mnist数据集是一款比较经典的各种手写数字体的集合,一般被初学者用来练手
我使用的框架为tensorflow,前面主要未解决的问题是BP神经网络的反向传播,所以在这里使用了tf框架。
下面开始我的项目搭建过程。
步骤:
0加载数据集的数据
import numpy as np
import matplotlib.pyplot as plt
from keras.datasets import mnist
(X_train_orig, Y_train_orig), (X_test_orig, Y_test_orig) = mnist.load_data()
print(X_train_orig.shape)
print(Y_train_orig.shape)
#(60000, 28, 28)
#(60000,)
0.1对数据的相关处理
说明:用python识别mnist的项目本质上是一个分类器项目,通过识别不同图片内部的特征,将其归类,所以,在最后我们输出的是[10,?]的一个张量,?代表样本数量,10是0-9的十个分类的个数。
这里需要对标签进行one-hot编码。
def one_hot_matrix(lables,C):
"""
创建一个矩阵,其中第i行对应第i个类号,第j列对应第j个训练样本
所以如果第j个样本对应着第i个标签,那么entry (i,j)将会是1
参数:
lables - 标签向量
C - 分类数
返回:
one_hot - 独热矩阵
"""
#创建一个tf.constant,赋值为C,名字叫C
C = tf.constant(C,name="C")
#使用tf.one_hot,注意一下axis
one_hot_matrix = tf.one_hot(indices=lables , depth=C , axis=0)
#创建一个session
sess = tf.Session()
#运行session
one_hot = sess.run(one_hot_matrix)
#关闭session
sess.close()
return one_hot
X_train_flatten = X_train_orig.reshape(X_train_orig.shape[0],-1).T #每一列就是一个样本
X_test_flatten = X_test_orig.reshape(X_test_orig.shape[0],-1).T
#归一化数据
X_train = X_train_flatten / 255
X_test = X_test_flatten / 255
#转换为独热矩阵
Y_train = one_hot_matrix(Y_train_orig,10)
Y_test = one_hot_matrix(Y_test_orig,10)
print("训练集样本数 = " + str(X_train.shape[1]))
print("测试集样本数 = " + str(X_test.shape[1]))
print("X_train.shape: " + str(X_train.shape))
print("Y_train.shape: " + str(Y_train.shape))
print("X_test.shape: " + str(X_test.shape))
print("Y_test.shape: " + str(Y_test.shape))
#训练集样本数 = 60000
#测试集样本数 = 10000
#X_train.shape: (784, 60000)
#Y_train.shape: (10, 60000)
#X_test.shape: (784, 10000)
#Y_test.shape: (10, 10000)
1 初始化网络参数
1.1
layers_dims = [784, 100, 50, 10]
def initialize_parameters(layers_dims):
parameters = {}
L = len(layers_dims)
for l in range(1, L):
parameters["W" + str(l)] = tf.get_variable("W"+str(l),[layers_dims[l],layers_dims[l-1]],initializer=tf.contrib.layers.xavier_initializer(seed=1))
parameters["b" + str(l)] = tf.get_variable("b"+str(l),[layers_dims[l],1],initializer=tf.zeros_initializer())
assert(parameters["W"+str(l)].shape == (layers_dims[l],layers_dims[l-1]))
assert(parameters["b"+str(l)].shape == (layers_dims[l],1))
return parameters#包含的参数为(W,b)
这里用tf的相关函数保证了数据类型的一致?这里我不确定,但前面用常规的初始化确实遇到了数据不匹配的问题。
这里神经网络的层数可以根据layers_dims中数量的变化来发生变化,比较灵活。
当然,也有直接定义几层参数的方法,三层神经网络的权重初始化如下代码所示:
def initialize_parameters():
"""
初始化神经网络的参数,参数的维度如下:
W1 : [25, 784]
b1 : [25, 1]
W2 : [12, 25]
b2 : [12, 1]
W3 : [10, 12]
b3 : [10, 1]
返回:
parameters - 包含了W和b的字典
"""
tf.set_random_seed(1) #指定随机种子
W1 = tf.get_variable("W1",[25,784],initializer=tf.contrib.layers.xavier_initializer(seed=1))
b1 = tf.get_variable("b1",[25,1],initializer=tf.zeros_initializer())
W2 = tf.get_variable("W2", [12, 25], initializer = tf.contrib.layers.xavier_initializer(seed=1))
b2 = tf.get_variable("b2", [12, 1], initializer = tf.zeros_initializer())
