吴恩达深度学习笔记一：神经网络和深度学习

1. 梯度下降中的一些计算

     第一个框：logistic回归方程的代价函数的导数dj/dz 可认为是a-y
     第二个框：前一级变量的导数为后一级变量导数乘以其系数。
     另外求dZ和dA不需要进行 ”/m” 操作

# dA/dA_pre = (dA/dZ * dZ/dA_pre) = (dA/dZ * w), 为了表示方便去掉了"dA/", 故乘法不变
    dA_pre = np.dot(W.T, dZ)

     或者说在起前面的变量通过其系数来“放大”对最终结果的影响。例如：y=2a + b, 令v=2a+b, 则 y=v, dy/dv=1 , dy/da = (dy/dv) * (dv/da) = 1*2 = 2

     有关维度的计算，假设有[5,3,2,1]的网络,m个样本则：
W1[3,5], W2[2,3], W3[1,2]; Z1=A1=[3,m], Z2=A2=[2,m], Z3=A3=[1,m], 所以：

$$dZ3 = A3 - Y -> [1, m]$$ $$dW3[1,2] = dZ3[1, m] * A2[2,m].T$$ $$dA2[2,m] = W3[1,2].T * dZ3[1, m]$$
可以看出dZ的位置及转置情况有所不同，因为1、W的行列分别为本层的单元数和上一层的单元数且与m无关，而A的行为上一层的单元数，故求dW时转置取A行为列，其列消去。2、A的列和Z的列总为m, 故求dA时Z不变取其列。3、 总结为：Z为纽带不转置

2. 激活函数的选择

   sigmoid函数一般只用于二分类的输出层，tanh函数类似，其缺点在于当z的绝对值太大时导数很小，训练的过程就会很慢
   ReLU函数是在神经网络中应用最多的激活函数，其优点在于当z为正值时导数值一般较大，训练过程较快，而在实际中z处于负值的情况很少（此时函数不可导）。

3. 损失函数和准确度的计算

一般来说，损失函数是对单个样本而言，使用 L 表示，代价（成本）函数是损失函数在整个样本空间的扩展，用 J 表示，但为了方便也可以使用损失函数来表达代价函数的意思。

   交叉熵损失函数：

$J(θ)=-\frac{1}{m} ∑_{i=1}^{m}[y(i)log(hθ(x(i)))+(1−y(i))log(1−hθ(x(i))]$

m = Y.shape[1]
cost = -1 / m * np.sum(np.dot(Y, np.log(AL).T) + np.dot(1 - Y, np.log(1 - AL).T))
cost = np.squeeze(cost)

   准确率，AL表示最后一层的输出：

acc = (np.dot(y_train, AL.T) + np.dot(1 - y_train, 1 - AL.T)) / y_train.shape[1] * 100

4. 神经网络模型的建立

   代码示例：

def initialize_deep_parameters(layer_dims=np.array([5, 4, 1])):
    """
    初始化任意层的神经网络参数
    :param layer_dims: 记录每层单元数的数组,包括输入输出，如[2,4,1]则有输入层2，W1[4,2],b1[4,1], W2[1,4].b2[1,1]
                        最后一个1表示只有一个类别，即二分类识别
    :return:
    """
    parameters = {}
    num_layers = len(layer_dims)
    for l in range(1, num_layers):
        parameters['W' + str(l)] = np.random.rand(layer_dims[l], layer_dims[l - 1]) * 0.01
        parameters['b' + str(l)] = np.zeros((layer_dims[l], 1))

    # 测试
    # for key in parameters.keys():
    #     print(key, parameters[key].shape)

    return parameters


def linear_forward(A, W, b):
    """
    一层神经网络中的线性计算
    :param A: 上一层激活函数的输出,第一层即为X
    :param W:
    :param b: 
    :return: cache,其实就是Z或A的复制值，用了计算反向传播，Z和A变量进行正向传播
    """
    Z = np.dot(W, A) + b
    assert (Z.shape == (W.shape[0], A.shape[1]))
    cache = (A, W, b)
    return Z, cache


def linear_activation_forward(A_pre, W, b, activation='relu'):
    """
    激活函数
    :param A_pre: 上一层的输出
    :param W:
    :param b:
    :param activation:
    :return:
    """
    Z, linear_cache = linear_forward(A_pre, W, b)

    if activation == 'sigmoid':
        A, activation_cache = dnn_utils_v2.sigmoid(Z)
        assert (A.shape == (W.shape[0], A_pre.shape[1]))
        cache = (linear_cache, activation_cache)
        return A, cache

    elif activation == 'relu':
        A, activation_cache = dnn_utils_v2.relu(Z)
        assert (A.shape == (W.shape[0], A_pre.shape[1]))
        cache = (linear_cache, activation_cache)
        return A, cache


