Tensorflow快速入门——知识点总结

Tensorflow快速入门

知识点总结

• 梯度下降方法总结
• 模型保存和恢复

前两篇博文我们实现了线性回归与手写数字识别
Tensorflow快速入门——线性回归
Tensorflw快速入门——运用逻辑回归算法手写数字识别
本文具体数据集与源代码可从我的GitHub地址获取
https://github.com/liuzuoping/Deep_Learning_note

梯度下降方法总结

• Tensorflow直接计算，满足公式：
  θ = θ − η ▽ θ M S E ( θ ) \theta = \theta- \eta \bigtriangledown_{\theta} MSE(\theta) θ=θ−η▽θ​MSE(θ)
• tf.random_uniform()：在图中创建一个包含随机值的节点，类似于NumPy中的random()函数
• tf.assign()：创建一个将新值赋给变量的一个节点，为variable更新值
  更加具体的可以回顾上篇博文
  梯度下降法详解

#--coding--:utf-8
import numpy as np
import tensorflow as tf
from sklearn.datasets import fetch_california_housing
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()

# 获取数据
housing = fetch_california_housing()
m,n = housing.data.shape
print (housing.keys())
print (housing.feature_names)
print (housing.target)
print (housing.DESCR)

housing_data_plus_bias = np.c_[np.ones((m,1)), housing.data]
scaled_data = scaler.fit_transform(housing.data)
data = np.c_[np.ones((m,1)),scaled_data]

# 设置参数
n_epoch = 1000
learning_rate = 0.01

# 设置placeholder
X = tf.constant(data,dtype = tf.float32,name = "X")
y = tf.constant(housing.target.reshape(-1,1),dtype=tf.float32,name='y')

theta = tf.Variable(tf.random_uniform([n+1, 1], -1, 1),name='theta')
y_pred = tf.matmul(X,theta,name='prediction')
error = y_pred - y
mse = tf.reduce_mean(tf.square(error),name='mse') #Mean Squared Error 

# 计算梯度公式
gradient = 2/m * tf.matmul(tf.transpose(X),error)
training_op = tf.assign(theta,theta - learning_rate * gradient)

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    for epoch in range(n_epoch):
        if epoch % 100 == 0:
            print ("Epoch",epoch, "MSE =", mse.eval())

    print ('best theta:',theta.eval())

自动求导autodiff

手动求导方法有两个缺陷：

• 深度神经网络中公式较长
• 计算效率较低
  可以采取以下命令进行自动求导：

gradients = tf.gradients(mse,[theta])[0]

gradient有两个参数：

• op，损失函数，这里是mse
• variable lists，变量列表，这里是theta值

使用Optimizer

可以使用各种优化器,如梯度下降优化器

optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
training_op = optimizer.minimize(mse)

还有很多其他优化器，例如收敛更快的MomentumOptimizer优化器等

梯度下降中传输数据的方式

• mini-batchbatch
• 方法：使用占位符placeholder

# 传统方式
A = tf.placeholder(tf.float32,shape=(None,3))
B = A + 5
with tf.Session() as sess:
    test_b_1 = B.eval(feed_dict={A:[[1,2,3]]})
    test_b_2 = B.eval(feed_dict={A:[[4,5,6],[7,8,9]]})
print(test_b_1)  
print(test_b_2) 

[[6. 7. 8.]]
[[ 9. 10. 11.]
[12. 13. 14.]]

# 使用mini-batch方式
X = tf.placeholder(tf.float32, shape=(None, n + 1), name="X")
y = tf.placeholder(tf.float32, shape=(None, 1), name="y")

# 与上文例子相同，这里采用的是使用optimizer计算梯度
n_epochs = 100
display_step = 10
learning_rate = 0.01
batch_size = 100
n_batches = int(np.ceil(m / batch_size))
theta = tf.Variable(tf.random_uniform([n + 1, 1], -1.0, 1.0, seed=42), name="theta") #X0
y_pred = tf.matmul(X, theta, name="predictions")
error = y_pred - y
mse = tf.reduce_mean(tf.square(error), name="mse")
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
training_op = optimizer.minimize(mse)

init = tf.global_variables_initializer()


# 定义mini-batch取数据方式
def fetch_batch(epoch, batch_index, batch_size):
    np.random.seed(epoch * n_batches + batch_index) 
    indices = np.random.randint(m, size=batch_size)  
    X_batch = data[indices] 
    y_batch = housing.target.reshape(-1, 1)[indices] 
    return X_batch, y_batch

# mini-batch计算过程
with tf.Session() as sess:
    sess.run(init)
    for epoch in range(n_epochs):#迭代的次数
        avg_cost = 0.
        for batch_index in range(n_batches):
            X_batch, y_batch = fetch_batch(epoch, batch_index, batch_size)
            sess.run(training_op, feed_dict={X: X_batch, y: y_batch})

            if epoch % 10 == 0:
                total_loss = 0
                acc_train = mse.eval(feed_dict={X: X_batch, y: y_batch})
                total_loss += acc_train 
                #print acc_train,total_loss
                print ( "epoch:",epoch, "batch_id:",batch_index, "Batch loss:", acc_train)

模型的保存和恢复

• 在创建图阶段创建一个Saver的节点，在执行阶段需要保存模型的地方调用Save()函数即可

# 模型的保存
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session() as sess:
    sess.run(init)
    for epoch in range(n_epochs):
        if epoch % 100 == 0:
            print("Epoch", epoch, "MSE =", mse.eval())  # 保存运行过程
            save_path = saver.save(sess, "/my_model.ckpt")
        sess.run(training_op)
    
    best_theta = theta.eval()
    save_path = saver.save(sess, "/my_model_final.ckpt")#保存最后的结果
    
# 模型的加载
with tf.Session() as sess:
    saver.restore(sess, "/tmp/my_model_final.ckpt")
    best_theta_restored = theta.eval() 
    
saver = tf.train.Saver({"weights":theta})


reset_graph()
saver = tf.train.import_meta_graph("/tmp/my_model_final.ckpt.meta")
theta = tf.get_default_graph().get_tensor_by_name("theta:0")