train.py 文件源码

def makeDNN(hidden_layer):
    # input from X
    prevLayer = X

    # make layers
    for i in range(hidden_layer):
        if i==0:
            newWeight = tf.get_variable("W0%d" % i, shape=[features, wide], initializer=tf.contrib.layers.xavier_initializer())
        else:
            newWeight = tf.get_variable("W0%d" % i, shape=[wide, wide], initializer=tf.contrib.layers.xavier_initializer())
        newBias = tf.Variable(tf.random_normal([wide]))
        newLayer = tf.nn.relu(tf.matmul(prevLayer, newWeight) + newBias)
        newDropLayer = tf.nn.dropout(newLayer, dropout_rate)
        prevLayer = newDropLayer

    # make output layers
    Wo = tf.get_variable("Wo", shape=[wide, labels], initializer=tf.contrib.layers.xavier_initializer())
    bo = tf.Variable(tf.random_normal([labels]))
    return tf.matmul(prevLayer, Wo) + bo

# tf Graph Input