W3 = tf.get_variable("W3", [10, 12], initializer = tf.contrib.layers.xavier_initializer(seed=1))
b3 = tf.get_variable("b3", [10, 1], initializer = tf.zeros_initializer())
parameters = {"W1": W1,
"b1": b1,
"W2": W2,
"b2": b2,
"W3": W3,
"b3": b3}
return parameters
tf.reset_default_graph() #用于清除默认图形堆栈并重置全局默认图形。
with tf.Session() as sess:
parameters = initialize_parameters(layers_dims)
print("W1 = " + str(parameters["W1"]))
print("b1 = " + str(parameters["b1"]))
print("W2 = " + str(parameters["W2"]))
print("b2 = " + str(parameters["b2"]))
#W1 = <tf.Variable 'W1:0' shape=(100, 784) dtype=float32_ref>
#b1 = <tf.Variable 'b1:0' shape=(100, 1) dtype=float32_ref>
#W2 = <tf.Variable 'W2:0' shape=(50, 100) dtype=float32_ref>
#b2 = <tf.Variable 'b2:0' shape=(50, 1) dtype=float32_ref>
#向前传播第一次初始化的输出
2 向前传播
def forward_propagation(X,parameters):
"""
实现一个模型的前向传播,模型结构为LINEAR -> RELU -> LINEAR -> RELU -> LINEAR -> SOFTMAX
参数:
X - 输入数据的占位符,维度为(输入节点数量,样本数量)
parameters - 包含了W和b的参数的字典
返回:
Z3 - 最后一个LINEAR节点的输出
"""
W1 = parameters['W1']
b1 = parameters['b1']
W2 = parameters['W2']
b2 = parameters['b2']
W3 = parameters['W3']
b3 = parameters['b3']
Z1 = tf.add(tf.matmul(W1,X),b1) # Z1 = np.dot(W1, X) + b1
#Z1 = tf.matmul(W1,X) + b1 #也可以这样写
A1 = tf.nn.relu(Z1) # A1 = relu(Z1)
Z2 = tf.add(tf.matmul(W2, A1), b2) # Z2 = np.dot(W2, a1) + b2
A2 = tf.nn.relu(Z2) # A2 = relu(Z2)
Z3 = tf.add(tf.matmul(W3, A2), b3) # Z3 = np.dot(W3,Z2) + b3
return Z3
tf.reset_default_graph() #用于清除默认图形堆栈并重置全局默认图形。
with tf.Session() as sess:
X,Y = create_placeholders(784,10)
parameters = initialize_parameters(layers_dims)
Z3 = forward_propagation(X,parameters)
print("Z3 = " + str(Z3))
3 计算误差
def compute_cost(Z3,Y):
"""
计算成本
参数:
Z3 - 前向传播的结果
Y - 标签,一个占位符,和Z3的维度相同
返回:
cost - 成本值
"""
logits = tf.transpose(Z3) #转置
labels = tf.transpose(Y) #转置
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits,labels=labels))
#这里进行softmax激活函数的处理了。
return cost
tf.reset_default_graph()
with tf.Session() as sess:
X,Y = create_placeholders(784,10)
parameters = initialize_parameters(layers_dims)
Z3 = forward_propagation(X,parameters)
print("Z3 = " + str(Z3))
print("Y = " + str(Y))
cost = compute_cost(Z3,Y)
print("cost = " + str(cost))
4 反向传播,最小化损失函数
optimizer = tf.train.GradientDescentOptimizer(learning_rate = learning_rate).minimize(cost)
_ , c = sess.run([optimizer,cost],feed_dict={X:mini_batch_X,Y:mini_batch_Y})
5 构建神经网络,训练参数
#小批量训练
def random_mini_batches(X, Y, mini_batch_size = 64, seed = 0):
"""
Creates a list of random minibatches from (X, Y)
Arguments:
X -- input data, of shape (input size, number of examples)
Y -- true "label" vector (containing 0 if cat, 1 if non-cat), of shape (1, number of examples)
mini_batch_size - size of the mini-batches, integer
seed -- this is only for the purpose of grading, so that you're "random minibatches are the same as ours.
