此答案和/或对其进行投票。谢谢!