-
Notifications
You must be signed in to change notification settings - Fork 2
/
deep_gaussian_process_5.py
333 lines (276 loc) · 15.6 KB
/
deep_gaussian_process_5.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
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
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
from __future__ import absolute_import
from __future__ import print_function
import matplotlib.pyplot as plt
import autograd.numpy as np
import autograd.numpy.random as npr
from autograd import value_and_grad,grad
from scipy.optimize import minimize
from autograd.numpy.linalg import solve
import autograd.scipy.stats.multivariate_normal as mvn
from builtins import range
from autograd.scipy.misc import logsumexp
import itertools
import sys
from optparse import OptionParser
from autograd.util import quick_grad_check
def build_checker_dataset(n_data = 6, noise_std =0.1):
rs = npr.RandomState(0)
inputs = np.array([np.array([x,y]) for x in np.linspace(-1,1,n_data) for y in np.linspace(-1,1,n_data)])
targets = np.sign([np.prod(input) for input in inputs]) + rs.randn(n_data**2)*noise_std
return inputs, targets
def build_step_function_dataset(D=1, n_data=40, noise_std=0.1):
rs = npr.RandomState(0)
inputs = np.linspace(-2, 2, num=n_data)
targets = np.sign(inputs) + rs.randn(n_data) * noise_std
inputs = inputs.reshape((len(inputs), D))
return inputs, targets
def rbf_covariance(cov_params, x, xp):
output_scale = np.exp(cov_params[0])
lengthscales = np.exp(cov_params[1:])
diffs = np.expand_dims(x /lengthscales, 1)\
- np.expand_dims(xp/lengthscales, 0)
return output_scale * np.exp(-0.5 * np.sum(diffs**2, axis=2))
def build_single_gp(cov_func, num_cov_params, num_pseudo_params, input_dimension):
"""Functions that perform Gaussian process regression.
cov_func has signature (cov_params, x, x')"""
def unpack_gp_params(params):
mean = params[0]
noise_scale = np.exp(params[1]) + 0.001
cov_params = params[2:2+num_cov_params]
pseudo_params = params[2+num_cov_params:]
x0, y0 = np.split(pseudo_params,[num_pseudo_params*input_dimension])
x0 = x0.reshape((num_pseudo_params,input_dimension))
gp_details = {'mean': mean, 'noise_scale': noise_scale, 'cov_params': cov_params, 'x0': x0, 'y0': y0}
#x0 = X
#y0 = y
return mean, cov_params, noise_scale, x0, y0
def pack_gp_params(mean, cov_params, noise_scale, x0, y0):
params = np.append(mean,noise_scale)
params = np.concatenate([params,cov_params])
params = np.concatenate([params,np.ndarray.flatten(np.array(x0))])
params = np.concatenate([params,np.ndarray.flatten(np.array(y0))])
return params
def predict(params, xstar):
"""Returns the predictive mean and covariance at locations xstar,
of the latent function value f (without observation noise)."""
mean, cov_params, noise_scale, x0, y0 = unpack_gp_params(params)
cov_f_f = cov_func(cov_params, xstar, xstar)
cov_y_f = cov_func(cov_params, x0, xstar)
cov_y_y = cov_func(cov_params, x0, x0) + noise_scale * np.eye(len(y0))
pred_mean = mean + np.dot(solve(cov_y_y, cov_y_f).T, y0 - mean)
pred_cov = cov_f_f - np.dot(solve(cov_y_y, cov_y_f).T, cov_y_f)
return pred_mean, pred_cov
def predict_with_noise(params, xstar):
pred_mean, pred_cov = predict(params, xstar)
mean, cov_params, noise_scale, x0, y0 = unpack_gp_params(params)
return pred_mean, pred_cov + noise_scale*np.eye(len(xstar))
num_gp_params = 2 + num_cov_params + num_pseudo_params*input_dimension + num_pseudo_params
return num_gp_params, predict, predict_with_noise, unpack_gp_params, pack_gp_params
def build_single_layer(input_dimension, output_dimension, num_pseudo_params, covariance_function, random):
layer_details = [build_single_gp(covariance_function, input_dimension + 1, num_pseudo_params, input_dimension) for i in xrange(output_dimension)]
num_params_each_output, predict_layer_funcs, predict_funcs_with_noise, unpack_gp_params_layer, pack_gp_params_layer = zip(*layer_details)
total_params_layer = sum(num_params_each_output)
def unpack_layer_params(params):
