質問編集履歴

書式の改善

2020/10/05 00:28

投稿

mizu103

スコア1

test CHANGED Viewed

File without changes

test CHANGED Viewed

@@ -1,4 +1,270 @@
+コード
+```
+import numpy as np
+import matplotlib.pyplot as plt
+class RBFkernel:
+    def __init__(self,*param):
+        self.param = list(param)
+    def __call__(self,x1,x2):
+        a,s,w = self.param
+        return a**2*np.exp(-((x1-x2)/s)**2) + w*(x1==x2)
+x0 = (0,5,10,20,25,33,50)
+y0 = (156,152,148,135,156,145,158)
+x1 = np.linspace(0,50,201)
+kernel = RBFkernel(10,5,3) # 適当なパラメータを使うカーネル関数
+k00 = kernel(*np.meshgrid(x0,x0))
+k00_1 = np.linalg.inv(k00) # 逆行列
+k01 = kernel(*np.meshgrid(x0,x1,indexing='ij'))
+k10 = k01.T
+k11 = kernel(*np.meshgrid(x1,x1))
+mu = k10.dot(k00_1.dot(y0))
+sigma = k11 - k10.dot(k00_1.dot(k01))
+plt.scatter(x0,y0,c='#ff77aa')
+plt.plot(x1,mu,'g') # 推測された平均
+std = np.sqrt(sigma.diagonal()) # 各点の標準偏差は共分散行列の対角成分
+plt.fill_between(x1,mu-std,mu+std,alpha=0.2,color='g') # 推測された標準偏差の中の領域
+plt.show()
+class Kernel:
+    def __init__(self,param,bound=None):
+        self.param = np.array(param)
+        if(bound==None):
+            bound = np.zeros([len(param),2])
+            bound[:,1] = np.inf
+        self.bound = np.array(bound)
+    def __call__(self,x1,x2):
+        a,s,w = self.param
+        return a**2*np.exp(-0.5*((x1-x2)/s)**2) + w**2*(x1==x2)
+class Gausskatei:
+    def __init__(self,kernel):
+        self.kernel = kernel
+    def gakushuu(self,x0,y0): # パラメータを調整せず学習
+        self.x0 = x0
+        self.y0 = y0
+        self.k00 = self.kernel(*np.meshgrid(x0,x0))
+        self.k00_1 = np.linalg.inv(self.k00)
+    def yosoku(self,x): # xからyを予測
+        k00_1 = self.k00_1
+        k01 = self.kernel(*np.meshgrid(self.x0,x,indexing='ij'))
+        k10 = k01.T
+        k11 = self.kernel(*np.meshgrid(x,x))
+        mu = k10.dot(k00_1.dot(self.y0))
+        sigma = k11 - k10.dot(k00_1.dot(k01))
+        std = np.sqrt(sigma.diagonal())
+        return mu,std
+    def logyuudo(self,param=None): # 対数尤度
+        if(param is None):
+            k00 = self.k00
+            k00_1 = self.k00_1
+        else:
+            self.kernel.param = param
+            k00 = self.kernel(*np.meshgrid(self.x0,self.x0))
+            k00_1 = np.linalg.inv(k00)
+        return -(np.linalg.slogdet(k00)[1]+self.y0.dot(k00_1.dot(self.y0)))
+    def saitekika(self,x0,y0,kurikaeshi=1000): # パラメータを調整して学習
+        self.x0 = x0
+        self.y0 = y0
+        param = self.kernel.param
+        logbound = np.log(self.kernel.bound)
+        s = (logbound[:,1]-logbound[:,0])/10.
+        n_param = len(param)
+        theta0 = np.log(param)
+        p0 = self.logyuudo(param)
+        lis_theta = []
+        lis_p = []
+        for i in range(kurikaeshi):
+            idou = np.random.normal(0,s,n_param)
+            hazure = (theta0+idou<logbound[:,0])|(theta0+idou>logbound[:,1])
+            while(np.any(hazure)):
+                idou[hazure] = np.random.normal(0,s,n_param)[hazure]
+                hazure = (theta0+idou<logbound[:,0])|(theta0+idou>logbound[:,1])
