From aee8dc20e6e5c58f6cba3c99202017a057b48231 Mon Sep 17 00:00:00 2001
From: corvettettt <corvettettt@gmail.com>
Date: Thu, 30 Apr 2020 00:28:38 -0500
Subject: [PATCH] new
---
face.rec.klda.homework.py | 222 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
homework2/œÓƒø÷∏µº£∫Kernel trick with LDA.docx | Bin 0 -> 528387 bytes
homework2/Ω·π˚’π æÕº.jpg | Bin 0 -> 101799 bytes
homework2/◊˜“µÃ· æ1.jpg | Bin 0 -> 326256 bytes
homework2/◊˜“µÃ· æ2.jpg | Bin 0 -> 251135 bytes
5 files changed, 222 insertions(+)
create mode 100644 face.rec.klda.homework.py
create mode 100755 homework2/œÓƒø÷∏µº£∫Kernel trick with LDA.docx
create mode 100755 homework2/Ω·π˚’π æÕº.jpg
create mode 100755 homework2/◊˜“µÃ· æ1.jpg
create mode 100755 homework2/◊˜“µÃ· æ2.jpg
diff --git a/face.rec.klda.homework.py b/face.rec.klda.homework.py
new file mode 100644
index 0000000..75a33b6
--- /dev/null
+++ b/face.rec.klda.homework.py
@@ -0,0 +1,222 @@
+"""
+==============================================================
+基于 Kernel LDA + KNN 的人脸识别
+使用 Kernel Discriminant Analysis 做特征降维
+使用 K-Nearest-Neighbor 做分类
+
+数据:
+ 人脸图像来自于 Olivetti faces data-set from AT&T (classification)
+ 数据集包含 40 个人的人脸图像, 每个人都有 10 张图像
+ 我们只使用其中标签(label/target)为 0 和 1 的前 2 个人的图像
+
+算法:
+ 需要自己实现基于 RBF Kernel 的 Kernel Discriminant Analysis 用于处理两个类别的数据的特征降维
+ 代码的框架已经给出, 需要学生自己补充 KernelDiscriminantAnalysis 的 fit() 和 transform() 函数的内容
+==============================================================
+"""
+
+# License: BSD 3 clause
+
+import numpy as np
+import matplotlib.pyplot as plt
+import math
+from sklearn import datasets
+from sklearn.model_selection import train_test_split
+from sklearn.datasets import fetch_olivetti_faces
+from sklearn.neighbors import KNeighborsClassifier
+from sklearn.pipeline import make_pipeline
+from sklearn.preprocessing import StandardScaler
+
+print(__doc__)
+################################################
+"""
+Scikit-learn-compatible Kernel Discriminant Analysis.
+"""
+
+import numpy as np
+from scipy import linalg
+from sklearn.base import BaseEstimator, ClassifierMixin, TransformerMixin
+from sklearn.preprocessing import OneHotEncoder
+from sklearn.utils.validation import check_array, check_is_fitted, check_X_y
+from sklearn.metrics.pairwise import rbf_kernel
+from numpy.linalg import matrix_rank, eigh, inv
+
+class KernelDiscriminantAnalysis(BaseEstimator, ClassifierMixin,
+ TransformerMixin):
+ """Kernel Discriminant Analysis.
+
+ Parameters
+ ----------
+ n_components: integer.
+ The dimension after transform.
+ gamma: float.
+ Parameter to RBF Kernel
+ lmb: float (>= 0.0), default=0.001.
+ Regularization parameter
+
+ """
+
+ def __init__(self, n_components, gamma, lmb=0.001):
+ self.n_components = n_components
+ self.gamma = gamma
+ self.lmb = lmb
+ self.X = None # 用于存放输入的训练数据的 X
+ self.K = None # 用于存放训练数据 X 产生的 Kernel Matrix
+ self.M = None # 用于存放 Kernel LDA 最优化公式中的 M
+ self.N = 0 # 用于存放 Kernel LDA 最优化公式中的 N
+ self.EigenVectors = None # 用于存放 Kernel LDA 最优化公式中的 M 对应的广义特征向量, 每一列为一个特征向量, 按照对应特征值大小排序
+
+ def Kernal(self,x,y,gamma):
+ return math.exp(-(x-y).T.dot(x-y)*gamma)
+
+ def Kernal_matrix(self,x1,x2,gamma):
+ #print(x1.shape,x2.shape,gamma)
+ #print(self.Kernal([1,1],[2,2],gamma=gamma))
+ return np.fromfunction(np.vectorize(lambda i,j: self.Kernal(x1[int(i),:],x2[int(j),:],gamma=gamma)), (x1.shape[0],x2.shape[0]), dtype=np.float32)
+
+ def fit(self, X, y):
+ """Fit KDA model.
+
+ Parameters
+ ----------
+ X: numpy array of shape [n_samples, n_features]
+ Training set.
+ y: numpy array of shape [n_samples]
+ Target values. Only works for 2 classes with label/target 0 and 1.
+
+ Returns
+ -------
+ self
+
+ """
+
+ self.X = X
+
+ ### n is the # of samples, and m is the dimension of data
+ n,m = X.shape
+
+ ### X0 are those whose tag ==0, X1 are those whose tag ==1
+ X0 = X[y==0]
+ X1 = X[y==1]
+
+ ### Computing the number of samples of each tag
+ l=[]
+ l.append(X0.shape[0])
+ l.append(X1.shape[0])
+
+ ### Computing M, based on M = (M0-M1) @ (M1-M1).T, M is symmetric
+ M0 = np.mean(self.Kernal_matrix(X,X0,gamma=self.gamma),axis=-1).reshape(-1,1)
+ M1 = np.mean(self.Kernal_matrix(X,X1,gamma=self.gamma),axis=-1).reshape(-1,1)
+ self.M = (M0-M1) @ (M0-M1).T
+
+ ### Computing kernal matrix K0 and K1
+ self.K = []
+ self.K.append(rbf_kernel(X,X0,gamma=self.gamma))
+ self.K.append(rbf_kernel(X,X1,gamma=self.gamma))
+ ### Computing N, N is symmetric
+ for i in range(1):
+ self.N += self.K[i] @ (np.eye(l[i]) - np.full((l[i],l[i]),1./l[i])) @ self.K[i].T
+
+ ### rebuild N if it's not invertable
+ AddOn = 0.01
+ while matrix_rank(self.N) != self.N.shape[0]:
+ self.N += np.eye(n)*AddOn
+ AddOn *= 5
+
+ ### Computing the EigenValue of matrix N^-1 * M, because both M and N are symmetric
+ ### we can use the eigenvalue of N^-1 * M
+ E_values, E_vectors = eigh(inv(self.N) @ self.M)
+
+ ### sort eigenvalues and associated eigenvectors
+ idx = E_values.argsort()[::-1]
+ E_values = E_values[idx]
+ E_vectors = E_vectors[:,idx]
+
+ ### store the eigenvectors
+ self.EigenVectors = E_vectors
+
+ def transform(self, X_test):
+ """Transform data with the trained KernelLDA model.
+
+ Parameters
+ ----------
+ X_test: numpy array of shape [n_samples, n_features]
+ The input data.
+
+ Returns
+ -------
+ y_pred: array-like, shape (n_samples, n_components)
+ Transformations for X.