def L_model_forward(X, parameters):
    """
    实现L层模型的前向传播
    :param X:
    :param parameters:
    :return:
    """
    caches = []
    A = X
    L = len(parameters) // 2  # 根据参数W或b的个数确定层数，一层两个参数，Wl.bl
    for l in range(1, L):
        A_pre = A
        A, cache = linear_activation_forward(A_pre, parameters['W' + str(l)], parameters['b' + str(l)],
                                             activation='relu')
        caches.append(cache)
    # 最后一层
    AL, cache = linear_activation_forward(A, parameters['W' + str(L)], parameters['b' + str(L)], activation='sigmoid')
    caches.append(cache)

    assert (AL.shape == (1, X.shape[1]))
    return AL, caches


def compute_cost(AL, Y):
    """
    代价函数计算
    :param AL: L层激活函数的输出即最好的输出
    :param Y:
    :return:
    """
    m = Y.shape[1]
    cost = -1 / m * np.sum(np.dot(Y, np.log(AL).T) + np.dot(1 - Y, np.log(1 - AL).T))
    cost = np.squeeze(cost)
    return cost


def linear_backward(dZ, cache):
    """
    反向传播计算梯度
    :param dZ: 当前层损失函数的导数
    :param cache:
    :return:
    """
    A_pre, W, b = cache
    m = A_pre.shape[1]

    dW = np.dot(dZ, A_pre.T) / m
    db = np.sum(dZ, axis=1, keepdims=True) / m
    # dA/dA_pre = (dA/dZ * dZ/dA_pre) = (dA/dZ * w), 为了表示方便去掉了"dA/", 故乘法不变
    dA_pre = np.dot(W.T, dZ)
    return dA_pre, dW, db


def linear_activation_backward(dA, cache, activation='relu'):
    """
    激活函数部分梯度计算
    :param dA:
    :param cache:
    :param activation:
    :return:
    """
    linear_cache, activation_cache = cache
    if activation == 'relu':
        dZ = dnn_utils_v2.relu_backward(dA, activation_cache)
        dA_pre, dW, db = linear_backward(dZ, linear_cache)
        return dA_pre, dW, db
    elif activation == 'sigmoid':
        dZ = dnn_utils_v2.sigmoid_backward(dA, activation_cache)
        dA_pre, dW, db = linear_backward(dZ, linear_cache)
        return dA_pre, dW, db


def L_model_backward(AL, Y, caches):
    """
    反向传播整合
    :param AL:
    :param Y:
    :param caches:
    :return:
    """
    grads = {}
    L = len(caches)  # 网络层数
    m = AL.shape[1]
    Y = Y.reshape(AL.shape)

    dAL = - (np.divide(Y, AL) - np.divide(1 - Y, 1 - AL))
    current_cache = caches[L - 1]
    grads['dA' + str(L)], grads['dW' + str(L)], grads['db' + str(L)] = linear_activation_backward(
        dAL, current_cache, activation='sigmoid')

    for l in reversed(range(L - 1)):
        current_cache = caches[l]
        dA_pre, dW, db = linear_activation_backward(grads['dA' + str(l + 2)], current_cache, activation='relu')
        grads['dA' + str(l + 1)] = dA_pre
        grads['dW' + str(l + 1)] = dW
        grads['db' + str(l + 1)] = db

    return grads


def update_parameters(parametes, grads, learning_rate=0.1):
    """
    更新参数
    :param parametes:
    :param grads:
    :param learning_rate:
    :return:
    """
    L = len(parametes) // 2  # 一层有w和b两个参数
    for l in range(L):
        parametes['W' + str(l + 1)] = parametes['W' + str(l + 1)] - learning_rate * grads['dW' + str(l + 1)]
        parametes['b' + str(l + 1)] = parametes['b' + str(l + 1)] - learning_rate * grads['db' + str(l + 1)]
    return parametes


if __name__ == '__main__':
    np.random.rand(1)

    # 线性单元测试
    # A, W, b = testCases_v2.linear_forward_test_case()
    # Z, linear_cache = linear_forward(A, W, b)
    # print(Z)

    # 激活函数测试
    # A_pre, W, b = testCases_v2.linear_activation_forward_test_case()
    # A, cache = linear_activation_forward(A_pre, W, b, activation='sigmoid')
    # print('sigmoid： ' + str(A))
    # A, cache = linear_activation_forward(A_pre, W, b, activation='relu')
    # print('relu: ' + str(A))

    # 正向传播测试
    # X, parameters = testCases_v2.L_model_forward_test_case()
    # AL, caches = L_mode_forward(X, parameters)
    # print(str(AL), str(len(caches)))

