Tensorflow执行时间

Question

我在Python脚本中有多次调用的函数（https://github.com/sankhaMukherjee/NNoptExpt/blob/dev/src/lib/NNlib/NNmodel.py）：为此，我已大大简化了该函数例。

def errorValW(self, X, y, weights):

    errVal = None

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())

        nW = len(self.allW)
        W = weights[:nW] 
        B = weights[nW:]

        for i in range(len(W)):
            sess.run(tf.assign( self.allW[i], W[i] ))

        for i in range(len(B)):
            sess.run(tf.assign( self.allB[i], B[i] ))

        errVal = sess.run(self.err, 
            feed_dict = {self.Inp: X, self.Op: y})

    return errVal

我从另一个函数多次调用此函数。当我看到程序日志时，似乎此功能花费的时间越来越长。显示了部分日志：

21:37:12,634 - ... .errorValW ... - Finished the function [errorValW] in 1.477610e+00 seconds
21:37:14,116 - ... .errorValW ... - Finished the function [errorValW] in 1.481470e+00 seconds
21:37:15,608 - ... .errorValW ... - Finished the function [errorValW] in 1.490914e+00 seconds
21:37:17,113 - ... .errorValW ... - Finished the function [errorValW] in 1.504651e+00 seconds
21:37:18,557 - ... .errorValW ... - Finished the function [errorValW] in 1.443876e+00 seconds
21:37:20,183 - ... .errorValW ... - Finished the function [errorValW] in 1.625608e+00 seconds
21:37:21,719 - ... .errorValW ... - Finished the function [errorValW] in 1.534915e+00 seconds
... many lines later  
22:59:26,524 - ... .errorValW ... - Finished the function [errorValW] in 9.576592e+00 seconds
22:59:35,991 - ... .errorValW ... - Finished the function [errorValW] in 9.466405e+00 seconds
22:59:45,708 - ... .errorValW ... - Finished the function [errorValW] in 9.716456e+00 seconds
22:59:54,991 - ... .errorValW ... - Finished the function [errorValW] in 9.282923e+00 seconds
23:00:04,407 - ... .errorValW ... - Finished the function [errorValW] in 9.415035e+00 seconds

有没有其他人经历过这样的事情？这完全让我感到困惑…

编辑 ：这是供参考…

作为参考，该类的初始化器如下所示。我怀疑result变量的图的大小正在逐渐增加。我尝试保存模型时看到了这个问题，tf.train.Saver(tf.trainable_variables())并且此文件的大小一直在增加。我不确定在以任何方式定义模型时是否犯了错误…

def __init__(self, inpSize, opSize, layers, activations):

    self.inpSize = inpSize
    self.Inp     = tf.placeholder(dtype=tf.float32, shape=inpSize, name='Inp')
    self.Op      = tf.placeholder(dtype=tf.float32, shape=opSize, name='Op')

    self.allW    = []
    self.allB    = []

    self.result  = None

    prevSize = inpSize[0]
    for i, l in enumerate(layers):
        tempW = tf.Variable( 0.1*(np.random.rand(l, prevSize) - 0.5), dtype=tf.float32, name='W_{}'.format(i) )
        tempB = tf.Variable( 0, dtype=tf.float32, name='B_{}'.format(i) )

        self.allW.append( tempW )
        self.allB.append( tempB )

        if i == 0:
            self.result = tf.matmul( tempW, self.Inp ) + tempB
        else:
            self.result = tf.matmul( tempW, self.result ) + tempB

        prevSize = l

        if activations[i] is not None:
            self.result = activations[i]( self.result )

    self.err = tf.sqrt(tf.reduce_mean((self.Op - self.result)**2))


    return

Answer 1

您正在tf.assign会话上下文中进行呼叫。每次执行errorValW函数时，都会在图表上不断添加操作，随着图表的变大，执行速度会变慢。根据经验，在对数据执行模型时，应避免调用Tensorflow操作（因为这通常会在循环内，从而导致图形不断增长）。根据我的个人经验，即使您在执行期间仅添加“几个”操作，也可能导致极慢的运行速度。

请注意，这tf.assign是任何其他操作。您应该事先定义一次（在创建模型/构建图形时），然后在启动会话后重复运行相同的op。

我不知道您在代码段中到底要实现什么，但是请考虑以下几点：

...
with tf.Session() as sess:
    sess.run(tf.assign(some_var, a_value))

可以替换为

a_placeholder = tf.placeholder(type_for_a_value, shape_for_a_value)
assign_op = tf.assign(some_var, a_placeholder)
...
with tf.Session() as sess:
    sess.run(assign_op, feed_dict={a_placeholder: a_value})

其中a_placeholder应具有相同的D型/形状some_var。我必须承认我还没有测试过此代码段，所以如果有问题，请告诉我，但这应该是正确的。