李宏毅ML2020作业2（源代码在最后）

李宏毅ML2020作业2（源代码在最后）

kaggle 的online judge需要科学上网，链接如下：

ML2020spring - hw2 | Kaggle

另附一个在作业时看到的很好的文章：

机器学习之逻辑斯蒂回归_梅菜扣肉-CSDN博客

2020李宏毅机器学习课程作业——Homework2：classification（Logistic Regression）_梅菜扣肉-CSDN博客

文章中提到;

判别式分类器才是首选。当训练集增加，尽管两种算法都会表现更好，但效果好的表现形式却是不同的。从反复实验中观察到，训练集增加后，判别式方法会达到更低的渐进误差，而生成式方法会更快的达到渐进误差，但这个误差比判别方法的渐进误差大。

而从我们上述推到过程中也可以发现，生成模型我们进行了诸多假设，这些假设可以使噪声的影响被抵消，所以生成式模型受样本数据的影响较小，而判别式模型受数据和噪声影响更大。

助教给的原始源代码scores

修改代码

HW2

1.将优化方式改为Adagrad

Deep Learning 最优化方法之AdaGrad_BVL的博客-CSDN博客_adagrad

下图为助教给出的优化方式： 

# Calcuate the number of parameter updates
step = 1

# Iterative training
for epoch in range(max_iter):
    # Random shuffle at the begging of each epoch
    X_train, Y_train = _shuffle(X_train, Y_train)
        
    # Mini-batch training
    for idx in range(int(np.floor(train_size / batch_size))):
        X = X_train[idx*batch_size:(idx+1)*batch_size]
        Y = Y_train[idx*batch_size:(idx+1)*batch_size]

        # Compute the gradient
        w_grad, b_grad = _gradient(X, Y, w, b)
            
        # gradient descent update
        # learning rate decay with time
        w = w - learning_rate/np.sqrt(step) * w_grad
        b = b - learning_rate/np.sqrt(step) * b_grad

        step = step + 1

 用adagrad优化的代码

# adagrad所需的累加和
adagrad_w = 0
adagrad_b = 0
# 防止adagrad除零
eps = 1e-8
 
# 迭代训练
for epoch in range(max_iter):
    # 在每个epoch开始时，随机打散训练数据
    X_train, Y_train = _shuffle(X_train, Y_train)
 
    # Mini-batch训练
    for idx in range(int(np.floor(train_size / batch_size))):
        X = X_train[idx * batch_size:(idx + 1) * batch_size]
        Y = Y_train[idx * batch_size:(idx + 1) * batch_size]
 
        # 计算梯度
        w_grad, b_grad = _gradient(X, Y, w, b)
        adagrad_w += w_grad**2
        adagrad_b += b_grad**2
 
        # 梯度下降法adagrad更新w和b
        w = w - learning_rate / (np.sqrt(adagrad_w + eps)) * w_grad
        b = b - learning_rate / (np.sqrt(adagrad_b + eps)) * b_grad

adagrad更新公式：

2.添加二次特征项来增加拟合效果

（举个例子，淘宝上有卖纸箱子的，各种型号，价格不等：

给你一组训练数据（a=长，b=宽，c=高）和他们对应的价格，让你训练出一个模型来预测箱子的价格。（假如卖家是根据箱子的体积来给箱子定价的）

这时候如果你直接用a,b,c的一次项来做线性回归，效果肯定不好；但如果你增加一个新变量d=a*b*c，也就是箱子的体积，再做线性回归，你会发现拟合效果会非常好。当箱子都是立方体的时候（a=b=c），那就相当于增加一个变量d=a^3表示体积来拟合（也就是你说的三次方，本质都是多项式）。

至于什么时候要用变量的多项式来拟合，要看具体问题和你的先验知识了，有时候需要不断的尝试。
链接：https://www.zhihu.com/question/264245010/answer/278623526
来源：知乎）

我把加入的特征平方项改为了 立方项，结果提升甚微 private score略有提高，public score略有下降。

下图为结果。

最后调参

# Some parameters for training    
max_iter = 272
batch_size = 128
learning_rate = 0.1

优化明显

也是打到了public strong baseline（0.89052）

 