Returns:
mini_batches -- list of synchronous (mini_batch_X, mini_batch_Y)
"""
m = X.shape[1] # number of training examples
mini_batches = []
np.random.seed(seed)
# Step 1: Shuffle (X, Y)
permutation = list(np.random.permutation(m))
shuffled_X = X[:, permutation]
shuffled_Y = Y[:, permutation].reshape((Y.shape[0],m))
# Step 2: Partition (shuffled_X, shuffled_Y). Minus the end case.
num_complete_minibatches = math.floor(m/mini_batch_size) # number of mini batches of size mini_batch_size in your partitionning
for k in range(0, num_complete_minibatches):
mini_batch_X = shuffled_X[:, k * mini_batch_size : k * mini_batch_size + mini_batch_size]
mini_batch_Y = shuffled_Y[:, k * mini_batch_size : k * mini_batch_size + mini_batch_size]
mini_batch = (mini_batch_X, mini_batch_Y)
mini_batches.append(mini_batch)
# Handling the end case (last mini-batch < mini_batch_size)
if m % mini_batch_size != 0:
mini_batch_X = shuffled_X[:, num_complete_minibatches * mini_batch_size : m]
mini_batch_Y = shuffled_Y[:, num_complete_minibatches * mini_batch_size : m]
mini_batch = (mini_batch_X, mini_batch_Y)
mini_batches.append(mini_batch)
return mini_batches
def model(X_train,Y_train,X_test,Y_test,
learning_rate=0.0001,num_epochs=1500,minibatch_size=32,
print_cost=True,is_plot=True):
"""
实现一个三层的TensorFlow神经网络:LINEAR->RELU->LINEAR->RELU->LINEAR->SOFTMAX
参数:
X_train - 训练集,维度为(输入大小(输入节点数量) = 12288, 样本数量 = 1080)
Y_train - 训练集分类数量,维度为(输出大小(输出节点数量) = 6, 样本数量 = 1080)
X_test - 测试集,维度为(输入大小(输入节点数量) = 12288, 样本数量 = 120)
Y_test - 测试集分类数量,维度为(输出大小(输出节点数量) = 6, 样本数量 = 120)
learning_rate - 学习速率
num_epochs - 整个训练集的遍历次数
mini_batch_size - 每个小批量数据集的大小
print_cost - 是否打印成本,每100代打印一次
is_plot - 是否绘制曲线图
返回:
parameters - 学习后的参数
"""
ops.reset_default_graph() #能够重新运行模型而不覆盖tf变量
tf.set_random_seed(1)
seed = 3
(n_x , m) = X_train.shape #获取输入节点数量和样本数
n_y = Y_train.shape[0] #获取输出节点数量
costs = [] #成本集
#给X和Y创建placeholder
X,Y = create_placeholders(n_x,n_y)
#初始化参数
parameters = initialize_parameters(layers_dims)
#前向传播
Z3 = forward_propagation(X,parameters)
#计算成本
cost = compute_cost(Z3,Y)
#反向传播,使用Adam优化
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
#初始化所有的变量
init = tf.global_variables_initializer()
#开始会话并计算
with tf.Session() as sess:
#初始化
sess.run(init)
#正常训练的循环
for epoch in range(num_epochs):
epoch_cost = 0 #每代的成本
num_minibatches = int(m / minibatch_size) #minibatch的总数量
seed = seed + 1
minibatches = random_mini_batches(X_train,Y_train,minibatch_size,seed)
for minibatch in minibatches:
#选择一个minibatch
(minibatch_X,minibatch_Y) = minibatch
#数据已经准备好了,开始运行session
_ , minibatch_cost = sess.run([optimizer,cost],feed_dict={X:minibatch_X,Y:minibatch_Y})
#计算这个minibatch在这一代中所占的误差
epoch_cost = epoch_cost + minibatch_cost / num_minibatches
#记录并打印成本
## 记录成本
if epoch % 5 == 0:
costs.append(epoch_cost)
#是否打印:
if print_cost and epoch % 5 == 0:
print("epoch = " + str(epoch) + " epoch_cost = " + str(epoch_cost))
#是否绘制图谱
if is_plot:
plt.plot(np.squeeze(costs))
plt.ylabel('cost')
plt.xlabel('iterations (per tens)')
