Tensorflow：TypeError：Fetch参数None的类型无效<class'NoneType'>

Question

我正在尝试运行这个简单的程序来计算梯度，但是我得到了 None 错误：

import tensorflow as tf
import numpy as np

batch_size = 5
dim = 3
hidden_units = 8


sess = tf.Session()

with sess.as_default():
    x = tf.placeholder(dtype=tf.float32, shape=[None, dim], name="x")
    y = tf.placeholder(dtype=tf.int32, shape=[None], name="y")
    w = tf.Variable(initial_value=tf.random_normal(shape=[dim, hidden_units]), name="w")
    b = tf.Variable(initial_value=tf.zeros(shape=[hidden_units]), name="b")
    logits = tf.nn.tanh(tf.matmul(x, w) + b)

    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, y,name="xentropy")
    # define model end


    # begin training
    optimizer = tf.train.GradientDescentOptimizer(1e-5)
    grads_and_vars = optimizer.compute_gradients(cross_entropy, tf.trainable_variables())

    # generate data
    data = np.random.randn(batch_size, dim)
    labels = np.random.randint(0, 10, size=batch_size)

    sess.run(tf.initialize_all_variables())
    gradients_and_vars = sess.run(grads_and_vars, feed_dict={x:data, y:labels})
    for g, v in gradients_and_vars:
        if g is not None:
            print "****************this is variable*************"
            print "variable's shape:", v.shape
            print v
            print "****************this is gradient*************"
            print "gradient's shape:", g.shape
            print g

sess.close()

错误：

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-14-8096b2e21e06> in <module>()
     29 
     30     sess.run(tf.initialize_all_variables())
---> 31     outnet = sess.run(grads_and_vars, feed_dict={x:data, y:labels})
     32 #     print(gradients_and_vars)
     33 #         if g is not None:

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
    893     try:
    894       result = self._run(None, fetches, feed_dict, options_ptr,
--> 895                          run_metadata_ptr)
    896       if run_metadata:
    897         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
   1107     # Create a fetch handler to take care of the structure of fetches.
   1108     fetch_handler = _FetchHandler(
-> 1109         self._graph, fetches, feed_dict_tensor, feed_handles=feed_handles)
   1110 
   1111     # Run request and get response.

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in __init__(self, graph, fetches, feeds, feed_handles)
    411     """
    412     with graph.as_default():
--> 413       self._fetch_mapper = _FetchMapper.for_fetch(fetches)
    414     self._fetches = []
    415     self._targets = []

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in for_fetch(fetch)
    231     elif isinstance(fetch, (list, tuple)):
    232       # NOTE(touts): This is also the code path for namedtuples.
--> 233       return _ListFetchMapper(fetch)
    234     elif isinstance(fetch, dict):
    235       return _DictFetchMapper(fetch)

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in __init__(self, fetches)
    338     """
    339     self._fetch_type = type(fetches)
--> 340     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
    341     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)
    342 

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in <listcomp>(.0)
    338     """
    339     self._fetch_type = type(fetches)
--> 340     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
    341     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)
    342 

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in for_fetch(fetch)
    231     elif isinstance(fetch, (list, tuple)):
    232       # NOTE(touts): This is also the code path for namedtuples.
--> 233       return _ListFetchMapper(fetch)
    234     elif isinstance(fetch, dict):
    235       return _DictFetchMapper(fetch)

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in __init__(self, fetches)
    338     """
    339     self._fetch_type = type(fetches)
--> 340     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
    341     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)
    342 

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in <listcomp>(.0)
    338     """
    339     self._fetch_type = type(fetches)
--> 340     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]
    341     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)
    342 

//anaconda/lib/python3.5/site-packages/tensorflow/python/client/session.py in for_fetch(fetch)
    228     if fetch is None:
    229       raise TypeError('Fetch argument %r has invalid type %r' %
--> 230                       (fetch, type(fetch)))
    231     elif isinstance(fetch, (list, tuple)):
    232       # NOTE(touts): This is also the code path for namedtuples.

TypeError: Fetch argument None has invalid type <class 'NoneType'>

为什么会出现这个错误？版本问题？

Answer 1

Gradients 返回None 如果图中它们之间没有显式连接。在您的代码中，似乎所有声明的变量都有连接，因此可能是从其他图表加载变量的情况。您可以使用：

 print([v.name for v in tf.all_variables()])

并仅检查预期变量是否是此图表的一部分。

试试这样的：

sess.run(tf.initialize_all_variables())
gradients_and_vars = sess.run([variable for grad,variable in grads_and_vars], feed_dict={x:data, y:labels})
print(gradients_and_vars)