【连载】深度学习笔记12:卷积神经网络的Tensorflow实现AI学院

来源: / 作者:塔卡拉 / 2018-10-30 17:24
本节就继续和大家一起学习如何利用 Tensorflow 搭建一个卷积神经网络。

      在上一讲中,我们学习了如何利用 numpy 手动搭建卷积神经网络。但在实际的图像识别中,使用 numpy 去手写 CNN 未免有些吃力不讨好。在 DNN 的学习中,我们也是在手动搭建之后利用 Tensorflow 去重新实现一遍,一来为了能够对神经网络的传播机制能够理解更加透彻,二来也是为了更加高效使用开源框架快速搭建起深度学习项目。本节就继续和大家一起学习如何利用 Tensorflow 搭建一个卷积神经网络。

      我们继续以 NG 课题组提供的 sign 手势数据集为例,学习如何通过 Tensorflow 快速搭建起一个深度学习项目。数据集标签共有零到五总共 6 类标签,示例如下:

      先对数据进行简单的预处理并查看训练集和测试集维度:
 

X_train = X_train_orig/255.
X_test = X_test_orig/255.
Y_train = convert_to_one_hot(Y_train_orig, 6).T Y_test = convert_to_one_hot(Y_test_orig, 6).T
print ("number of training examples = " + str(X_train.shape[0]))
print ("number of test examples = " + str(X_test.shape[0]))
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))


      可见我们总共有 1080 张 64643 训练集图像,120 张 64643 的测试集图像,共有 6 类标签。下面我们开始搭建过程。

创建 placeholder

      首先需要为训练集预测变量和目标变量创建占位符变量 placeholder ,定义创建占位符变量函数:

def create_placeholders(n_H0, n_W0, n_C0, n_y):    
    """
    Creates the placeholders for the tensorflow session.
 
    Arguments:
    n_H0 -- scalar, height of an input image
    n_W0 -- scalar, width of an input image
    n_C0 -- scalar, number of channels of the input
    n_y -- scalar, number of classes
 
    Returns:
    X -- placeholder for the data input, of shape [None, n_H0, n_W0, n_C0] and dtype "float"
    Y -- placeholder for the input labels, of shape [None, n_y] and dtype "float"
    """
    X = tf.placeholder(tf.float32, shape=(None, n_H0, n_W0, n_C0), name='X')
    Y = tf.placeholder(tf.float32, shape=(None, n_y), name='Y')    
    return X, Y
参数初始化

      然后需要对滤波器权值参数进行初始化:

def initialize_parameters():    
    """
    Initializes weight parameters to build a neural network with tensorflow.
    Returns:
    parameters -- a dictionary of tensors containing W1, W2
    """
 
    tf.set_random_seed(1)                            
 
    W1 = tf.get_variable("W1", [4,4,3,8], initializer = tf.contrib.layers.xavier_initializer(seed = 0))
    W2 = tf.get_variable("W2", [2,2,8,16], initializer = tf.contrib.layers.xavier_initializer(seed = 0))
 
    parameters = {"W1": W1,                  
                  "W2": W2}    
    return parameters

执行卷积网络的前向传播过程

      前向传播过程如下所示:
CONV2D -> RELU -> MAXPOOL -> CONV2D -> RELU -> MAXPOOL -> FLATTEN -> FULLYCONNECTED


      可见我们要搭建的是一个典型的 CNN 过程,经过两次的卷积-relu激活-最大池化,然后展开接上一个全连接层。利用 Tensorflow  搭建上述传播过程如下:

def forward_propagation(X, parameters):    
    """
    Implements the forward propagation for the model
 
    Arguments:
    X -- input dataset placeholder, of shape (input size, number of examples)
    parameters -- python dictionary containing your parameters "W1", "W2"
                  the shapes are given in initialize_parameters
 
    Returns:
    Z3 -- the output of the last LINEAR unit
    """
 
    # Retrieve the parameters from the dictionary "parameters"
    W1 = parameters['W1']
    W2 = parameters['W2']    
    # CONV2D: stride of 1, padding 'SAME'
    Z1 = tf.nn.conv2d(X,W1, strides = [1,1,1,1], padding = 'SAME')    
    # RELU
    A1 = tf.nn.relu(Z1)    
    # MAXPOOL: window 8x8, sride 8, padding 'SAME'
    P1 = tf.nn.max_pool(A1, ksize = [1,8,8,1], strides = [1,8,8,1], padding = 'SAME')    
    # CONV2D: filters W2, stride 1, padding 'SAME'
    Z2 = tf.nn.conv2d(P1,W2, strides = [1,1,1,1], padding = 'SAME')    
    # RELU
    A2 = tf.nn.relu(Z2)  
    # MAXPOOL: window 4x4, stride 4, padding 'SAME'
    P2 = tf.nn.max_pool(A2, ksize = [1,4,4,1], strides = [1,4,4,1], padding = 'SAME')    
    # FLATTEN
    P2 = tf.contrib.layers.flatten(P2)
 