gp_params = np.array_split(params, output_dimension) # assuming all parameters have equal dims, change to what we had below
return gp_params
def sample_from_mvn(mu, sigma):
rs = npr.RandomState(0)
return np.dot(np.linalg.cholesky(sigma+1e-6*np.eye(len(sigma))*np.max(np.diag(sigma))),rs.randn(len(sigma)))+mu if random == 1 else mu
def sample_mean_cov_from_layer(layer_params, xstar, with_noise = False):
predict = predict_funcs_with_noise if with_noise else predict_layer_funcs
gp_params = unpack_layer_params(layer_params)
samples = [predict[i](gp_params[i],xstar) for i in xrange(output_dimension)]
return samples
def sample_values_from_layer(layer_params, xstar, with_noise = False): # should return len(x*)
samples = sample_mean_cov_from_layer(layer_params, xstar, with_noise)
outputs = [sample_from_mvn(mean,cov) for mean,cov in samples]
return np.array(outputs).T
return total_params_layer, sample_mean_cov_from_layer, sample_values_from_layer, predict_layer_funcs, unpack_gp_params_layer, unpack_layer_params, pack_gp_params_layer
def build_deep_gp(dimensions, covariance_function, num_pseudo_params, random):
deep_details = [build_single_layer(dimensions[i],dimensions[i+1],num_pseudo_params,covariance_function,random) for i in xrange(len(dimensions)-1)]
num_params_each_layer, sample_mean_cov_funcs, sample_value_funcs, predict_layer_funcs, unpack_gp_params_all, unpack_layer_params, pack_gp_params_all = zip(*deep_details)
total_params_gp = sum(num_params_each_layer)
def unpack_all_params(all_params):
all_layer_params = np.array_split(all_params,np.cumsum(num_params_each_layer))
return all_layer_params
def sample_mean_cov_from_deep_gp(all_params, xstar, with_noise = False):
xtilde = xstar
all_layer_params = unpack_all_params(all_params)
for layer in xrange(len(dimensions)-2):
layer_params = all_layer_params[layer]
xtilde = sample_value_funcs[layer](layer_params, xtilde, with_noise)
final_layer = len(dimensions)-2
final_layer_params = all_layer_params[final_layer]
final_mean, final_cov = sample_mean_cov_funcs[final_layer](final_layer_params, xtilde, with_noise)[0] # index into 0 because final layer has one unit
#print("Function Means",final_mean)
return final_mean, final_cov
def evaluate_prior(all_params): # clean up code so we don't compute matrices twice
all_layer_params = unpack_all_params(all_params)
log_prior = 0
for layer in xrange(n_layers):
layer_params = all_layer_params[layer]
layer_gp_params = unpack_layer_params[layer](layer_params)
for dim in xrange(dimensions[layer+1]):
gp_params = layer_gp_params[dim]
mean, cov_params, noise_scale, x0, y0 = unpack_gp_params_all[layer][dim](gp_params)
cov_y_y = covariance_function(cov_params,x0,x0) + noise_scale * np.eye(len(y0))
log_prior += mvn.logpdf(y0,np.ones(len(cov_y_y))*mean,cov_y_y+np.eye(len(cov_y_y))*10)
return log_prior
def log_likelihood(all_params):
samples = [sample_mean_cov_from_deep_gp(all_params, X, True) for i in xrange(n_samples)]
return logsumexp(np.array([mvn.logpdf(y,mean,var+1e-6*np.eye(len(var))*np.max(np.diag(var))) for mean,var in samples])) - np.log(n_samples) \
+ evaluate_prior(all_params)
def squared_error(all_params):
samples = np.array([sample_mean_cov_from_deep_gp(all_params, X, False)[0] for i in xrange(n_samples)])
return np.mean((y - np.mean(samples,axis = 0)) ** 2)
return total_params_gp, log_likelihood, sample_mean_cov_from_deep_gp, predict_layer_funcs, squared_error, unpack_gp_params_all, unpack_layer_params, unpack_all_params,\
pack_gp_params_all
if __name__ == '__main__':
random = 0
n_samples = 1
dimensions = [2,1] # Architecture of the GP. Last layer should always be 1
n_data = 20
input_dimension = dimensions[0]
n_layers = len(dimensions)-1
num_pseudo_params = 50
#X, y = build_step_function_dataset(D=input_dimension, n_data=20)
X, y = build_checker_dataset(n_data=16)
total_num_params, log_likelihood, sample_mean_cov_from_deep_gp, predict_layer_funcs, squared_error, unpack_gp_params_all, unpack_layer_params, unpack_all_params, \
pack_gp_params_all = build_deep_gp(dimensions, rbf_covariance, num_pseudo_params, random)
# Set up figure.
if dimensions[0] == 1:
fig = plt.figure(figsize=(12,8), facecolor='white')
ax_first = fig.add_subplot(411, frameon=False)
ax_end_to_end = fig.add_subplot(412, frameon=False)
ax_x_to_h = fig.add_subplot(413, frameon=False)
ax_h_to_y = fig.add_subplot(414, frameon=False)
plt.show(block=False)
else:
fig = plt.figure(figsize=(12,8), facecolor='white')
ax = fig.add_subplot(111, frameon=False)
plt.show(block=False)
def plot_deep_gp_2d(ax,params,plot_xs):
ax.cla()
rs = npr.RandomState(0)
sampled_means_and_covs = [sample_mean_cov_from_deep_gp(params, plot_xs) for i in xrange(n_samples)]
sampled_means, sampled_covs = zip(*sampled_means_and_covs)
avg_pred_mean = np.mean(sampled_means, axis = 0)
avg_pred_cov = np.mean(sampled_covs, axis = 0)
#print("X*",avg_pred_mean)
#rint("X*",plot_xs[0:4])
#sampled_means_and_covs_orig = [sample_mean_cov_from_deep_gp(params, X) for i in xrange(n_samples)]
#sampled_means_orig, sampled_covs_orig = zip(*sampled_means_and_covs_orig)
#avg_pred_mean_orig = np.mean(sampled_means_orig, axis = 0)
#print("Orignal Xs",avg_pred_mean_orig)
X0 = params[5:5+num_pseudo_params*2].reshape(num_pseudo_params,2)
y0 = params[5+num_pseudo_params*2:5+num_pseudo_params*3]
#ax.scatter(X0[:,0],X0[:,1],c = y0)
avg_pred_mean = avg_pred_mean.reshape(40,40)
ax.contourf(np.linspace(-1,1,40),np.linspace(-1,1,40), avg_pred_mean)
ax.scatter(X[:,0],X[:,1],c=y)
ax.set_xticks([])
ax.set_yticks([])
ax.set_title("Full Deep GP")
def plot_deep_gp(ax, params, plot_xs):
ax.cla()
rs = npr.RandomState(0)
sampled_means_and_covs = [sample_mean_cov_from_deep_gp(params, plot_xs) for i in xrange(n_samples)]
sampled_means, sampled_covs = zip(*sampled_means_and_covs)
avg_pred_mean = np.mean(sampled_means, axis = 0)
avg_pred_cov = np.mean(sampled_covs, axis = 0)
marg_std = np.sqrt(np.diag(avg_pred_cov))
if n_samples > 1:
ax.fill(np.concatenate([plot_xs, plot_xs[::-1]]),
np.concatenate([avg_pred_mean - 1.96 * marg_std,
(avg_pred_mean + 1.96 * marg_std)[::-1]]),
alpha=.15, fc='Blue', ec='None')
ax.plot(plot_xs, avg_pred_mean, 'b')
sampled_funcs = np.array([rs.multivariate_normal(mean, cov*(random)) for mean,cov in sampled_means_and_covs])
ax.plot(plot_xs,sampled_funcs.T)
ax.plot(X, y, 'kx')
#ax.set_ylim([-1.5,1.5])
ax.set_xticks([])
ax.set_yticks([])
ax.set_title("Full Deep GP, inputs to outputs")
def plot_single_gp(ax, x0, y0, pred_mean, pred_cov, plot_xs):
ax.cla()
marg_std = np.sqrt(np.diag(pred_cov))
if n_samples > 1:
ax.plot(plot_xs, pred_mean, 'b')
ax.fill(np.concatenate([plot_xs, plot_xs[::-1]]),
np.concatenate([pred_mean - 1.96 * marg_std,
(pred_mean + 1.96 * marg_std)[::-1]]),
alpha=.15, fc='Blue', ec='None')