plt.title("Learning rate =" + str(learning_rate))
plt.show()
#保存学习后的参数
parameters = sess.run(parameters)
print("参数已经保存到session。")
#计算当前的预测结果
correct_prediction = tf.equal(tf.argmax(Z3),tf.argmax(Y))
#计算准确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction,"float"))
print("训练集的准确率:", accuracy.eval({X: X_train, Y: Y_train}))
print("测试集的准确率:", accuracy.eval({X: X_test, Y: Y_test}))
return parameters
#开始时间
start_time = time.clock()
#开始训练
parameters = model(X_train, Y_train, X_test, Y_test,num_epochs=150, minibatch_size=512)
#结束时间
end_time = time.clock()
#计算时差
print("CPU的执行时间 = " + str(end_time - start_time) + " 秒" )
'''
epoch = 0 epoch_cost = 1.8048073939788036
epoch = 5 epoch_cost = 0.3332515982990592
epoch = 10 epoch_cost = 0.2444374926834026
epoch = 15 epoch_cost = 0.20180034395466495
epoch = 20 epoch_cost = 0.17208938682690644
epoch = 25 epoch_cost = 0.1487552207122501
epoch = 30 epoch_cost = 0.13145502363769412
epoch = 35 epoch_cost = 0.11671997498498006
epoch = 40 epoch_cost = 0.1042079112978063
epoch = 45 epoch_cost = 0.09372433128519962
epoch = 50 epoch_cost = 0.08588403874100783
epoch = 55 epoch_cost = 0.07700667132297136
epoch = 60 epoch_cost = 0.07071055720249812
epoch = 65 epoch_cost = 0.06447908312528051
epoch = 70 epoch_cost = 0.0591392962851076
epoch = 75 epoch_cost = 0.054782480272090345
epoch = 80 epoch_cost = 0.05065463378261298
epoch = 85 epoch_cost = 0.04652624101274543
epoch = 90 epoch_cost = 0.042965341470817195
epoch = 95 epoch_cost = 0.03951436747661513
epoch = 100 epoch_cost = 0.036161916457817085
epoch = 105 epoch_cost = 0.03350572805437777
epoch = 110 epoch_cost = 0.03098719434924104
epoch = 115 epoch_cost = 0.028340620108139824
epoch = 120 epoch_cost = 0.02633996693910952
epoch = 125 epoch_cost = 0.023921261095784154
epoch = 130 epoch_cost = 0.021953271526811466
epoch = 135 epoch_cost = 0.020309688396051403
epoch = 140 epoch_cost = 0.018723787532912355
epoch = 145 epoch_cost = 0.01693216629294503
参数已经保存到session。
训练集的准确率: 0.9978833
测试集的准确率: 0.9775
CPU的执行时间 = 312.2200849999999 秒
'''
6 预测
import matplotlib.pyplot as plt # plt 用于显示图片
import matplotlib.image as mpimg # mpimg 用于读取图片
import numpy as np
#这是博主自己拍的图片
my_image1 = "0-5.bmp" #定义图片名称
fileName1 = "/Users/weijinqian/Documents/python/test/" + my_image1 #图片地址
image1 = mpimg.imread(fileName1) #读取图片
plt.imshow(image1) #显示图片
my_image1 = image1.reshape(1,28 * 28).T #重构图片
my_image_prediction = predict(my_image1, parameters) #开始预测
print("预测结果: y = " + str(np.squeeze(my_image_prediction)))
plt.show(image1)
预测结果: y = 5
在这里,plt中的show()与imshow()函数出现了一些问题,我暂时还不是很理解。
写到这里就已经完成了识别手写数字体这个简单的项目了。在该过程中用到的数据集为:
10k-labels-idx1-ubyte.gz
train-labels-idx1-ubyte.gz
t10k-images-idx3-ubyte.gz
train-images-idx3-ubyte.gz
网上很容易下载的。
具体加载数据集的函数,大家可以参考keras框架的load_mnist()函数,我对具体实现不太了解,后期可能会去研究。
对于项目的优化或者改进,大家可以从以下几个方面进行:
1. 正则化
2. dropout随机失活
3. 优化函数,比如说adam等
4. 各种超参数的调试,比如:学习率,神经网络层数,迭代次数等
5. 可以使用卷积神经网络(CNN)来对数据进行处理。
卷积神经网络的识别数字体项目会在下一个博客中展示,期待以后的更新吧。