因爲近來一直再看kaggle的入門書(sklearn入門手冊的感受233),感受對機器學習的理解加深了很多(實際上就只是調包能力增強了),聯想到假期在python科學計算上也算是進行了一些嘗試學習,以爲仍是須要學習一下機器學習原理的,因此從新啃起了吳恩達的cs229,上次(5月份的時候?)就是在多元高斯分佈這裏吃的癟,看不下去了,此次覺定穩紮穩打,不求速度多實踐實踐,儘可能理解數學原理,因此再次看到這部分時決定把這個分佈復現出來,吳恩達大佬用的matlab,我用的python,畫的還不錯,代碼以下,python
import numpy as np import matplotlib.pyplot as plt from mpl_toolkits.mplot3d import axes3d from matplotlib import cm import matplotlib as mpl num = 200 l = np.linspace(-5,5,num) X, Y =np.meshgrid(l, l) u = np.array([0, 0]) o = np.array([[1, 0.5], [0.5, 1]]) pos = np.concatenate((np.expand_dims(X,axis=2),np.expand_dims(Y,axis=2)),axis=2) a = (pos-u).dot(np.linalg.inv(o)) b = np.expand_dims(pos-u,axis=3) Z = np.zeros((num,num), dtype=np.float32) for i in range(num): Z[i] = [np.dot(a[i,j],b[i,j]) for j in range(num)] Z = np.exp(Z*(-0.5))/(2*np.pi*np.linalg.det(o)) fig = plt.figure() ax = fig.add_subplot(111,projection='3d') ax.plot_surface(X, Y, Z, rstride=5, cstride=5, alpha=0.3, cmap=cm.coolwarm) cset = ax.contour(X,Y,Z,10,zdir='z',offset=0,cmap=cm.coolwarm) cset = ax.contour(X, Y, Z, zdir='x', offset=-5,cmap=mpl.cm.winter) cset = ax.contour(X, Y, Z, zdir='y', offset= 5,cmap= mpl.cm.winter) ''' mpl.cm.rainbow mpl.cm.winter mpl.cm.bwr # 藍,白,紅 cm.coolwarm ''' ax.set_xlabel('X') ax.set_ylabel('Y') ax.set_zlabel('Z') plt.show()
實際操做中,能夠看到我在Z生成部分使用了雙層循環,我本意是使用numpy廣播機制優化掉循環,實際操做不太順利,(20,20,2)去叉乘(20,20,2,1),結果shape不是我指望的(20,20,1),而是(20,20,20,20,1),也就是說在高維叉乘時其實廣播機制不太好用,畢竟實際上兩個不一樣維度矩陣是能夠直接叉乘的(雖然對維度有要求),這一點值得注意(高維矩陣叉乘不要依賴numpy的廣播機制)。機器學習
參數:ide
u = np.array([0, 0])
o = np.array([[1, 0.5],
[0.5, 1]])
參數:學習
u = np.array([1, 1])
o = np.array([[1, 0],
[0, 1]])
參數:優化
u = np.array([1, 1])
o = 3*np.array([[1, 0],
[0, 1]])
咱們單獨繪製一下等高線圖,spa
# 前面添加圖的位置修改以下, # ax = fig.add_subplot(211,projection='3d') ax2 = fig.add_subplot(212) cs = ax2.contour(X,Y,Z) ax2.clabel(cs, inline=1, fontsize=20)
如今咱們在上面代碼的基礎上可視化吳恩達老大的下一節的圖,高斯判別分析模型可視化,這裏面咱們僅僅可視化基礎的雙高斯獨立分佈,代碼以下,3d
import numpy as np import matplotlib.pyplot as plt from mpl_toolkits.mplot3d import axes3d from matplotlib import cm import matplotlib as mpl num = 200 l = np.linspace(-5,5,num) X, Y =np.meshgrid(l, l) pos = np.concatenate((np.expand_dims(X,axis=2),np.expand_dims(Y,axis=2)),axis=2) u1 = np.array([2, 2]) o1 = 3*np.array([[1, 0], [0, 1]]) a1 = (pos-u1).dot(np.linalg.inv(o1)) b1 = np.expand_dims(pos-u1,axis=3) Z1 = np.zeros((num,num), dtype=np.float32) u2 = np.array([-2, -2]) o2 = 3*np.array([[1, 0], [0, 1]]) a2 = (pos-u2).dot(np.linalg.inv(o2)) b2 = np.expand_dims(pos-u2,axis=3) Z2 = np.zeros((num,num), dtype=np.float32) for i in range(num): Z1[i] = [np.dot(a1[i,j],b1[i,j]) for j in range(num)] Z2[i] = [np.dot(a2[i,j],b2[i,j]) for j in range(num)] Z1 = np.exp(Z1*(-0.5))/(2*np.pi*np.linalg.det(o1)) Z2 = np.exp(Z2*(-0.5))/(2*np.pi*np.linalg.det(o1)) Z = Z1 + Z2 fig = plt.figure() ax = fig.add_subplot(211,projection='3d') ax.plot_surface(X, Y, Z, rstride=5, cstride=5, alpha=0.3, cmap=cm.coolwarm) cset = ax.contour(X,Y,Z,10,zdir='z',offset=0,cmap=cm.coolwarm) cset = ax.contour(X, Y, Z, zdir='x', offset=-5,cmap=mpl.cm.winter) cset = ax.contour(X, Y, Z, zdir='y', offset= 5,cmap= mpl.cm.winter) ''' mpl.cm.rainbow mpl.cm.winter mpl.cm.bwr # 藍,白,紅 cm.coolwarm ''' ax.set_xlabel('X') ax.set_ylabel('Y') ax.set_zlabel('Z') plt.show() ax2 = fig.add_subplot(212) cs = ax2.contour(X,Y,Z) ax2.clabel(cs, inline=1, fontsize=20)
不過吳老大的圖兩個高斯分佈投影是分開的,因此咱們再次小改繪圖部分,blog
cset = ax.contour(X,Y,Z1,10,zdir='z',offset=0,cmap=cm.coolwarm) cset = ax.contour(X,Y,Z2,10,zdir='z',offset=0,cmap=cm.coolwarm) cset = ax.contour(X, Y, Z, zdir='x', offset=-5,cmap=mpl.cm.winter) cset = ax.contour(X, Y, Z, zdir='y', offset= 5,cmap= mpl.cm.winter) ''' mpl.cm.rainbow mpl.cm.winter mpl.cm.bwr # 藍,白,紅 cm.coolwarm ''' ax.set_xlabel('X') ax.set_ylabel('Y') ax.set_zlabel('Z') plt.show() ax2 = fig.add_subplot(212) cs = ax2.contour(X,Y,Z1) ax2.clabel(cs, inline=1, fontsize=20) cs2 = ax2.contour(X,Y,Z2) ax2.clabel(cs2, inline=1, fontsize=20)
顯示以下,框子不夠標準致使圓有點變形,不過這個能夠經過手動拉伸獲得優化,因此問題不大,博客
有關多元正態分佈的數學原理建議自行百度(cs229的學習不會在博客上更新,主要是由於我很是很是討厭打數學公式233)。數學