吴恩达DeepLearning第二部分作业week1（3）Grad-check梯度检查

开始完全整蒙了，还是半夜头脑清醒了。按着公式一点点打出来了，由于网上也没找到对照的也不知道对不对。

首先搬来自己写的多层神经网络

import scipy.io as scio
import numpy as np
import matplotlib.pyplot as plt


def init(x):
    first_num = 10
    second_num = 7
    third_num = 3
    np.random.seed(5)
    w1 = np.random.randn(first_num, x.shape[0]) * np.sqrt(2 / x.shape[0])
    b1 = np.zeros((first_num, 1))
    w2 = np.random.randn(second_num, first_num) * np.sqrt(2 / first_num)
    b2 = np.zeros((second_num, 1))
    w3 = np.random.randn(third_num, second_num) * np.sqrt(2 / second_num)
    b3 = np.zeros((third_num, 1))
    w4 = np.random.randn(1, third_num) * np.sqrt(2 / third_num)
    b4 = np.zeros((1, 1))
    ini_param = {
        "w1": w1,
        "b1": b1,
        "w2": w2,
        "b2": b2,
        "w3": w3,
        "b3": b3,
        "w4": w4,
        "b4": b4
    }
    return ini_param


# 构建前向
def cal_z(w, a, b):  # a维度：上层节点数*例子数 w维度：该层节点数*上层节点数
    return np.dot(w, a) + b


def cal_sigma(z):  # z维度：该层节点数*例子数
    return 1 / (1 + np.exp(-z))


def cal_tan_h(z):  # z1维度hidden_num*x.shape[1]
    return np.tanh(z)


def cal_relu(z):  # z维度：该层节点数*例子数
    return np.maximum(0, z)


def forward_f(x, p):
    z1 = cal_z(p["w1"], x, p["b1"])
    a1 = cal_relu(z1)
    z2 = cal_z(p["w2"], a1, p["b2"])
    a2 = cal_relu(z2)
    z3 = cal_z(p["w3"], a2, p["b3"])
    a3 = cal_relu(z3)
    z4 = cal_z(p["w4"], a3, p["b4"])
    a4 = cal_sigma(z4)
    forward_param = {
        "z1": z1,
        "a1": a1,
        "z2": z2,
        "a2": a2,
        "z3": z3,
        "a3": a3,
        "z4": z4,
        "a4": a4
    }
    return forward_param


# 计算损失函数
def cost_f(a, y):  # a维度: 1*例子数 y维度：1*例子数
    m = y.shape[1]
    return -np.sum(y * np.log(a) + (1 - y) * np.log(1 - a)) / m


# 后向
def cal_dz_last(a, y):  # ai维度：第i层节点数*例子数 y维度：1*例子数
    return a - y


def cal_dw_db(dz, a, m):  # dzi维度：第i层节点数*例子数 ai维度：第i层节点数*例子数
    return np.dot(dz, a.T) / m, np.sum(dz, axis=1, keepdims=True) / m


def cal_da(dz, w):  # dzi维度：第i层节点数*例子数 wi维度：第i层节点数*第i-1层节点数
    return np.dot(w.T, dz)


def cal_drelu(da, z):  # dai维度：第i层节点数*例子数 zi维度：第i层节点数*例子数
    t = np.ones(z.shape)
    t[z <= 0] = 0
    return da * t


def back_f(p, f_p, x, y):
    dz4 = cal_dz_last(f_p["a4"], y)
    dw4, db4 = cal_dw_db(dz4, f_p["a3"], y.shape[1])

    da3 = cal_da(dz4, p["w4"])
    dz3 = cal_drelu(da3, f_p["z3"])
    dw3, db3 = cal_dw_db(dz3, f_p["a2"], y.shape[1])

    da2 = cal_da(dz3, p["w3"])
    dz2 = cal_drelu(da2, f_p["z2"])
    dw2, db2 = cal_dw_db(dz2, f_p["a1"], y.shape[1])

    da1 = cal_da(dz2, p["w2"])
    dz1 = cal_drelu(da1, f_p["z1"])
    dw1, db1 = cal_dw_db(dz1, x, y.shape[1])

    back_param = {
        "dw1": dw1,
        "db1": db1,
        "dw2": dw2,
        "db2": db2,
        "dw3": dw3,
        "db3": db3,
        "dw4": dw4,
        "db4": db4
    }
    return back_param