+            theta1 = theta0 + idou
+            param = np.exp(theta1)
+            p1 = self.logyuudo(param)
+            r = np.exp(p1-p0)
+            if(r>=1 or r>np.random.random()):
+                theta0 = theta1
+                p0 = p1
+                lis_theta.append(theta0)
+                lis_p.append(p0)
+        self.ar_theta = np.array(lis_theta)
+        self.ar_p = np.array(lis_p)
+        self.kernel.param = np.exp(lis_theta[np.argmax(lis_p)])
+        self.k00 = self.kernel(*np.meshgrid(x0,x0))
+        self.k00_1 = np.linalg.inv(self.k00)
+n = 7 # 既知の点の数
+x0 = (0,5,10,20,25,33,50) # 既知の点
+y0 = (156,152,148,135,156,145,158)
+param0 = [10,5,1] # パラメータの初期値
+bound = [[1e-2,1e2],[1e-2,1e2],[1e-2,1e2]] # 下限上限
+kernel = Kernel(param0,bound)
+x1 = np.linspace(0,50,200)
+gp = Gausskatei(kernel)
+gp.gakushuu(x0,y0) # パラメータを調整せずに学習
+plt.figure(figsize=[5,8])
+for i in [0,1]:
+    if(i):
+        gp.saitekika(x0,y0,1000) # パラメータを調整する
+    plt.subplot(211+i)
+    plt.plot(x0,y0,'. ')
+    mu,std = gp.yosoku(x1)
+    plt.plot(x1,mu,'g')
+    plt.fill_between(x1,mu-std,mu+std,alpha=0.2,color='g')
+    plt.title('a=%.3f, s=%.3f, w=%.3f'%tuple(gp.kernel.param))
+plt.tight_layout()
+plt.show()
+```
+コード
-### 前提・実現したいこと
+```### 前提・実現したいこと
 https://qiita.com/phyblas/items/d756803ec932ab621c56
@@ -62,266 +328,52 @@
 ### 該当のソースコード
-import numpy as np
-import matplotlib.pyplot as plt
-class RBFkernel:
-    def __init__(self,*param):
-        self.param = list(param)
-    def __call__(self,x1,x2):
-        a,s,w = self.param
+n = 7 # 既知の点の数
-        return a**2*np.exp(-((x1-x2)/s)**2) + w*(x1==x2)
-x0 = (0,5,10,20,25,33,50)
+x0 = (0,5,10,20,25,33,50) # 既知の点
 y0 = (156,152,148,135,156,145,158)
+param0 = [10,5,1] # パラメータの初期値
+bound = [[1e-2,1e2],[1e-2,1e2],[1e-2,1e2]] # 下限上限
+kernel = Kernel(param0,bound)
-x1 = np.linspace(0,50,201)
+x1 = np.linspace(0,50,200)
+gp = Gausskatei(kernel)
-kernel = RBFkernel(10,5,3) # 適当なパラメータを使うカーネル関数
+gp.gakushuu(x0,y0) # パラメータを調整せずに学習
-k00 = kernel(*np.meshgrid(x0,x0))
+plt.figure(figsize=[5,8])
-k00_1 = np.linalg.inv(k00) # 逆行列
-k01 = kernel(*np.meshgrid(x0,x1,indexing='ij'))
-k10 = k01.T
+for i in [0,1]:
-k11 = kernel(*np.meshgrid(x1,x1))
-mu = k10.dot(k00_1.dot(y0))
+    if(i):
-sigma = k11 - k10.dot(k00_1.dot(k01))
+        gp.saitekika(x0,y0,1000) # パラメータを調整する
+    plt.subplot(211+i)
-plt.scatter(x0,y0,c='#ff77aa')
+    plt.plot(x0,y0,'. ')
+    mu,std = gp.yosoku(x1)
-plt.plot(x1,mu,'g') # 推測された平均
+    plt.plot(x1,mu,'g')
-std = np.sqrt(sigma.diagonal()) # 各点の標準偏差は共分散行列の対角成分
-plt.fill_between(x1,mu-std,mu+std,alpha=0.2,color='g') # 推測された標準偏差の中の領域
+    plt.fill_between(x1,mu-std,mu+std,alpha=0.2,color='g')
+    plt.title('a=%.3f, s=%.3f, w=%.3f'%tuple(gp.kernel.param))
+plt.tight_layout()
 plt.show()
-class Kernel:
-    def __init__(self,param,bound=None):
-        self.param = np.array(param)
-        if(bound==None):
-            bound = np.zeros([len(param),2])
-            bound[:,1] = np.inf
-        self.bound = np.array(bound)
-    def __call__(self,x1,x2):