最后附上源代码：

为方便查找、优化代码，采取jupyter notebook的段式结构

import numpy as np

np.random.seed(0)
X_train_fpath = '../input/ml2020spring-hw2/data/X_train'
Y_train_fpath = '../input/ml2020spring-hw2/data/Y_train'
X_test_fpath = '../input/ml2020spring-hw2/data/X_test'
output_fpath = './output_{}.csv'

# Parse csv files to numpy array
with open(X_train_fpath) as f:
    next(f)
    X_train = np.array([line.strip('n').split(',')[1:] for line in f], dtype = float)
with open(Y_train_fpath) as f:
    next(f)
    Y_train = np.array([line.strip('n').split(',')[1] for line in f], dtype = float)
with open(X_test_fpath) as f:
    next(f)
    X_test = np.array([line.strip('n').split(',')[1:] for line in f], dtype = float)

def _normalize(X, train = True, specified_column = None, X_mean = None, X_std = None):
    # This function normalizes specific columns of X.
    # The mean and standard variance of training data will be reused when processing testing data.
    #
    # Arguments:
    #     X: data to be processed
    #     train: 'True' when processing training data, 'False' for testing data
    #     specific_column: indexes of the columns that will be normalized. If 'None', all columns
    #         will be normalized.
    #     X_mean: mean value of training data, used when train = 'False'
    #     X_std: standard deviation of training data, used when train = 'False'
    # Outputs:
    #     X: normalized data
    #     X_mean: computed mean value of training data
    #     X_std: computed standard deviation of training data

    if specified_column == None:
        specified_column = np.arange(X.shape[1])
    if train:
        X_mean = np.mean(X[:, specified_column] ,0).reshape(1, -1)
        X_std  = np.std(X[:, specified_column], 0).reshape(1, -1)

    X[:,specified_column] = (X[:, specified_column] - X_mean) / (X_std + 1e-8)
     
    return X, X_mean, X_std



def _add_feature(X):
    X_2 = np.power(X,3)
    X = np.concatenate([X,X_2], axis=1)
    return X
 
# 引入二次项
X_train = _add_feature(X_train)
X_test = _add_feature(X_test)


def _train_dev_split(X, Y, dev_ratio = 0.25):
    # This function spilts data into training set and development set.
    train_size = int(len(X) * (1 - dev_ratio))
    return X[:train_size], Y[:train_size], X[train_size:], Y[train_size:]

# Normalize training and testing data
X_train, X_mean, X_std = _normalize(X_train, train = True)
X_test, _, _= _normalize(X_test, train = False, specified_column = None, X_mean = X_mean, X_std = X_std)
    
# Split data into training set and development set
dev_ratio = 0.1
X_train, Y_train, X_dev, Y_dev = _train_dev_split(X_train, Y_train, dev_ratio = dev_ratio)

train_size = X_train.shape[0]
dev_size = X_dev.shape[0]
test_size = X_test.shape[0]
data_dim = X_train.shape[1]
print('Size of training set: {}'.format(train_size))
print('Size of development set: {}'.format(dev_size))
print('Size of testing set: {}'.format(test_size))
print('Dimension of data: {}'.format(data_dim))

def _shuffle(X, Y):
    # This function shuffles two equal-length list/array, X and Y, together.
    randomize = np.arange(len(X))
    np.random.shuffle(randomize)
    return (X[randomize], Y[randomize])

def _sigmoid(z):
    # Sigmoid function can be used to calculate probability.
    # To avoid overflow, minimum/maximum output value is set.
    return np.clip(1 / (1.0 + np.exp(-z)), 1e-8, 1 - (1e-8))

def _f(X, w, b):
    # This is the logistic regression function, parameterized by w and b
    #
    # Arguements:
    #     X: input data, shape = [batch_size, data_dimension]
    #     w: weight vector, shape = [data_dimension, ]
    #     b: bias, scalar
    # Output:
    #     predicted probability of each row of X being positively labeled, shape = [batch_size, ]
    return _sigmoid(np.matmul(X, w) + b)

