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])