    # # 损失函数测试
    # Y, AL = testCases_v2.compute_cost_test_case()
    # print(compute_cost(AL, Y))

    # 梯度测试
    # dZ, linear_cache = testCases_v2.linear_backward_test_case()
    # print(linear_backward(dZ, linear_cache))
    # AL, cache = testCases_v2.linear_activation_backward_test_case()
    # print(linear_activation_backward(AL, cache, activation='sigmoid'))
    # print(linear_activation_backward(AL, cache, activation='relu'))

    # 完整的反向出传播梯度测试
    # AL, Y, caches = testCases_v2.L_model_backward_test_case()
    # grads = L_mode_backward(AL, Y, caches)
    # print(grads)

    # 更新参数
    parameters, grads = testCases_v2.update_parameters_test_case()
    parameters = update_parameters(parameters, grads, learning_rate=0.1)
    print(parameters)

   使用：

def load_data():
    train_dataset = h5py.File('datasets/train_catvnoncat.h5', "r")
    train_set_x_orig = np.array(train_dataset["train_set_x"][:])  # your train set features
    train_set_y_orig = np.array(train_dataset["train_set_y"][:])  # your train set labels

    test_dataset = h5py.File('datasets/test_catvnoncat.h5', "r")
    test_set_x_orig = np.array(test_dataset["test_set_x"][:])  # your test set features
    test_set_y_orig = np.array(test_dataset["test_set_y"][:])  # your test set labels

    classes = np.array(test_dataset["list_classes"][:])  # the list of classes

    train_set_y_orig = train_set_y_orig.reshape((1, train_set_y_orig.shape[0]))
    test_set_y_orig = test_set_y_orig.reshape((1, test_set_y_orig.shape[0]))

    return train_set_x_orig, train_set_y_orig, test_set_x_orig, test_set_y_orig, classes


def L_layer_model_train(x_train, y_train, layers_dims, train_num=2000, learning_rate=0.02, print_cost=True):
    """
    多层模型
    :param x_train:
    :param y_train:
    :param layers_dims: 整个网络模型的维度数组
    :param train_num:
    :param learning_rate:
    :param print_cost:
    :return:
    """
    parameters = dnn_app_utils_v2.initialize_parameters_deep(layers_dims)
    costs = []
    for i in range(1, train_num + 1):
        AL, cache = deep_neural_network.L_model_forward(x_train, parameters)
        cost = deep_neural_network.compute_cost(AL, y_train)
        grads = deep_neural_network.L_model_backward(AL, y_train, cache)
        parameters = deep_neural_network.update_parameters(parameters, grads, learning_rate)

        if print_cost and i % 100 == 0:
            costs.append(cost)
            acc = (np.dot(y_train, AL.T) + np.dot(1 - y_train, 1 - AL.T)) / y_train.shape[1] * 100
            print("num {} cost is: {} and acc is : {}".format(i, cost, np.squeeze(acc)))

    if print_cost:
        plt.plot(np.squeeze(costs))
        plt.ylabel('cost')
        plt.xlabel('iterations (per tens)')
        plt.title("Learning rate = " + str(learning_rate))
        plt.show()

    return parameters


def L_layers_model_test(x_test, y_test, parameters):
    """
    多层模型的检验
    :param x_test:
    :param y_test:
    :param parameters:
    :return:
    """
    y_pre, cache = deep_neural_network.L_model_forward(x_test, parameters)
    cost = deep_neural_network.compute_cost(y_pre, y_test)
    acc = (np.dot(y_test,y_pre.T) + np.dot(1-y_test, 1-y_pre.T)) / y_test.shape[1] * 100
    print(cost, np.squeeze(acc))
    return cost, acc


if __name__ == '__main__':
    np.random.rand(1)

    # 从文件中读取数据
    train_x_orig, train_y, test_x_orig, test_y, classes = deep_neural_network.load_data()
    print(train_x_orig.shape, test_x_orig.shape)

    # 将 [m,px,px,3] 转为 [px *  px * 3, m]， 并归一化
    x_train = train_x_orig.reshape(train_x_orig.shape[0], -1).T
    # 将类数目作为输入的列数，可保持不变
    x_test = test_x_orig.reshape(test_x_orig.shape[0], -1).T
    x_train = x_train / 255
    x_test = x_test / 255
    print(x_train.shape)

    # 两层模型
    # parameters = two_layer_model_train(x_train, train_y)
    # two_layer_model_test(x_train, train_y, parameters)
    # two_layer_model_test(x_test, test_y, parameters)

    # 任意层模型
    layer_dims = np.array([12288, 20, 10, 5, 1])
    parameters = L_layer_model_train(x_train, train_y, layer_dims, learning_rate=0.01)
    L_layers_model_test(x_test, test_y, parameters)