def _predict(X, w, b):
    # This function returns a truth value prediction for each row of X 
    # by rounding the result of logistic regression function.
    return np.round(_f(X, w, b)).astype(np.int)
    
def _accuracy(Y_pred, Y_label):
    # This function calculates prediction accuracy
    acc = 1 - np.mean(np.abs(Y_pred - Y_label))
    return acc

def _cross_entropy_loss(y_pred, Y_label):
    # This function computes the cross entropy.
    #
    # Arguements:
    #     y_pred: probabilistic predictions, float vector
    #     Y_label: ground truth labels, bool vector
    # Output:
    #     cross entropy, scalar
    cross_entropy = -np.dot(Y_label, np.log(y_pred)) - np.dot((1 - Y_label), np.log(1 - y_pred))
    return cross_entropy

def _gradient(X, Y_label, w, b):
    # This function computes the gradient of cross entropy loss with respect to weight w and bias b.
    y_pred = _f(X, w, b)
    pred_error = Y_label - y_pred
    w_grad = -np.sum(pred_error * X.T, 1)
    b_grad = -np.sum(pred_error)
    return w_grad, b_grad

# Zero initialization for weights ans bias
w = np.zeros((data_dim,)) 
b = np.zeros((1,))

# Some parameters for training    
max_iter = 272
batch_size = 128
learning_rate = 0.1

# Keep the loss and accuracy at every iteration for plotting
train_loss = []
dev_loss = []
train_acc = []
dev_acc = []

# adagrad所需的累加和
adagrad_w = 0
adagrad_b = 0
# 防止adagrad除零
eps = 1e-8
 
# 迭代训练
for epoch in range(max_iter):
    # 在每个epoch开始时，随机打散训练数据
    X_train, Y_train = _shuffle(X_train, Y_train)
 
    # Mini-batch训练
    for idx in range(int(np.floor(train_size / batch_size))):
        X = X_train[idx * batch_size:(idx + 1) * batch_size]
        Y = Y_train[idx * batch_size:(idx + 1) * batch_size]
 
        # 计算梯度
        w_grad, b_grad = _gradient(X, Y, w, b)
        adagrad_w += w_grad**2
        adagrad_b += b_grad**2
 
        # 梯度下降法adagrad更新w和b
        w = w - learning_rate / (np.sqrt(adagrad_w + eps)) * w_grad
        b = b - learning_rate / (np.sqrt(adagrad_b + eps)) * b_grad



            
    # Compute loss and accuracy of training set and development set
    y_train_pred = _f(X_train, w, b)
    Y_train_pred = np.round(y_train_pred)
    train_acc.append(_accuracy(Y_train_pred, Y_train))
    train_loss.append(_cross_entropy_loss(y_train_pred, Y_train) / train_size)

    y_dev_pred = _f(X_dev, w, b)
    Y_dev_pred = np.round(y_dev_pred)
    dev_acc.append(_accuracy(Y_dev_pred, Y_dev))
    dev_loss.append(_cross_entropy_loss(y_dev_pred, Y_dev) / dev_size)

print('Training loss: {}'.format(train_loss[-1]))
print('Development loss: {}'.format(dev_loss[-1]))
print('Training accuracy: {}'.format(train_acc[-1]))
print('Development accuracy: {}'.format(dev_acc[-1]))

import matplotlib.pyplot as plt

# Loss curve
plt.plot(train_loss)
plt.plot(dev_loss)
plt.title('Loss')
plt.legend(['train', 'dev'])
plt.savefig('loss.png')
plt.show()

# Accuracy curve
plt.plot(train_acc)
plt.plot(dev_acc)
plt.title('Accuracy')
plt.legend(['train', 'dev'])
plt.savefig('acc.png')
plt.show()

# Predict testing labels
predictions = _predict(X_test, w, b)
with open(output_fpath.format('logistic'), 'w') as f:
    f.write('id,labeln')
    for i, label in  enumerate(predictions):
        f.write('{},{}n'.format(i, label))

本人刚接触机器学习不久，以写博客的方式记录学习过程同时也为了帮助和我一起在学习过程的遇到困难的小伙伴,文中如有错误，希望大家指正。