机器学习练习题-MNIST数据增强

题目： 写一个可以将MNIST图片向任意方向（上、下、左、右）移动一个像素的功能。然后，对训练集中的每张图片，创建四个位移后的副本（每个方向一个），添加到训练集。最后，在这个扩展过的训练集上训练模型，衡量其在测试集上的精度。你应该能注意到，模型的表现甚至变得更好了！这种人工扩展训练集的技术称为数据增广或训练集扩展。

参考代码：

from sklearn.datasets import fetch_mldata
import numpy as np
from sklearn.neighbors import KNeighborsClassifier


mnist = fetch_mldata(MNIST original, data_home=./mnist1)
X, y = mnist["data"], mnist["target"]
#print(X.shape)
X_train = X[:60000]
y_train = y[:60000]
X_test = X[60000:]
y_test = y[60000:]
shuffle_index = np.random.permutation(60000)
X_train, y_train = X_train[shuffle_index], y_train[shuffle_index]
X_train, y_train = X_train[:60000], y_train[:60000]

train_array = []
train_size = X_train.shape[0]
for i in range(train_size):
#for i in range(122):
    train_array.append(X_train[i].reshape(28, 28))
    
print(train_size=, train_size)

def data_en(data_s, direc=u):
    size=len(data_s)
    en_ret = np.zeros((size, 784))
    if direc == u:        
        for i in range(size):
            trans_data = np.append(data_s[i][1:,:], data_s[i][0:1,:],axis=0)
            #print(trans_data.shape)
            en_ret[i] = trans_data.reshape(1, -1)
    elif direc == d:
        for i in range(size):
            trans_data = np.append(data_s[i][-1:,:], data_s[i][:-1,:],axis=0)
            #print(trans_data.shape)
            en_ret[i] = trans_data.reshape(1, -1)
    elif direc == l:
        for i in range(size):
            trans_data = np.append(data_s[i][:,1:], data_s[i][:,0:1],axis=1)
            #print(trans_data.shape)
            en_ret[i] = trans_data.reshape(1, -1)
    elif direc == r:
        for i in range(size):
            trans_data = np.append(data_s[i][:,-1:], data_s[i][:,:-1],axis=1)
            #plt.imshow(trans_data, cmap = matplotlib.cm.binary,interpolation="nearest")
            en_ret[i] = trans_data.reshape(1, -1)
    return en_ret

X_trainu = data_en(train_array, u)
X_traind = data_en(train_array, d)
X_trainl = data_en(train_array, l)
X_trainr = data_en(train_array, r)

#X_trainA = np.append(X_train, X_trainu, axis=0)
X_trainA = np.concatenate((X_train, X_trainu, X_traind, X_trainl, X_trainr), axis=0)
y_trainA = np.concatenate((y_train, y_train, y_train, y_train, y_train), axis=0)
print(X_trainA.shape)
print(y_trainA.shape)

knn_clf = KNeighborsClassifier()
knn_clf.fit(X_train, y_train)
y_pred = knn_clf.predict(X_test)
from sklearn.metrics import precision_score, recall_score,confusion_matrix
ps = precision_score(y_test, y_pred, average=None)
print(
ps=, ps, np.average(ps))

knn_clfA = KNeighborsClassifier()
knn_clfA.fit(X_trainA, y_trainA)
y_pred = knn_clfA.predict(X_test)
psA = precision_score(y_test, y_pred, average=None)
print(
psA=, psA, np.average(psA))

输出结果：

(train_size=, 60000)
(300000, 784)
(300000,)
(
ps=, array([ 0.96340257,  0.95450716,  0.98216056,  0.96442688,  0.9762151 ,
        0.96528555,  0.98130841,  0.96108949,  0.98809524,  0.95626243]), 0.96927533865705706)
(
psA=, array([ 0.97887324,  0.95850974,  0.98624754,  0.97628458,  0.97841727,
        0.9698324 ,  0.98130841,  0.96605238,  0.99353448,  0.96915423]), 0.97582142737164035)

精度由0.9692提升到0.9758。