莫烦tensorflow Batch Normalization 代码加注释

莫烦tensorflow Baych Normalization 代码加注释。

"""

Know more, visit my Python tutorial page: https://morvanzhou.github.io/tutorials/
My Youtube Channel: https://www.youtube.com/user/MorvanZhou


Dependencies:
tensorflow: 1.1.0
matplotlib
numpy
"""
#import package
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt


tf.set_random_seed(1)
np.random.seed(1)


# Hyper parameters
N_SAMPLES = 2000
BATCH_SIZE = 64
EPOCH = 12
LR = 0.03
N_HIDDEN = 8
ACTIVATION = tf.nn.tanh
B_INIT = tf.constant_initializer(-0.2)      # use a bad bias initialization


# training data size is N_SAMPLES 
x = np.linspace(-7, 10, N_SAMPLES)[:, np.newaxis]
np.random.shuffle(x)
noise = np.random.normal(0, 2, x.shape)
y = np.square(x) - 5 + noise
train_data = np.hstack((x, y))


# hstack表示轴1合并。hstack的字母h来自于horizontal，表示两个数组是水平的，hstack((a,b))将把b排在a的右边的意思。
# test data size is 200,由于hstack()所以后边有b_x, b_y = train_data[b_s: b_f, 0:1], train_data[b_s: b_f, 1:2]这一操作
test_x = np.linspace(-7, 10, 200)[:, np.newaxis]
noise = np.random.normal(0, 2, test_x.shape)
test_y = np.square(test_x) - 5 + noise


# plot input data
#==============================================================================
# plt.scatter(x, y, c='#FF9359', s=50, alpha=0.5, label='train')
# plt.legend(loc='upper left')
# 
#==============================================================================
# tensorflow placeholder
tf_x = tf.placeholder(tf.float32, [None, 1])
tf_y = tf.placeholder(tf.float32, [None, 1])
tf_is_train = tf.placeholder(tf.bool, None)     # flag for using BN on training or testing




class NN(object):
    def __init__(self, batch_normalization=False):
        self.is_bn = batch_normalization


        self.w_init = tf.random_normal_initializer(0., .1)  # weights initialization
        # 保存activation之前的输入，如果是标准化网络net[1]，就是标准输入数据列表，如果是不标准化网络net[0]，就是原始输入数据列表.
        self.pre_activation = [tf_x]  
        #对输入层的输入数据判断是否需要标准化，如果标准化网络，就每层输入都标准化，都则就都不标准化
        if self.is_bn:
            self.layer_input = [tf.layers.batch_normalization(tf_x, training=tf_is_train)]  # for input data
        else:
            self.layer_input = [tf_x]
        # 添加隐藏层，每一层的输入就是前一层的输出
        for i in range(N_HIDDEN):  # adding hidden layers
            self.layer_input.append(self.add_layer(self.layer_input[-1], 10, ac=ACTIVATION))
        #添加输出层
        self.out = tf.layers.dense(self.layer_input[-1], 1, kernel_initializer=self.w_init, bias_initializer=B_INIT)
        #损失计算操作
        self.loss = tf.losses.mean_squared_error(tf_y, self.out)


        # !! IMPORTANT !! the moving_mean and moving_variance need to be updated,
        # pass the update_ops with control_dependencies to the train_op
        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
        with tf.control_dependencies(update_ops):
            self.train = tf.train.AdamOptimizer(LR).minimize(self.loss)


    def add_layer(self, x, out_size, ac=None):
        x = tf.layers.dense(x, out_size, kernel_initializer=self.w_init, bias_initializer=B_INIT)
        self.pre_activation.append(x)
        # the momentum plays important rule. the default 0.99 is too high in this case!
        # 数据标准化函数tf.layers.batch_normalization()
        if self.is_bn: x = tf.layers.batch_normalization(x, momentum=0.4, training=tf_is_train)    # when have BN
        out = x if ac is None else ac(x) #是否需要激活函数
        return out


# 创建两个网络，保存在列表里
nets = [NN(batch_normalization=False), NN(batch_normalization=True)]    # two nets, with and without BN


sess = tf.Session()
sess.run(tf.global_variables_initializer())