-        a,s,w = self.param
-        return a**2*np.exp(-0.5*((x1-x2)/s)**2) + w**2*(x1==x2)
-class Gausskatei:
-    def __init__(self,kernel):
-        self.kernel = kernel
-    def gakushuu(self,x0,y0): # パラメータを調整せず学習
-        self.x0 = x0
-        self.y0 = y0
-        self.k00 = self.kernel(*np.meshgrid(x0,x0))
-        self.k00_1 = np.linalg.inv(self.k00)
-    def yosoku(self,x): # xからyを予測
-        k00_1 = self.k00_1
-        k01 = self.kernel(*np.meshgrid(self.x0,x,indexing='ij'))
-        k10 = k01.T
-        k11 = self.kernel(*np.meshgrid(x,x))
-        mu = k10.dot(k00_1.dot(self.y0))
-        sigma = k11 - k10.dot(k00_1.dot(k01))
-        std = np.sqrt(sigma.diagonal())
-        return mu,std
-    def logyuudo(self,param=None): # 対数尤度
-        if(param is None):
-            k00 = self.k00
-            k00_1 = self.k00_1
-        else:
-            self.kernel.param = param
-            k00 = self.kernel(*np.meshgrid(self.x0,self.x0))
-            k00_1 = np.linalg.inv(k00)
-        return -(np.linalg.slogdet(k00)[1]+self.y0.dot(k00_1.dot(self.y0)))
-    def saitekika(self,x0,y0,kurikaeshi=1000): # パラメータを調整して学習
-        self.x0 = x0
-        self.y0 = y0
-        param = self.kernel.param
-        logbound = np.log(self.kernel.bound)
-        s = (logbound[:,1]-logbound[:,0])/10.
-        n_param = len(param)
-        theta0 = np.log(param)
-        p0 = self.logyuudo(param)
-        lis_theta = []
-        lis_p = []
-        for i in range(kurikaeshi):
-            idou = np.random.normal(0,s,n_param)
-            hazure = (theta0+idou<logbound[:,0])|(theta0+idou>logbound[:,1])
-            while(np.any(hazure)):
-                idou[hazure] = np.random.normal(0,s,n_param)[hazure]
-                hazure = (theta0+idou<logbound[:,0])|(theta0+idou>logbound[:,1])
-            theta1 = theta0 + idou
-            param = np.exp(theta1)
-            p1 = self.logyuudo(param)
-            r = np.exp(p1-p0)
-            if(r>=1 or r>np.random.random()):
-                theta0 = theta1
-                p0 = p1
-                lis_theta.append(theta0)
-                lis_p.append(p0)
-        self.ar_theta = np.array(lis_theta)
-        self.ar_p = np.array(lis_p)
-        self.kernel.param = np.exp(lis_theta[np.argmax(lis_p)])
-        self.k00 = self.kernel(*np.meshgrid(x0,x0))
-        self.k00_1 = np.linalg.inv(self.k00)
-n = 7 # 既知の点の数
-x0 = (0,5,10,20,25,33,50) # 既知の点
-y0 = (156,152,148,135,156,145,158)
-param0 = [10,5,1] # パラメータの初期値
-bound = [[1e-2,1e2],[1e-2,1e2],[1e-2,1e2]] # 下限上限
-kernel = Kernel(param0,bound)
-x1 = np.linspace(0,50,200)
-gp = Gausskatei(kernel)
-gp.gakushuu(x0,y0) # パラメータを調整せずに学習
-plt.figure(figsize=[5,8])
-for i in [0,1]:
-    if(i):
-        gp.saitekika(x0,y0,1000) # パラメータを調整する
-    plt.subplot(211+i)
-    plt.plot(x0,y0,'. ')
-    mu,std = gp.yosoku(x1)
-    plt.plot(x1,mu,'g')
-    plt.fill_between(x1,mu-std,mu+std,alpha=0.2,color='g')
-    plt.title('a=%.3f, s=%.3f, w=%.3f'%tuple(gp.kernel.param))
-plt.tight_layout()
-plt.show()
 ### 試したこと

タグ変更

2020/10/05 00:28

投稿

mizu103

スコア1

test CHANGED Viewed

File without changes

test CHANGED Viewed

File without changes

書式の改善

2020/10/02 00:36

投稿

mizu103

スコア1

test CHANGED Viewed

File without changes

test CHANGED Viewed

@@ -108,7 +108,7 @@
-# ここでは上述の方程式の通りのμとΣ
 mu = k10.dot(k00_1.dot(y0))

コード補足

2020/09/30 06:30

投稿

mizu103

スコア1

test CHANGED Viewed

File without changes

test CHANGED Viewed

@@ -62,6 +62,72 @@
 ### 該当のソースコード
+import numpy as np
+import matplotlib.pyplot as plt
+class RBFkernel:
+    def __init__(self,*param):
+        self.param = list(param)
+    def __call__(self,x1,x2):
+        a,s,w = self.param
+        return a**2*np.exp(-((x1-x2)/s)**2) + w*(x1==x2)
+x0 = (0,5,10,20,25,33,50)
+y0 = (156,152,148,135,156,145,158)
+x1 = np.linspace(0,50,201)
+kernel = RBFkernel(10,5,3) # 適当なパラメータを使うカーネル関数
+k00 = kernel(*np.meshgrid(x0,x0))
+k00_1 = np.linalg.inv(k00) # 逆行列
+k01 = kernel(*np.meshgrid(x0,x1,indexing='ij'))
+k10 = k01.T
+k11 = kernel(*np.meshgrid(x1,x1))
+# ここでは上述の方程式の通りのμとΣ
+mu = k10.dot(k00_1.dot(y0))
+sigma = k11 - k10.dot(k00_1.dot(k01))
+plt.scatter(x0,y0,c='#ff77aa')
+plt.plot(x1,mu,'g') # 推測された平均
+std = np.sqrt(sigma.diagonal()) # 各点の標準偏差は共分散行列の対角成分
+plt.fill_between(x1,mu-std,mu+std,alpha=0.2,color='g') # 推測された標準偏差の中の領域
+plt.show()
 class Kernel:
     def __init__(self,param,bound=None):

