Question

我正在尝试在笔记本中实例化ESN类的两个对象，如下所示：

esn_1 = ESN(ESN_arch, activation, leak_rate, weights_variance, sparsity, sparseness) esn_2 = ESN(ESN_arch, activation, leak_rate, weights_variance, sparsity, sparseness)

运行会话时，在sess.run(tf.global_variables_initializer())行弹出以下错误：

FailedPreconditionError: Attempting to use uninitialized value initializers/ReservoirWeights
     [[node initializers/ReservoirWeights/read (defined at /home/tah/Documents/Computation_EOC/esn-neuroevolution/ESN_Cell.py:38) ]]

我认为该错误基本上是由于我在类内部使用了变量作用域。但我似乎无法弄清楚到底是什么。我已经在这里进行了检查：https://stackoverflow.com/a/36016117。

此外，我担心删除variable_scope会导致一个更老的问题：-我将variable_scope与tf.AUTO_RESUSE一起使用（最初是成功的）的最初原因是要获得所有变量的所有权重矩阵对象ESN的实例化具有相同的值。在打开variable_scope之前，请注意我对tf.set_random_seed(1234)的使用。

ESN_Cell.py ：

class ESN(rnn_cell_impl.RNNCell):

     def __init__(...):
            self.in_units = ESN_arch[0]
            self.res_units = ESN_arch[1]
            self.activation = activation
            self.alpha = tf.cast(leak_rate, dtype=tf.float64)
            self.weights_std = tf.cast(weights_std, dtype=tf.float64)
            self.sparsity = tf.cast(sparsity, dtype=tf.float64)
            self.sparseness = sparseness

            tf.set_random_seed(1234)

            with tf.variable_scope('initializers', reuse=tf.AUTO_REUSE):

                  self.weights_in = tf.get_variable("InputWeights", \
                                      initializer=self.init_weights_in(self.weights_std),\
                                      trainable=False, dtype=tf.float64)
                  # 'weights_in' is: [in_units x res_units]

                  self.weights_res = self.normalize_weights_res(tf.get_variable("ReservoirWeights", \
                                       initializer=self.init_weights_res(self.weights_std),\
                                       trainable=False, dtype=tf.float64))
                  # 'weights_res' is: [res_units x res_units]

                  self.bias = tf.get_variable("Bias", \
                                    initializer=self.init_bias(self.weights_std),\
                                    trainable=False, dtype=tf.float64)
                  # 'bias' is: [1, res_units]

                  self.spectral_radius = tf.get_variable("SpectralRadius",\
                                               initializer=self.get_spectral_radius(self.weights_res),\
                                               trainable=False, dtype=tf.float64)

                 if self.sparseness:
                        self.sparse_mask = tf.get_variable("SparseMatrix",\
                                               initializer=self.init_sparse_matrix(self.weights_res), \
                                               trainable=False, dtype=tf.float64)
                        self.weights_res = tf.multiply(self.weights_res, self.sparse_mask)
       .
       .
       .

我的会话如下：

以tf.Session（）作为会话：

# res_units is an int with value 100.
for p_neuron in range(res_units):

    sess.run(tf.global_variables_initializer())

    init_esn_state = np.zeros([1, res_units], dtype="float64")

    print(type(p_neuron))

    dist, initial, init_esn_2 = sess.run([dist_esn_1_2, initial, init_esn_2], \
                                          feed_dict={leak_rate: alpha,\
                                                     inputs:esn_input,\
                                                     init_state:init_esn_state,\
                                                     pert_neuron:p_neuron})

使用变量范围存在问题：“ FailedPreconditionError：尝试使用未初始化的值初始化程序...”

0 个答案: