TF之BN：BN算法对多层中的每层神经网络加快学习QuadraticFunction_InputData+Histogram+BN的Error_curve

TF之BN：BN算法对多层中的每层神经网络加快学习QuadraticFunction_InputData+Histogram+BN的Error_curve

目录

输出结果

代码设计

输出结果

代码设计

 23 Batch Normalization
 
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
 
 
ACTIVATION = tf.nn.tanh
N_LAYERS = 7           
N_HIDDEN_UNITS = 30   
 
 
def fix_seed(seed=1): 
     reproducible
    np.random.seed(seed)
    tf.set_random_seed(seed)
 
 
def plot_his(inputs, inputs_norm):
     plot histogram for the inputs of every layer
     
    for j, all_inputs in enumerate([inputs, inputs_norm]):
        for i, input in enumerate(all_inputs):
            plt.subplot(2, len(all_inputs), j*len(all_inputs)+(i+1))
            plt.cla()
            if i == 0:
                the_range = (-7, 10)
            else:
                the_range = (-1, 1)
            plt.hist(input.ravel(), bins=15, range=the_range, color='0000FF')
            plt.yticks(())
            if j == 1:
                plt.xticks(the_range)
            else:
                plt.xticks(())
            ax = plt.gca()
            ax.spines['right'].set_color('none')
            ax.spines['top'].set_color('none')
        plt.title("%s normalizing" % ("Without" if j == 0 else "With"))
    plt.title('Matplotlib,BN,histogram--Jason Niu')
    plt.draw()
    plt.pause(0.001)
 
 
def built_net(xs, ys, norm): 
    def add_layer(inputs, in_size, out_size, activation_function=None, norm=False):
         weights and biases (bad initialization for this case)
        Weights = tf.Variable(tf.random_normal([in_size, out_size], mean=0., stddev=1.))
        biases = tf.Variable(tf.zeros([1, out_size]) + 0.1)
 
         fully connected product
        Wx_plus_b = tf.matmul(inputs, Weights) + biases
 
         normalize fully connected product
        if norm:
             Batch Normalize
            fc_mean, fc_var = tf.nn.moments( 
                Wx_plus_b,
                axes=[0],  
                             
            )
            scale = tf.Variable(tf.ones([out_size]))
            shift = tf.Variable(tf.zeros([out_size]))
            epsilon = 0.001
 
             apply moving average for mean and var when train on batch
            ema = tf.train.ExponentialMovingAverage(decay=0.5)
            def mean_var_with_update():
                ema_apply_op = ema.apply([fc_mean, fc_var])
                with tf.control_dependencies([ema_apply_op]):
                    return tf.identity(fc_mean), tf.identity(fc_var)
            mean, var = mean_var_with_update()
 
            Wx_plus_b = tf.nn.batch_normalization(Wx_plus_b, mean, var, shift, scale, epsilon)
             
             Wx_plus_b = (Wx_plus_b - fc_mean) / tf.sqrt(fc_var + 0.001)  进行BN一下
             Wx_plus_b = Wx_plus_b * scale + shift
  
         activation
        if activation_function is None:
            outputs = Wx_plus_b
        else:
            outputs = activation_function(Wx_plus_b)
 
        return outputs  输出激活结果
 
    fix_seed(1)
 
    if norm:
         BN for the first input
        fc_mean, fc_var = tf.nn.moments(
            xs,
            axes=[0],
        )
        scale = tf.Variable(tf.ones([1]))
        shift = tf.Variable(tf.zeros([1]))
        epsilon = 0.001
         apply moving average for mean and var when train on batch
        ema = tf.train.ExponentialMovingAverage(decay=0.5)
        def mean_var_with_update():
            ema_apply_op = ema.apply([fc_mean, fc_var])
            with tf.control_dependencies([ema_apply_op]):
                return tf.identity(fc_mean), tf.identity(fc_var)
        mean, var = mean_var_with_update()
        xs = tf.nn.batch_normalization(xs, mean, var, shift, scale, epsilon)
 
     record inputs for every layer
    layers_inputs = [xs] 
 
     build hidden layers
    for l_n in range(N_LAYERS):
        layer_input = layers_inputs[l_n]
        in_size = layers_inputs[l_n].get_shape()[1].value
 
        output = add_layer(
            layer_input,     input
            in_size,         input size
            N_HIDDEN_UNITS,  output size
            ACTIVATION,      activation function
            norm,            normalize before activation
        )
        layers_inputs.append(output)  
 
     build output layer
    prediction = add_layer(layers_inputs[-1], 30, 1, activation_function=None)
 
    cost = tf.reduce_mean(tf.reduce_sum(tf.square(ys - prediction), reduction_indices=[1]))
    train_op = tf.train.GradientDescentOptimizer(0.001).minimize(cost)
    return [train_op, cost, layers_inputs] 
 
fix_seed(1)
x_data = np.linspace(-7, 10, 2500)[:, np.newaxis]  水平轴-7~10
np.random.shuffle(x_data)
noise = np.random.normal(0, 8, x_data.shape)
y_data = np.square(x_data) - 5 + noise
 
 
xs = tf.placeholder(tf.float32, [None, 1])   [num_samples, num_features]
ys = tf.placeholder(tf.float32, [None, 1])
 
建立两个神经网络作对比
train_op, cost, layers_inputs = built_net(xs, ys, norm=False) 
train_op_norm, cost_norm, layers_inputs_norm = built_net(xs, ys, norm=True)
 
 
sess = tf.Session()
if int((tf.__version__).split('.')[1]) < 12 and int((tf.__version__).split('.')[0]) < 1:
    init = tf.initialize_all_variables()
else:
    init = tf.global_variables_initializer()
sess.run(init)
 
 record cost
cost_his = []       
cost_his_norm = []  
record_step = 5     
 
plt.ion()
plt.figure(figsize=(7, 3)) 
for i in range(250):
    if i % 50 == 0: 
         plot histogram
        all_inputs, all_inputs_norm = sess.run([layers_inputs, layers_inputs_norm], feed_dict={xs: x_data, ys: y_data})
        plot_his(all_inputs, all_inputs_norm)
 
     train on batch每一步都run一下
    sess.run([train_op, train_op_norm], feed_dict={xs: x_data[i*10:i*10+10], ys: y_data[i*10:i*10+10]})
 
    if i % record_step == 0:
         record cost
        cost_his.append(sess.run(cost, feed_dict={xs: x_data, ys: y_data}))
        cost_his_norm.append(sess.run(cost_norm, feed_dict={xs: x_data, ys: y_data}))
 
以下是绘制误差值Cost误差曲线的方法
plt.ioff()
plt.figure()
plt.title('Matplotlib,BN,Error_curve--Jason Niu')
plt.plot(np.arange(len(cost_his))*record_step, np.array(cost_his), label='no BN')      no norm
plt.plot(np.arange(len(cost_his))*record_step, np.array(cost_his_norm), label='BN')    norm
plt.legend()
plt.show()

相关文章
TF之BN：BN算法对多层中的每层神经网络加快学习QuadraticFunction_InputData+Histogram+BN的Error_curve