コードとエラーメッセージ補足

2020/09/30 06:29

投稿

mizu103

スコア1

test CHANGED Viewed

File without changes

test CHANGED Viewed

@@ -8,13 +8,207 @@
 ### 発生している問題・エラーメッセージ
+---------------------------------------------------------------------------
+AttributeError                            Traceback (most recent call last)
+<ipython-input-63-3d592025a82f> in <module>()
+     11 for i in [0,1]:
+     12     if(i):
+---> 13         gp.saitekika(x0,y0,1000) # パラメータを調整する
+     14     plt.subplot(211+i)
+     15     plt.plot(x0,y0,'. ')
+1 frames
+<ipython-input-62-54f0617fec3e> in saitekika(self, x0, y0, kurikaeshi)
+     49         n_param = len(param)
+     50         theta0 = np.log(param)
+---> 51         p0 = self.logyuudo(param)
+     52         lis_theta = []
+     53         lis_p = []
+<ipython-input-62-54f0617fec3e> in logyuudo(self, param)
+     39             k00 = self.kernel(*np.meshgrid(self.x0,self.x0))
+     40             k00_1 = np.linalg.inv(k00)
+---> 41         return -(np.linalg.slogdet(k00)[1]+self.y0.dot(k00_1.dot(self.y0)))
+     42
+     43     def saitekika(self,x0,y0,kurikaeshi=1000): # パラメータを調整して学習
-https://qiita.com/phyblas/items/d756803ec932ab621c56
+AttributeError: 'tuple' object has no attribute 'dot'
 ### 該当のソースコード
+class Kernel:
+    def __init__(self,param,bound=None):
+        self.param = np.array(param)
+        if(bound==None):
+            bound = np.zeros([len(param),2])
+            bound[:,1] = np.inf
+        self.bound = np.array(bound)
+    def __call__(self,x1,x2):
+        a,s,w = self.param
+        return a**2*np.exp(-0.5*((x1-x2)/s)**2) + w**2*(x1==x2)
+class Gausskatei:
+    def __init__(self,kernel):
+        self.kernel = kernel
+    def gakushuu(self,x0,y0): # パラメータを調整せず学習
+        self.x0 = x0
+        self.y0 = y0
+        self.k00 = self.kernel(*np.meshgrid(x0,x0))
+        self.k00_1 = np.linalg.inv(self.k00)
+    def yosoku(self,x): # xからyを予測
+        k00_1 = self.k00_1
+        k01 = self.kernel(*np.meshgrid(self.x0,x,indexing='ij'))
+        k10 = k01.T
+        k11 = self.kernel(*np.meshgrid(x,x))
+        mu = k10.dot(k00_1.dot(self.y0))
+        sigma = k11 - k10.dot(k00_1.dot(k01))
+        std = np.sqrt(sigma.diagonal())
+        return mu,std
+    def logyuudo(self,param=None): # 対数尤度
+        if(param is None):
+            k00 = self.k00
+            k00_1 = self.k00_1
+        else:
+            self.kernel.param = param
+            k00 = self.kernel(*np.meshgrid(self.x0,self.x0))
+            k00_1 = np.linalg.inv(k00)
+        return -(np.linalg.slogdet(k00)[1]+self.y0.dot(k00_1.dot(self.y0)))
-上記urlの「カーネルとガウス過程のクラスを定義します。」の箇所は共通です。それ以降のコードを記します。
+    def saitekika(self,x0,y0,kurikaeshi=1000): # パラメータを調整して学習
+        self.x0 = x0
+        self.y0 = y0
+        param = self.kernel.param
+        logbound = np.log(self.kernel.bound)
+        s = (logbound[:,1]-logbound[:,0])/10.
+        n_param = len(param)
+        theta0 = np.log(param)
+        p0 = self.logyuudo(param)
+        lis_theta = []
+        lis_p = []
+        for i in range(kurikaeshi):
+            idou = np.random.normal(0,s,n_param)
+            hazure = (theta0+idou<logbound[:,0])|(theta0+idou>logbound[:,1])
+            while(np.any(hazure)):
+                idou[hazure] = np.random.normal(0,s,n_param)[hazure]
+                hazure = (theta0+idou<logbound[:,0])|(theta0+idou>logbound[:,1])
+            theta1 = theta0 + idou
+            param = np.exp(theta1)
+            p1 = self.logyuudo(param)
+            r = np.exp(p1-p0)
+            if(r>=1 or r>np.random.random()):
+                theta0 = theta1
+                p0 = p1
+                lis_theta.append(theta0)
+                lis_p.append(p0)
+        self.ar_theta = np.array(lis_theta)
+        self.ar_p = np.array(lis_p)
+        self.kernel.param = np.exp(lis_theta[np.argmax(lis_p)])
+        self.k00 = self.kernel(*np.meshgrid(x0,x0))
+        self.k00_1 = np.linalg.inv(self.k00)