# Show samples from posterior.
rs = npr.RandomState(0)
sampled_funcs = rs.multivariate_normal(pred_mean, pred_cov*(random), size=n_samples)
ax.plot(plot_xs, sampled_funcs.T)
ax.plot(x0, y0, 'ro')
ax.set_xticks([])
ax.set_yticks([])
def callback(params):
print("Log likelihood {}, Squared Error {}".format(-objective(params),squared_error(params)))
# Show posterior marginals.
if dimensions[0] == 1:
plot_xs = np.reshape(np.linspace(-5, 5, 300), (300,1))
plot_deep_gp(ax_first, init_params, plot_xs)
ax_first.set_title("Initial full predictions")
plot_deep_gp(ax_end_to_end, params, plot_xs)
if dimensions == [1,1]:
ax_end_to_end.plot(params[4:14],params[14:24], 'ro')
elif dimensions == [1,1,1]:
hidden_mean, hidden_cov = predict_layer_funcs[0][0](params[0:24], plot_xs)
plot_single_gp(ax_x_to_h, params[4:14], params[14:24], hidden_mean, hidden_cov, plot_xs)
ax_x_to_h.set_title("Inputs to hiddens, inducing points in red")
y_mean, y_cov = predict_layer_funcs[0][0](params[24:48], plot_xs)
plot_single_gp(ax_h_to_y, params[28:38], params[38:48], y_mean, y_cov, plot_xs)
ax_h_to_y.set_title("Hiddens to outputs, inducing points in red")
plt.draw()
plt.pause(1.0/60.0)
elif dimensions[0] == 2:
plot_xs = np.array([np.array([a,b]) for a in np.linspace(-1,1,40) for b in np.linspace(-1,1,40)])
plot_deep_gp_2d(ax, params, plot_xs)
#all_layer_params = unpack_all_params(init_params)
#for layer in xrange(n_layers):
# layer_params = all_layer_params[layer]
# layer_gp_params = unpack_layer_params[layer](layer_params)
# for dim in xrange(dimensions[layer+1]):
# gp_params = layer_gp_params[dim]
# mean, cov_params, noise_scale, x0, y0 = unpack_gp_params_all[layer][dim](gp_params)
# lengthscales = cov_params[1:]
##ax.scatter(X[:,0],x0[:,0])
plt.draw()
plt.pause(1.0/60.0)
rs = npr.RandomState(1234)
init_params = .1 * rs.randn(total_num_params)
# SMART INITIALIZATION
# Admittedly, this is not a good way to do it
# If you have any tips on how to make this better please let me know
smart_params = np.array([])
all_layer_params = unpack_all_params(init_params)
for layer in xrange(n_layers):
layer_params = all_layer_params[layer]
layer_gp_params = unpack_layer_params[layer](layer_params)
for dim in xrange(dimensions[layer+1]):
gp_params = layer_gp_params[dim]
mean, cov_params, noise_scale, x0, y0 = unpack_gp_params_all[layer][dim](gp_params)
lengthscales = cov_params[1:]
if layer == 0:
pairs = itertools.combinations(X, 2)
dists = np.array([np.abs(p1-p2) for p1,p2 in pairs])
smart_lengthscales = np.array([np.log(np.median(dists[:,i])) for i in xrange(len(lengthscales))])
smart_x0 = np.array(X)[rs.choice(len(X), num_pseudo_params, replace=False),:]
smart_y0 = np.ndarray.flatten(smart_x0)
else:
smart_x0 = x0
smart_y0 = np.ndarray.flatten(x0)
smart_lengthscales = np.array([np.log(1) for i in xrange(len(lengthscales))])
cov_params = np.append(cov_params[0],smart_lengthscales)
params = pack_gp_params_all[layer][dim](mean, cov_params, noise_scale, smart_x0, smart_y0)
smart_params = np.append(smart_params, params)
init_params = smart_params
print("Optimizing covariance parameters...")
objective = lambda params: -log_likelihood(params)
params = minimize(value_and_grad(objective), init_params, jac=True,
method='BFGS', callback=callback)
plt.pause(10.0)