kernel.py 10.9 KB
Newer Older
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
1
2
3
4
5
6
7
8
9
10
import numpy as np
from abc import ABC, abstractmethod

from scipy.spatial.distance import pdist, cdist, squareform

class kernel(ABC):
    def __init__(self):
        self._theta = None

    @abstractmethod
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
11
    def __call__(self):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
12
13
        pass

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
14
15
16
    @abstractmethod
    def kernel(self):
        pass
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
17

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
18
19
20
    @abstractmethod
    def kernel_vector(self):
        pass
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

    @abstractmethod
    def kernel_jacobian(self):
        pass

    @abstractmethod
    def kernel_hyperparameter_gradient(self):
        pass

    @property
    def theta(self):
        return self._theta

    @theta.setter
    def theta(self, theta):
        self._theta = theta
    
    def numerical_jacobian(self,x,y, dx=1.e-5):
        if np.ndim(y) == 1:
            y = y.reshape((1,-1))
        nx = len(x)
        ny = y.shape[0]
        f0 = self.kernel(x,y)
        f_jac = np.zeros((ny,nx))
        for i in range(nx):
            x_up = np.copy(x)
            x_down = np.copy(x)
            x_up[i] += 0.5*dx
            x_down[i] -= 0.5*dx
            
            f_up = self.kernel(x_up,y)
            f_down = self.kernel(x_down,y)
            f_jac[:,i] = (f_up - f_down)/dx
        return f_jac

    def numerical_hyperparameter_gradient(self,X, dx=1.e-5):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
57
58
59
        """Calculates the numerical derivative of the kernel with respect to the
        log transformed hyperparameters.
        """
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
60
61
62
63
64
65
66
67
68
69
70
        N_data = X.shape[0]
        theta = np.copy(self.theta)
        N_hyper = len(theta)
        dK_dTheta = np.zeros((N_hyper, N_data, N_data))
        for i in range(N_hyper):
            theta_up = np.copy(theta)
            theta_down = np.copy(theta)
            theta_up[i] += 0.5*dx
            theta_down[i] -= 0.5*dx
            
            self.theta = theta_up
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
71
            K_up = self(X, eval_gradient=False)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
72
            self.theta = theta_down
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
73
            K_down = self(X, eval_gradient=False)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
74
            dK_dTheta[i,:,:] = (K_up - K_down)/dx
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
75
        self.theta = theta
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
76
77
78
79
80
81
82
        return dK_dTheta


class gauss_kernel(kernel):
    def __init__(self, amplitude=10.0, length_scale=10.0, amplitude_bounds=(1e0, 1e3), length_scale_bounds=(1e-1, 1e1)):
        self.amplitude = amplitude
        self.length_scale = length_scale
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
83

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
84
85
86
        self.amplitude_bounds = amplitude_bounds
        self.length_scale_bounds = length_scale_bounds

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
87
88
89
90
91
92
93
94
95
96
97
98
99
100
        self.theta_bounds = np.array([amplitude_bounds, length_scale_bounds])

    def __call__(self, X, eval_gradient=False):
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
        d = cdist(X / self.length_scale,
                  X / self.length_scale, metric='sqeuclidean')
        K = self.amplitude * np.exp(-0.5 * d)
        
        if eval_gradient:
            K_gradient = self.kernel_hyperparameter_gradient(X)
            return K, K_gradient
        else:
            return K
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
101
        
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
102
    def kernel(self, X,Y=None):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
103
104
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
105
106
107
        if Y is None:
            Y = X
        elif np.ndim(Y) == 1:
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
108
109
110
111
112
113
            Y = Y.reshape((1,-1))
        d = cdist(X / self.length_scale,
                  Y / self.length_scale, metric='sqeuclidean')
        K = self.amplitude * np.exp(-0.5 * d)
        return K

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
114
115
116
117
    def kernel_value(self, x,y):
        K = self.kernel(x,y)
        return np.asscalar(K)
    
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
118
    def kernel_vector(self, x,Y):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
119
        K = self.kernel(x,Y).reshape(-1)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
        return K

    def kernel_jacobian(self, X,Y):
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
        if np.ndim(Y) == 1:
            Y = Y.reshape((1,-1))
        K = self.kernel(X,Y).T
        dK_dd = -1./(2*self.length_scale**2)*K
        dd_df = 2*(X - Y)

        dk_df = np.multiply(dK_dd, dd_df)
        return dk_df

    @property
    def theta(self):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
136
137
138
139
        """Returns the log-transformed hyperparameters of the kernel.
        """
        self._theta = np.array([self.amplitude, self.length_scale])
        return np.log(self._theta)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
140
141
142

