batch_normalization 中错误维度的 beta

Question

I am trying to implement batch_normalization in a neural Network in Tensorflow. Initialization is done with variable_scope. When I run my program, I get an error:

*** ValueError: Trying to share variable scope1/beta, but specified shape (100,) and found shape (100, 1).

As I understand it, I must initialize the beta (and the other stored parameters) with a correct dimension. I believe that my dimension should be (100,1) as there are 100 hidden units in my first hidden layer. How can I specify the shape as to fit the expectations of my batch_normalization?

Here is my code:

    Layers = [203,100,100,100,2]    

    def initialize_parameters(Layers,m):
        params = {}
        for i in range(len(Layers)):
            if i > 0:
            with tf.variable_scope("scope"+str(i), reuse=None) as sp:
                beta = tf.get_variable("beta", [Layers[i],1], initializer=tf.constant_initializer(0.0))
                gamma = tf.get_variable("gamma",[Layers[i],1], initializer=tf.constant_initializer(1.0))
                moving_avg = tf.get_variable("moving_mean", [Layers[i],1], initializer=tf.constant_initializer(0.0),trainable=False)
                moving_var = tf.get_variable("moving_variance", [Layers[i],1], initializer=tf.constant_initializer(1.0),trainable=False)                
            lastUnits = Layers[i]
        sp.reuse_variables()
        return params

    def forward_propagation(X, Layers, parameters, keep_prob):
        for i in range(len(Layers)):
            ...
            Z_BN = tf.contrib.layers.batch_norm(Z,is_training=True,updates_collections=ops.GraphKeys.UPDATE_OPS,scope="scope"+str(i), reuse=True)
           ...
        return Z_BN

Answer 1

The solution was to just give one number as shape:

    beta = tf.get_variable("beta", shape=Layers[i], initializer=tf.constant_initializer(0.0))
    gamma = tf.get_variable("gamma", shape=Layers[i], initializer=tf.constant_initializer(1.0))
    moving_avg = tf.get_variable("moving_mean", shape=Layers[i], initializer=tf.constant_initializer(0.0),trainable=False)
    moving_var = tf.get_variable("moving_variance", shape=Layers[i], initializer=tf.constant_initializer(1.0),trainable=False)                

Hope, this helps.