在Tensorflow中只计算一次子图

Question

我正在用Tensorflow构建一个深度学习模型。在训练之前，我会做一些计算，如反向传播。但只需要计算一次。下面是我的伪代码：

class residual_net() 
    def pseudo_bp(self): 
     # do something... 
     self.bp = ... 

    def build_net(self): 
     # build a residual_network.... 
     # utilize the variable in pseudo_bp 
     rn.output = func(self.bp) 

def run(): 
    rn = residual_net() 
    rn.pseudo_bp() 
    rn.deep_residual_network() 
    sess = tf.InteractiveSession() 
    sess.run(tf.initialize_all_variables()) 
    for i in range(1000): 
     err = tf.reduce_mean(rn.output, labels) 
     train = tf.train.GradientDescentOptimizer(learning_rate).minimize(err) 
     sess.run(train, feed_dict=train_feed_dict) 

不知pseudo_bp将在每次迭代运行？如果是的话，我怎么才能让它运行一次？提前致谢！

编辑： 最新的错误：

Traceback (most recent call last): 
    File "run.py", line 124, in <module> 
    sess.run(pseudo_bp, feed_dict=feed_dict) 
    File "/Users/yobichi/bigdata/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 717, in run 
    run_metadata_ptr) 
    File "/Users/yobichi/bigdata/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 902, in _run 
    fetch_handler = _FetchHandler(self._graph, fetches, feed_dict_string) 
    File "/Users/yobichi/bigdata/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 358, in __init__ 
    self._fetch_mapper = _FetchMapper.for_fetch(fetches) 
    File "/Users/yobichi/bigdata/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 178, in for_fetch 
    (fetch, type(fetch))) 
TypeError: Fetch argument None has invalid type <class 'NoneType'> 

你有什么想法？

Answer 1

在TensorFlow中，您从构建tf.Graph开始。该图由变量，操作和占位符组成。然后开始tf.Session()，您可以在其中执行操作并更新变量。

在这种情况下，我认为psuedo_bp最终需要您计算一些操作（如tf.matmul）。 sess就像一个指针，只要你运行sess.run(op)就会执行各种tf.Operation。您提供一些输入来填充占位符（feed_dict）。

因此，您只会执行sess.run(op) for for循环的第一次迭代。这里是结果代码 -

class residual_net() 
    def pseudo_bp(self): 
     # do something... 
     return op 

    def build_net(self): 
     # build a residual_network.... 
     rn.output = sth 

def run(): 
    rn = residual_net() 
    operation = rn.pseudo_bp() 
    rn.build_net() 
    err = tf.reduce_mean(rn.output, labels) 
    train = tf.train.GradientDescentOptimizer(learning_rate).minimize(err) 
    # Graph has been built completely. Begin tf.Session() 
    sess = tf.Session() 
    sess.run(tf.initialize_all_variables()) 
    for i in range(1000): 
     # Carry out the training in each iteration 
     # Note that train is an operation here 
     sess.run(train, feed_dict=feed_dict) 
     if i == 0: 
      # Execute `operation` for the first iteration 
      result = sess.run(operation, feed_dict=feed_dict)