# 更新参数
def update_p(p, b_p, learning_rate):
    upd_p = {
        "w1": p["w1"] - learning_rate * b_p["dw1"],
        "b1": p["b1"] - learning_rate * b_p["db1"],
        "w2": p["w2"] - learning_rate * b_p["dw2"],
        "b2": p["b2"] - learning_rate * b_p["db2"],
        "w3": p["w3"] - learning_rate * b_p["dw3"],
        "b3": p["b3"] - learning_rate * b_p["db3"],
        "w4": p["w4"] - learning_rate * b_p["dw4"],
        "b4": p["b4"] - learning_rate * b_p["db4"]
    }
    return upd_p

下面我按照梯度检查公式写的，想了一下午用向量化实现计算，以失败告终。只能以最笨的循环实现：

# 梯度检查
def cal_dcost(x, y, p, theta, str):
    dt = []
    for i in range(p[str].shape[0]):
        for j in range(p[str].shape[1]):
            p[str][i, j] = p[str][i, j] + theta
            p_plust = forward_f(x, p)
            p[str][i, j] = p[str][i, j] - 2 * theta
            p_subt = forward_f(x, p)
            p[str][i, j] = p[str][i, j] + theta
            dapprox = (cost_f(p_plust["a4"], y) - cost_f(p_subt["a4"], y)) / (2 * theta)
            # if dapprox==0:
            # print(str,i,j)
            dt.append(dapprox)
    return dt


def grad_check(x, y, p, b_p, theta=0.0000001):
    theta_arr = np.ones((1, 1))
    for i in p.keys():
        t = cal_dcost(x, y, p, theta, i)
        t = np.array(t)
        t = t.reshape(-1, 1)
        theta_arr = np.append(theta_arr, t, axis=0)
    dtheta_arr = np.ones((1, 1))
    for i in b_p.keys():
        t = np.copy(b_p[i])
        t = t.reshape(-1, 1)
        dtheta_arr = np.append(dtheta_arr, t, axis=0)
    return np.linalg.norm(theta_arr - dtheta_arr, axis=1) / (
                np.linalg.norm(theta_arr, axis=1) + np.linalg.norm(dtheta_arr, axis=1))

建模，绘图

# 建模
def model(x, y, learning_rate, loop_num):
    p = init(x)
    cost_t = []
    for i in range(loop_num):
        f_p = forward_f(x, p)
        b_p = back_f(p, f_p, x, y)
        p = update_p(p, b_p, learning_rate)
        print(grad_check(x, y, p, b_p))
        if i % 1 == 0:
            cost_t.append(cost_f(f_p["a4"], y))
    return p, cost_t


# 至此模型搭建完成


def print_figure(x, y, final_p, _cost):
    x_min, x_max = x[0, :].min() - 1, x[0, :].max() + 1
    y_min, y_max = x[1, :].min() - 1, x[1, :].max() + 1
    h = 0.01
    # Generate a grid of points with distance h between them
    xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
    # Predict the function value for the whole grid
    z = forward_f(np.vstack((xx.ravel(), yy.ravel())), final_p)
    z = np.round(z["a4"]).reshape(xx.shape)
    # Plot the contour and training examples
    plt.contourf(xx, yy, z, cmap=plt.cm.Spectral)
    plt.ylabel('x2')
    plt.xlabel('x1')
    plt.scatter(x[0, :], x[1, :], c=y, cmap=plt.cm.Spectral)
    plt.figure()
    plt.plot(_cost)
    plt.show()

这边读取mat数据作业里没给，需要自己读一下

data_path = r".\datasets\data.mat"
data = scio.loadmat(data_path)
print(data.keys())
x_train = data.get('X')
y_train = data.get('y')
x_test = data.get("Xval")
y_test = data.get("yval")
print(x_train.shape)
print(y_train.shape)
model_p, cost = model(x_train.T, y_train.T, 0.001, 1)
forward_f(x_train.T, model_p)
print_figure(x_train.T, y_train.T, model_p, cost)

结果有报错，RuntimeWarning: invalid value encountered in true_divide
  np.linalg.norm(theta_arr, axis=1) + np.linalg.norm(dtheta_arr, axis=1))

意思除数有0，我检查了这些数，两个导数确实都是0。这几个节点是不是在梯度下降过程中没有发挥作用。

老吴在课上说10^-7是完美，10^-4~在10^-5也没有大问题，我就算勉强过关了，不想在烧脑了，计算机真是让人头秃。