+
+ """
+ ### Take the eigenvectors corresponding to n_components highest eigenvalues
+ Transform_maxtrix = self.EigenVectors[:,:self.n_components]
+ ### transform data
+ X_transformed = rbf_kernel(self.X,X_test,gamma=self.gamma).T @ Transform_maxtrix
+
+ return X_transformed
+################################################
+
+# 指定 KNN 中最近邻的个数 (k 的值)
+n_neighbors = 3
+
+# 设置随机数种子让实验可以复现
+random_state = 0
+
+# 现在人脸数据集
+faces = fetch_olivetti_faces()
+targets = faces.target
+
+# show sample images
+images = faces.images[targets < 2] # save images
+
+features = faces.data # features
+targets = faces.target # targets
+
+fig = plt.figure() # create a new figure window
+for i in range(20): # display 20 images
+ # subplot : 4 rows and 5 columns
+ img_grid = fig.add_subplot(4, 5, i+1)
+ # plot features as image
+ img_grid.imshow(images[i], cmap='gray')
+
+plt.show()
+
+# Prepare data, 只限于处理类别 0 和 1 的人脸
+X, y = faces.data[targets < 2], faces.target[targets < 2]
+
+# Split into train/test
+X_train, X_test, y_train, y_test = \
+ train_test_split(X, y, test_size=0.5, stratify=y,
+ random_state=random_state)
+
+
+# Reduce dimension to 2 with KernelDiscriminantAnalysis
+# can adjust the value of 'gamma' as needed.
+kda = make_pipeline(StandardScaler(),
+ KernelDiscriminantAnalysis(n_components=2, gamma = 0.000005))
+
+# Use a nearest neighbor classifier to evaluate the methods
+knn = KNeighborsClassifier(n_neighbors=n_neighbors)
+
+
+plt.figure()
+# plt.subplot(1, 3, i + 1, aspect=1)
+
+# Fit the method's model
+kda.fit(X_train, y_train)
+
+# Fit a nearest neighbor classifier on the embedded training set
+knn.fit(kda.transform(X_train), y_train)
+
+# Compute the nearest neighbor accuracy on the embedded test set
+acc_knn = knn.score(kda.transform(X_test), y_test)
+# Embed the data set in 2 dimensions using the fitted model
+X_embedded = kda.transform(X)
+
+# Plot the projected points and show the evaluation score
+plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=y, s=30, cmap='Set1')
+plt.title("{}, KNN (k={})\nTest accuracy = {:.2f}".format('kda',
+ n_neighbors,
+ acc_knn))
+plt.show()
diff --git "a/homework2/\305\223\303\223\306\222\303\270\303\267\342\210\217\302\265\302\272\302\243\342\210\253Kernel trick with LDA.docx" "b/homework2/\305\223\303\223\306\222\303\270\303\267\342\210\217\302\265\302\272\302\243\342\210\253Kernel trick with LDA.docx"
new file mode 100755
index 0000000..1dd82b6
Binary files /dev/null and "b/homework2/\305\223\303\223\306\222\303\270\303\267\342\210\217\302\265\302\272\302\243\342\210\253Kernel trick with LDA.docx" differ
diff --git "a/homework2/\316\251\302\267\317\200\313\232\342\200\231\317\200\302\240\303\246\303\225\302\272.jpg" "b/homework2/\316\251\302\267\317\200\313\232\342\200\231\317\200\302\240\303\246\303\225\302\272.jpg"
new file mode 100755
index 0000000..2f86574
Binary files /dev/null and "b/homework2/\316\251\302\267\317\200\313\232\342\200\231\317\200\302\240\303\246\303\225\302\272.jpg" differ
diff --git "a/homework2/\342\227\212\313\234\342\200\234\302\265\303\203\302\267\302\240\303\2461.jpg" "b/homework2/\342\227\212\313\234\342\200\234\302\265\303\203\302\267\302\240\303\2461.jpg"
new file mode 100755
index 0000000..2b70d0a
Binary files /dev/null and "b/homework2/\342\227\212\313\234\342\200\234\302\265\303\203\302\267\302\240\303\2461.jpg" differ
diff --git "a/homework2/\342\227\212\313\234\342\200\234\302\265\303\203\302\267\302\240\303\2462.jpg" "b/homework2/\342\227\212\313\234\342\200\234\302\265\303\203\302\267\302\240\303\2462.jpg"
new file mode 100755
index 0000000..dd0af80
Binary files /dev/null and "b/homework2/\342\227\212\313\234\342\200\234\302\265\303\203\302\267\302\240\303\2462.jpg" differ
--
libgit2 0.26.0