Question

使用tf.print，我们收到错误：

In [1]: import tensorflow as tf

In [2]: # using print

In [3]: entcoeff =  tf.Variable([0], dtype=tf.float32, trainable=False)
   ...: entcoeff = tf.Print(entcoeff,[entcoeff,"printing"])

In [4]: tf.assign(entcoeff, [-1.])
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-4-dd57efca5923> in <module>()
----> 1 tf.assign(entcoeff, [-1.])

/nohome/jaan/abhishek/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/state_ops.py in assign(ref, value, validate_shape, use_locking, name)
    270         ref, value, use_locking=use_locking, name=name,
    271         validate_shape=validate_shape)
--> 272   return ref.assign(value)

AttributeError: 'Tensor' object has no attribute 'assign'

不使用tf.print。它看起来像预期的那样正常

In [5]: # not using print

In [6]: entcoeff =  tf.Variable([0], dtype=tf.float32, trainable=False)

In [7]: tf.assign(entcoeff, [-1.])
Out[7]: <tf.Tensor 'Assign:0' shape=(1,) dtype=float32_ref>

tf.Print会将变量转换为常量吗？试图调试：

In [8]: entcoeff = tf.Print(entcoeff,[entcoeff,"printing"])

In [9]: type(entcoeff)
Out[9]: tensorflow.python.framework.ops.Tensor
In [10]: dir(entcoeff)
Out[10]:
['OVERLOADABLE_OPERATORS',
 '__abs__',
 '__add__',
 '__and__',
 '__array_priority__',
 '__bool__',
 '__class__',
 '__delattr__',
 '__dict__',
 '__dir__',
 '__div__',
 '__doc__',
 '__eq__',
 '__floordiv__',
 '__format__',
 '__ge__',
 '__getattribute__',
 '__getitem__',
 '__gt__',
 '__hash__',
 '__init__',
 '__init_subclass__',
 '__invert__',
 '__iter__',
 '__le__',
 '__lt__',
 '__matmul__',
 '__mod__',
 '__module__',
 '__mul__',
 '__ne__',
 '__neg__',
 '__new__',
 '__nonzero__',
 '__or__',
 '__pow__',
 '__radd__',
 '__rand__',
 '__rdiv__',
 '__reduce__',
 '__reduce_ex__',
 '__repr__',
 '__rfloordiv__',
 '__rmatmul__',
 '__rmod__',
 '__rmul__',
 '__ror__',
 '__rpow__',
 '__rsub__',
 '__rtruediv__',
 '__rxor__',
 '__setattr__',
 '__sizeof__',
 '__str__',
 '__sub__',
 '__subclasshook__',
 '__truediv__',
 '__weakref__',
 '__xor__',
 '_add_consumer',
 '_as_node_def_input',
 '_as_tf_output',
 '_consumers',
 '_dtype',
 '_handle_dtype',
 '_handle_shape',
 '_op',
 '_override_operator',
 '_shape',
 '_shape_as_list',
 '_value_index',
 'consumers',
 'device',
 'dtype',
 'eval',
 'get_shape',
 'graph',
 'name',
 'op',
 'set_shape',
 'shape',
 'value_index']

Answer 1

问题是tf.Print会返回张量而不是tf.Variable。您可以通过将结果分配给一个新变量来解决这个问题，虽然这对我来说有点麻烦，并且可能有更优雅的方式。如果您运行以下内容：

import tensorflow as tf

entcoeff = tf.Variable([0], dtype=tf.float32, trainable=False)
print(entcoeff)
entcoeff = tf.Variable(tf.Print(entcoeff, [entcoeff], message="\n\nprinting"))
print(entcoeff)

entcoeff = tf.assign(entcoeff, [-1.])   # This returns an op which you need to run for the assignment to happen
print(entcoeff)
init_op = tf.global_variables_initializer()

sess = tf.Session()
sess.run(init_op)           # This triggers the tf.Print
print(sess.run(entcoeff))   # This causes entcoeff to be assigned -1.

你应该得到

<tf.Variable 'Variable:0' shape=(1,) dtype=float32_ref>
<tf.Variable 'Variable_1:0' shape=(1,) dtype=float32_ref>
<tf.Variable 'Variable_1:0' shape=(1,) dtype=float32_ref>
[ -1.]
2017-11-23 00:49:56.450723: I C:\tf_jenkins\home\workspace\rel-win\M\windows\PY\35\tensorflow\core\kernels\logging_ops.cc:79]

printing[0]

请注意，tf.Print的输出不会显示到结尾，因为tf.Print通过写入标准错误来工作。另请注意我添加到输出消息中的\n\n以使其更加明显。

tf.Print与tf.assign并不是很好

1 个答案: