机器学习系列------除噪

alltolove

        还是把昨天代码复制过来：

from sklearn.datasets import fetch_mldata



mnist=fetch_mldata('MNIST original',data_home='.\datasets')


X,y=mnist["data"],mnist["target"]
%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt

X_train,X_test,y_train,y_test=X[:60000],X[60000:],y[:60000],y[60000:]

import numpy as np
shuffle_index=np.random.permutation(60000)
X_train,y_train=X_train[shuffle_index],y_train[shuffle_index]
some_digit=X[36000]

from sklearn.linear_model import SGDClassifier

sgd_clf=SGDClassifier(random_state=42)

        然后我们分别把输出项里大于等于7的数和奇数取2组布尔值：

from sklearn.neighbors import KNeighborsClassifier
y_train_large=(y_train>=7)
y_train_odd=(y_train%2==1)
y_multilabel=np.c_[y_train_large,y_train_odd]
y_multilabel

        输出为：
array([[False, False],
       [False, False],
       [False, False],
       ...,
       [False, False],
       [False, False],
       [False, False]], dtype=bool)
        以下这种分类方法就可以一起预测2个条件：

knn_clf=KNeighborsClassifier()
knn_clf.fit(X_train,y_multilabel)
knn_clf.predict([some_digit])

        输出为：array([[False,  True]], dtype=bool)。果然5是小于7而且是奇数。然后我们人为在图片里制造一些噪声：

from numpy import random as rnd

noise1=rnd.randint(0,100,(len(X_train),784))
noise2=rnd.randint(0,100,(len(X_test),784))
X_train_mod=X_train+noise1
X_test_mod=X_test+noise2
y_train_mod=X_train
y_test_mod=X_test

        随便找个数画出图来：

%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt

some_digit=X_train_mod[36000]
some_digit_imge=some_digit.reshape(28,28)
plt.imshow(some_digit_imge,cmap=matplotlib.cm.binary,interpolation="nearest")
plt.axis("off")
plt.show()

        输出图像：

        再看看原始的图片：

%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt

some_digit=y_train_mod[36000]
some_digit_imge=some_digit.reshape(28,28)
plt.imshow(some_digit_imge,cmap=matplotlib.cm.binary,interpolation="nearest")
plt.axis("off")
plt.show()

        输出为：

        以下代码是通过上面这种分类器除去随便一张图的噪音：

knn_clf.fit(X_train_mod,y_train_mod)
clean_digit=knn_clf.predict([X_test_mod[100]])

        然后画图：

%matplotlib inline
import matplotlib
import matplotlib.pyplot as plt

some_digit=clean_digit
some_digit_imge=some_digit.reshape(28,28)
plt.imshow(some_digit_imge,cmap=matplotlib.cm.binary,interpolation="nearest")
plt.axis("off")
plt.show()

        输出图片为：