深度学习笔记12：卷积神经网络的Tensorflow实现

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。

  注：本深度学习笔记系作者学习 Andrew NG 的 deeplearningai 五门课程所记笔记，其中代码为每门课的课后assignments作业整理而成。

参考资料：

https://www.coursera.org/learn/machine-learning

https://www.deeplearning.ai/