《机器学习》编程作业的Python实现【ex3_nn.py】

代码

import numpy as np
import matplotlib.pyplot as plt
from scipy.io import loadmat
import math


def displayData(X, *example_width):
    if example_width == ():
        example_width = round(np.sqrt(X.shape[1]))

    m, n = X.shape
    rows = math.floor(np.sqrt(m))  # 设定每行显示多少张图片
    cols = math.ceil(m / rows)    # 设定每列显示多少张图片
    fig, ax = plt.subplots(
        nrows=rows, ncols=cols, sharey=True, sharex=True, figsize=(8, 8))

    for row in range(rows):
        for column in range(cols):
            ax[row, column].matshow(
                X[rows*row+column].reshape((20, 20)), cmap='gray_r')
    plt.xticks([])
    plt.yticks([])
    plt.show()


def sigmoid(z):
    return 1 / (1 + np.exp(-z))


def predict(Theta1, Theta2, X):
    m, n = X.shape
    num_labels = Theta2.shape[0]
    p = np.zeros((m, 1))

    # Add ones to the X data matrix
    X = np.hstack((np.ones((m, 1)), X))  # 注意这里的hstack两个括号
    a2 = sigmoid(X.dot(Theta1.T))  # (m, 25)
    a2 = np.hstack((np.ones((m, 1)), a2))  # (m, 26)
    a3 = sigmoid(a2.dot(Theta2.T))  # (m, 10)
    p = np.argmax(a3, axis=1)
    return p


if __name__ == '__main__':

    # Setup the parameters you will use for this exercise

    input_layer_size = 400  # 20x20 Input Images of Digits
    hidden_layer_size = 25  # 25 hidden units
    num_labels = 10  # 10 labels, from 1 to 10
    # (note that we have mapped '0' to label '10')
    print('='*18, 'Beginning', '='*18)

# =========== Part 1: Loading and Visualizing Data =============
    # We start the exercise by first loading and visualizing the dataset.
    # You will be working with a dataset that contains handwritten digits.

    # load Training Data
    print('Loading and Visualizing Data ... ')

    file = 'ex3data1.mat'
    data = loadmat(file)  # data 是字典类型

    X = data['X']
    y = data['y']
    m, n = X.shape

    # Randomly select 100 data points to display
    index = np.random.choice(m, size=100, replace=False)
    # displayData(X[index])
    print('='*40)
# ================ Part 2: Loading Pameters ================
    # In this part of the exercise, we load some pre-initialized
    # neural network parameters.

    # Load the weights into variables Theta1 and Theta2
    print('Loading Saved Neural Network Parameters ...')
    para = loadmat('ex3weights.mat')
    Theta1 = para['Theta1']  # (25, 401)
    Theta2 = para['Theta2']  # (10, 26)

    print('='*40)

# ================= Part 3: Implement Predict =================
    # After training the neural network, we would like to use it to predict
    # the labels.
    pred = predict(Theta1, Theta2, X) + 1  # 预测出的标签值为0~9，表示1~10，因此要加1修正
    acc = np.mean(pred == y.flatten())  # y是(m,1),要展开为(m, )才能与pred比较
    print('Training Set Accuracy: ', acc)
    print('='*40)

    # To give you an idea of the network's output, you can also run
    #  through the examples one at the a time to see what it is predicting.

    # Randomly permute examples
    rp = np.random.choice(m, size=10, replace=False)
    for i in range(10):  # 测试10次
        # display
        print('Displaying Example Image')
        x = X[rp[i], :].reshape(1, n)  # (1, 400)
        pred = predict(Theta1, Theta2, x) + 1
        print('Neural Network Prediction:', np.mod(pred, 10))
        print('The Label is: ', y[rp[i]])
        fig = plt.figure()
        plt.imshow(x.reshape(20, 20), cmap='binary')
        plt.show()
        s = input('输入q退出，回车继续：')
        if s == 'q':
            break
    print('='*40)

运行结果

踩到的坑

1、在定义损失函数：lrCostFunction时，theta[0]不应参与正则化。在处理时应注意：
        theta = theta_t.copy()，此处如果不加copy()，结果就是将theta指向theta_t的内存单元，即共享同一个内存单元，如果修改theta：
        theta[0] = 0 ，则theta_t[0]也会变成0。

2、在运行oneVsAll函数时，报错：“operands could not be broadcast together with shapes (401,5000) (401,) ”
        报错位置：lrCostFunction_grad( )里的grad=...那行
        报错原因：h.shape=(5000, )
        解决办法：将h reshape为5000x1的

3、报错：“invalid gradient vector from minimized function”
        报错位置：op.minimize()函数所在行
        报错原因：在运行oneVsAll( ) 函数时，传入到 lrCostFunction( )的 theta_t 的shape发生变化，变为(401, )，应为(401, 1)
        解决办法：将theta_t reshape为(401, 1) ：theta_t = theta_t.reshape(theta_t.shape[0], 1)

        应加深对numpy中参数广播的理解。

4、在执行优化op.minimize()时，由于是分别训练10个分类器的参数theta，在每次训练时忘记处理标签，即未把需要训练的分类器的标签置为1，其余类别置为0
        导致训练出来的参数全为0。
        method='TNC' 指的是用牛顿截断(truncated Newton)法优化
        https://docs.scipy.org/doc/scipy/reference/optimize.minimize-tnc.html#optimize-minimize-tnc