    Z3 = tf.contrib.layers.fully_connected(P2, 6, activation_fn = None)    
    return Z3

计算当前损失

      在 Tensorflow  中计算损失函数非常简单,一行代码即可:

def compute_cost(Z3, Y):    
    """
    Computes the cost
    Arguments:
    Z3 -- output of forward propagation (output of the last LINEAR unit), of shape (6, number of examples)
    Y -- "true" labels vector placeholder, same shape as Z3
 
    Returns:
    cost - Tensor of the cost function
    """
 
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=Z3, labels=Y))    
    return cost

      定义好上述过程之后,就可以封装整体的训练过程模型。可能你会问为什么没有反向传播,这里需要注意的是 Tensorflow 帮助我们自动封装好了反向传播过程,无需我们再次定义,在实际搭建过程中我们只需将前向传播的网络结构定义清楚即可。

封装模型

def model(X_train, Y_train, X_test, Y_test, learning_rate = 0.009,
          num_epochs = 100, minibatch_size = 64, print_cost = True):    
    """
    Implements a three-layer ConvNet in Tensorflow:
    CONV2D -> RELU -> MAXPOOL -> CONV2D -> RELU -> MAXPOOL -> FLATTEN -> FULLYCONNECTED
 
    Arguments:
    X_train -- training set, of shape (None, 64, 64, 3)
    Y_train -- test set, of shape (None, n_y = 6)
    X_test -- training set, of shape (None, 64, 64, 3)
    Y_test -- test set, of shape (None, n_y = 6)
    learning_rate -- learning rate of the optimization
    num_epochs -- number of epochs of the optimization loop
    minibatch_size -- size of a minibatch
    print_cost -- True to print the cost every 100 epochs
 
    Returns:
    train_accuracy -- real number, accuracy on the train set (X_train)
    test_accuracy -- real number, testing accuracy on the test set (X_test)
    parameters -- parameters learnt by the model. They can then be used to predict.
    """
 
    ops.reset_default_graph()                        
    tf.set_random_seed(1)                            
    seed = 3                                        
    (m, n_H0, n_W0, n_C0) = X_train.shape            
    n_y = Y_train.shape[1]                            
    costs = []                                      
 
    # Create Placeholders of the correct shape
    X, Y = create_placeholders(n_H0, n_W0, n_C0, n_y)  
    # Initialize parameters
    parameters = initialize_parameters()    
    # Forward propagation
    Z3 = forward_propagation(X, parameters)    
    # Cost function
    cost = compute_cost(Z3, Y)    
    # Backpropagation
    optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost)    # Initialize all the variables globally
    init = tf.global_variables_initializer()    
    # Start the session to compute the tensorflow graph
    with tf.Session() as sess:        
        # Run the initialization
        sess.run(init)        
        # Do the training loop
        for epoch in range(num_epochs):
 
            minibatch_cost = 0.
            num_minibatches = int(m / minibatch_size)
            seed = seed + 1
            minibatches = random_mini_batches(X_train, Y_train, minibatch_size, seed)            
            for minibatch in minibatches:                
                # Select a minibatch
                (minibatch_X, minibatch_Y) = minibatch
                _ , temp_cost = sess.run([optimizer, cost], feed_dict={X: minibatch_X, Y: minibatch_Y})
                minibatch_cost += temp_cost / num_minibatches            
                # Print the cost every epoch
            if print_cost == True and epoch % 5 == 0:              
                print ("Cost after epoch %i: %f" % (epoch, minibatch_cost))            
            if print_cost == True and epoch % 1 == 0:
                costs.append(minibatch_cost)        
        # plot the cost
        plt.plot(np.squeeze(costs))
        plt.ylabel('cost')
        plt.xlabel('iterations (per tens)')
        plt.title("Learning rate =" + str(learning_rate))
        plt.show()        # Calculate the correct predictions
        predict_op = tf.argmax(Z3, 1)
        correct_prediction = tf.equal(predict_op, tf.argmax(Y, 1))        
        # Calculate accuracy on the test set
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
        print(accuracy)
        train_accuracy = accuracy.eval({X: X_train, Y: Y_train})
        test_accuracy = accuracy.eval({X: X_test, Y: Y_test})
        print("Train Accuracy:", train_accuracy)
        print("Test Accuracy:", test_accuracy)      
         
        return train_accuracy, test_accuracy, parameters

     对训练集执行模型训练:
_, _, parameters = model(X_train, Y_train, X_test, Y_test)

     训练迭代过程如下:


    我们在训练集上取得了 0.67 的准确率,在测试集上的预测准确率为 0.58 ,虽然效果并不显著,模型也有待深度调优,但我们已经学会了如何用 Tensorflow  快速搭建起一个深度学习系统了
 

阅读延展

1
3