# plot layer input distribution
f, axs = plt.subplots(4, N_HIDDEN+1, figsize=(10, 5))
plt.ion()   # something about plotting


def plot_histogram(l_in, l_in_bn, pre_ac, pre_ac_bn):
    for i, (ax_pa, ax_pa_bn, ax,  ax_bn) in enumerate(zip(axs[0, :], axs[1, :], axs[2, :], axs[3, :])):
        [a.clear() for a in [ax_pa, ax_pa_bn, ax, ax_bn]]
        if i == 0: p_range = (-7, 10); the_range = (-7, 10)
        else: p_range = (-4, 4); the_range = (-1, 1)
        ax_pa.set_title('L' + str(i))
        ax_pa.hist(pre_ac[i].ravel(), bins=10, range=p_range, color='#FF9359', alpha=0.5)
        ax_pa_bn.hist(pre_ac_bn[i].ravel(), bins=10, range=p_range, color='#74BCFF', alpha=0.5)
        ax.hist(l_in[i].ravel(), bins=10, range=the_range, color='#FF9359')
        ax_bn.hist(l_in_bn[i].ravel(), bins=10, range=the_range, color='#74BCFF')
        for a in [ax_pa, ax, ax_pa_bn, ax_bn]:
            a.set_yticks(()); a.set_xticks(())
        ax_pa_bn.set_xticks(p_range); ax_bn.set_xticks(the_range); axs[2, 0].set_ylabel('Act'); axs[3, 0].set_ylabel('BN Act')
    plt.pause(0.01)


losses = [[], []]   # record test loss
for epoch in range(EPOCH):
    print('Epoch: ', epoch)
    np.random.shuffle(train_data)
    step = 0
    in_epoch = True
    while in_epoch:
        #数据开始和结尾
        b_s, b_f = (step*BATCH_SIZE) % len(train_data), ((step+1)*BATCH_SIZE) % len(train_data) # batch index
        step += 1
        if b_f < b_s: #如果取完了所有数据in_epoch = False
            b_f = len(train_data)
            in_epoch = False
        b_x, b_y = train_data[b_s: b_f, 0:1], train_data[b_s: b_f, 1:2] # batch training data 分别输入的是x,y
        sess.run([nets[0].train, nets[1].train], {tf_x: b_x, tf_y: b_y, tf_is_train: True})     # train
        
        # 对每一个epoch,step==0 进行训练 tf_is_train: True，step==1进行测试 tf_is_train: False
        if step == 1:
            l0, l1, l_in, l_in_bn, pa, pa_bn = sess.run(
                [nets[0].loss, nets[1].loss, nets[0].layer_input, nets[1].layer_input,
                 nets[0].pre_activation, nets[1].pre_activation],
                {tf_x: test_x, tf_y: test_y, tf_is_train: False}) 
            
            # 保存loss到列表用于画图
            [loss.append(l) for loss, l in zip(losses, [l0, l1])]   # recode test loss
            plot_histogram(l_in, l_in_bn, pa, pa_bn)     # plot histogram


plt.ioff()


# plot test loss
plt.figure(2)
plt.plot(losses[0], c='#FF9359', lw=3, label='Original')
plt.plot(losses[1], c='#74BCFF', lw=3, label='Batch Normalization')
plt.ylabel('test loss'); plt.ylim((0, 2000)); plt.legend(loc='best')


# plot prediction line 进行预测，得到标准化网络预测值pred_bn和原始网络预测值pred，并画出数据点，标准化网络预测的效果很好.
pred, pred_bn = sess.run([nets[0].out, nets[1].out], {tf_x: test_x, tf_is_train: False})
plt.figure(3)
plt.plot(test_x, pred, c='#FF9359', lw=4, label='Original')
plt.plot(test_x, pred_bn, c='#74BCFF', lw=4, label='Batch Normalization')
plt.scatter(x[:200], y[:200], c='r', s=50, alpha=0.2, label='train')
plt.legend(loc='best'); plt.show()