如何正确设置tensorflow中占位符的值

Question

我用tensorflow编写了以下快速程序来打印斐波那契数。初始fib编号被初始化为占位符x1和x2。然而，当我试图输入session.run中占位符的值时，会导致错误：

InvalidArgumentError: You must feed a value for placeholder tensor 'x2' with dtype int64 and shape [1]

你能帮我理解并解决我的代码问题吗？

import tensorflow as tf
session = tf.InteractiveSession()
import numpy as np
ones = tf.ones((1,))
n = 10
x1 = tf.placeholder(tf.int64, [1], name='x1')
x2 = tf.placeholder(tf.int64, [1], name='x2')
temp = tf.Variable(ones, name='temp')
tf.initialize_all_variables().run()
fib_nums = [x1, x2]
for i in range(100):
temp = x1 + x2
x1 = x2
x2 = temp
fib_nums.append(temp)
series = tf.stack(fib_nums)
print(np.ones(1).astype(np.int64))
session.run(series, feed_dict={x1:np.ones(1).astype(np.int64), x2:np.ones(1).astype(np.int64)})
print(series.eval())

Answer 1

这里有几个错误。首先，由于您重用Python名称x1和x2，因此当您在feed_dict中给定它们时，它们不再引用占位符，而是引用循环的最后结果。因此，您应该更改代码，使您在feed_dict中给出的键真正成为占位符。其次，首先用feed_dict调用session.run，这是正确的，但随后调用series.eval()，这与前一行基本相同，只是在这种情况下没有提供feed_dict，所以它不起作用。您实际上不需要调用series.eval()，只需获取session.run返回的值即可。你的固定程序可能看起来像这样：

import tensorflow as tf
session = tf.InteractiveSession()
import numpy as np
ones = tf.ones((1,))
n = 10
# Reserve these Python names for the placeholders
x1_ph = tf.placeholder(tf.int64, [1], name='x1')
x2_ph = tf.placeholder(tf.int64, [1], name='x2')
temp = tf.Variable(ones, name='temp')
tf.initialize_all_variables().run()
# Use placeholders as initial values in the iterations
x1, x2 = x1_ph, x2_ph
fib_nums = [x1, x2]
for i in range(100):
temp = x1 + x2
x1 = x2
x2 = temp
fib_nums.append(temp)
series = tf.stack(fib_nums)
print(np.ones(1).astype(np.int64))
# You can just give lists as inputs and TensorFlow will convert their type
series_val = sess.run(series, feed_dict={x1_ph: [1], x2_ph: [1]})
print(series_val)

输出：

[1]
[[                   1]
[                   1]
[                   2]
[                   3]
[                   5]
[                   8]
[                  13]
[                  21]
[                  34]
[                  55]
[                  89]
[                 144]
[                 233]
[                 377]
[                 610]
[                 987]
[                1597]
[                2584]
...

Answer 2

我认为另一个变体是使用tf.identity.

y = tf.identity(x1)
y1 = tf.identity(x2)
fib_nums = [y, y1]
for i in range(100):
temp = y + y1
y = y1
y1 = temp
fib_nums.append(temp)

这个占位符修改问题也在这里讨论

这是获得该系列的又一种方式。

def cond(i, x_next, x_prev):
return x_next <= 100
def body( i, x_next, x_prev ):
nextinseries = x_next + x_prev
next = tf.Print(x_next, [x_next], message="Next is : ")
prev = tf.Print(x_prev, [x_prev], message="Previous is : ")
with tf.control_dependencies([nextinseries]):
prev = tf.identity( next )

return [i + 1, nextinseries , prev ]
sess = tf.Session()
sess.run(tf.global_variables_initializer())

sess.run(tf.while_loop(cond, body, [1, 1, 1]))

Answer 3

只需将tf.placeholder更改为张量，它就可以使用

x1 = tf.ones(dtype=tf.int64, shape=1, name='x1') 
x2 = tf.ones(dtype=tf.int64, shape=1, name='x2')