    @theta.setter
    def theta(self, theta):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
143
144
145
146
147
        """Sets the hyperparameters of the kernel.

        theta: log-transformed hyperparameters
        """
        self._theta = np.exp(theta)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
148
149
        self.amplitude = self._theta[0]
        self.length_scale = self._theta[1]
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
150
    
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
151
152
153
154
155
    def dK_da(self, X):
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
        d = cdist(X / self.length_scale,
                  X / self.length_scale, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
156
        dK_da = self.amplitude * np.exp(-0.5 * d)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
157
158
159
160
161
162
163
        return dK_da
        
    def dK_dl(self, X):
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
        d = cdist(X / self.length_scale,
                  X / self.length_scale, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
164
        dK_dl = self.amplitude * d * np.exp(-0.5 * d)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
165
166
167
        return dK_dl

    def kernel_hyperparameter_gradient(self, X):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
168
169
170
        """Calculates the derivative of the kernel with respect to the
        log transformed hyperparameters.
        """
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
171
172
173
174
        return np.array([self.dK_da(X), self.dK_dl(X)])
    

class double_gauss_kernel(kernel):
175
    def __init__(self, amplitude=100., amplitude_bounds=(1e1,1e5),
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
176
177
178
                 length_scale1=10.0, length_scale1_bounds=(1e0, 1e3),
                 length_scale2=10.0, length_scale2_bounds=(1e0, 1e3),
                 weight=0.01, weight_bounds=(0.01,0.01),
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
179
180
181
                 noise=1e-5, noise_bounds=(1e-5,1e-5),
                 eta=1, eta_bounds=(0.1,10),
                 Nsplit_eta=None):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
182
183
184
185
        self.amplitude = amplitude
        self.length_scale1 = length_scale1
        self.length_scale2 = length_scale2
        self.weight = weight
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
186
        self.noise = noise
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
187
188
        self.eta = eta
        self.Nsplit_eta = Nsplit_eta
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
189
190
191
192
193

        self.amplitude_bounds = amplitude_bounds
        self.length_scale1_bounds = length_scale1_bounds
        self.length_scale2_bounds = length_scale2_bounds
        self.weight_bounds = weight_bounds
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
194
        self.noise_bounds = noise_bounds
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
195
196
197
198
        if self.Nsplit_eta is None:
            self.eta_bounds = (eta,eta)
        else:
            self.eta_bounds = eta_bounds
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
199

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
200
        self.theta_bounds = np.log(np.array([amplitude_bounds, length_scale1_bounds, length_scale2_bounds, weight_bounds, noise_bounds, self.eta_bounds]))
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
201

Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
202
    def __call__(self, X, eval_gradient=False):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
203
        K = self.kernel(X, with_noise=True)
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
204
205
206
207
208
        if eval_gradient:
            K_gradient = self.kernel_hyperparameter_gradient(X)
            return K, K_gradient
        else:
            return K
209

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
210
211
212
213
    def kernel(self, X, Y=None, with_noise=False):
        if with_noise:
            assert Y is None
        X = self.apply_eta(X)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
214
215
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
216
217
218
        if Y is None:
            Y = X
        elif np.ndim(Y) == 1:
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
219
220
221
222
223
            Y = Y.reshape((1,-1))
        d1 = cdist(X / self.length_scale1,
                  Y / self.length_scale1, metric='sqeuclidean')
        d2 = cdist(X / self.length_scale2,
                  Y / self.length_scale2, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
224
225
226
227
        if with_noise:
            K = self.amplitude * (np.exp(-0.5 * d1) + self.weight*np.exp(-0.5 * d2) + self.noise*np.eye(X.shape[0]))
        else:
            K = self.amplitude * (np.exp(-0.5 * d1) + self.weight*np.exp(-0.5 * d2))
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
228
229
        return K

Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
230
231
232
233
    def kernel_value(self, x,y):
        K = self.kernel(x,y)
        return np.asscalar(K)
    
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
234
    def kernel_vector(self, x,Y):
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
235
        K = self.kernel(x,Y).reshape(-1)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
236
237
238
239
240
        return K

    def kernel_jacobian(self, X,Y):
        """ Jacobian of the kernel with respect to X
        """
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
241
242
        X = self.apply_eta(X)
        Y = self.apply_eta(Y)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
        if np.ndim(X) == 1:
            X = X.reshape((1,-1))
        if np.ndim(Y) == 1:
            Y = Y.reshape((1,-1))
        d1 = cdist(X / self.length_scale1,
                   Y / self.length_scale1, metric='sqeuclidean')
        d2 = cdist(X / self.length_scale2,
                   Y / self.length_scale2, metric='sqeuclidean')
        dK1_dd1 = -1/(2*self.length_scale1**2) * np.exp(-0.5 * d1)
        dK2_dd2 = -1/(2*self.length_scale2**2) * np.exp(-0.5 * d2)
        dK_dd = self.amplitude * (dK1_dd1 + self.weight*dK2_dd2)
        dd_df = 2*(X - Y)

        dk_df = np.multiply(dK_dd.T, dd_df)
        return dk_df

    @property
    def theta(self):
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
261
262
        """Returns the log-transformed hyperparameters of the kernel.
        """
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
263
        self._theta = np.array([self.amplitude, self.length_scale1, self.length_scale2, self.weight, self.noise, self.eta])
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
264
        return np.log(self._theta)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
265
266
267

    @theta.setter
    def theta(self, theta):
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
268
269
270
271
272
        """Sets the hyperparameters of the kernel.

        theta: log-transformed hyperparameters
        """
        self._theta = np.exp(theta)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
273
274
275
276
        self.amplitude = self._theta[0]
        self.length_scale1 = self._theta[1]
        self.length_scale2 = self._theta[2]
        self.weight = self._theta[3]
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
277
        self.noise = self._theta[4]
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
278
279
280
281
282
283
284
285
286
287
        self.eta = self._theta[5]

    def apply_eta(self, X):
        Xeta = np.copy(X)
        if self.Nsplit_eta is not None:
            if np.ndim(X) == 1:
                Xeta[self.Nsplit_eta:] *= self.eta
            else:
                Xeta[:,self.Nsplit_eta:] *= self.eta
        return Xeta
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
288
289
290
291
292
293

    def dK_da(self, X):
        d1 = cdist(X / self.length_scale1,
                   X / self.length_scale1, metric='sqeuclidean')
        d2 = cdist(X / self.length_scale2,
                   X / self.length_scale2, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
294
        dK_da = self.amplitude * (np.exp(-0.5 * d1) + self.weight*np.exp(-0.5 * d2) + self.noise*np.eye(X.shape[0]))
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
295
296
297
298
299
        return dK_da
        
    def dK_dl1(self, X):
        d1 = cdist(X / self.length_scale1,
                   X / self.length_scale1, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
300
        dK_dl1 = self.amplitude*d1 * np.exp(-0.5 * d1)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
301
302
303
304
305
        return dK_dl1

    def dK_dl2(self, X):
        d2 = cdist(X / self.length_scale2,
                   X / self.length_scale2, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
306
        dK_dl2 = self.amplitude*self.weight*d2 * np.exp(-0.5 * d2)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
307
308
309
310
311
        return dK_dl2

    def dK_dw(self, X):
        d2 = cdist(X / self.length_scale2,
                   X / self.length_scale2, metric='sqeuclidean')
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
312
        dK_dl2 = self.amplitude*self.weight*np.exp(-0.5 * d2)
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
313
314
        return dK_dl2

Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
315
    def dK_dn(self, X):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
316
        dK_dn = self.amplitude * self.noise * np.eye(X.shape[0])
Malthe Kjær Bisbo's avatar
update    
Malthe Kjær Bisbo committed
317
318
        return dK_dn

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
    def dK_deta(self, X, dx=1e-5):
        N_data = X.shape[0]
        theta = np.copy(self.theta)
        dK_deta = np.zeros((N_data, N_data))

        theta_up = np.copy(theta)
        theta_down = np.copy(theta)
        theta_up[-1] += 0.5*dx
        theta_down[-1] -= 0.5*dx
        
        self.theta = theta_up
        K_up = self(X, eval_gradient=False)
        self.theta = theta_down
        K_down = self(X, eval_gradient=False)
        dK_dTheta = (K_up - K_down)/dx

        self.theta = theta
        return dK_dTheta

Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
338
    def kernel_hyperparameter_gradient(self, X):
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
339
340
341
        """Calculates the derivative of the kernel with respect to the
        log transformed hyperparameters.
        """
Malthe Kjær Bisbo's avatar
Malthe Kjær Bisbo committed
342
343
        X = self.apply_eta(X)
        return np.array([self.dK_da(X), self.dK_dl1(X), self.dK_dl2(X), self.dK_dw(X), self.dK_dn(X), self.dK_deta(X)])