|
a |
|
b/data/warps.py |
|
|
1 |
import numpy as np |
|
|
2 |
import matplotlib.pyplot as plt |
|
|
3 |
import pandas as pd |
|
|
4 |
|
|
|
5 |
import seaborn as sns |
|
|
6 |
import sys |
|
|
7 |
|
|
|
8 |
sys.path.append("../..") |
|
|
9 |
|
|
|
10 |
|
|
|
11 |
# from gp_functions import rbf_covariance |
|
|
12 |
from gpsa.util import rbf_kernel_numpy as rbf_covariance |
|
|
13 |
from gpsa import polar_warp |
|
|
14 |
from scipy.stats import multivariate_normal as mvnpy |
|
|
15 |
|
|
|
16 |
|
|
|
17 |
def apply_gp_warp( |
|
|
18 |
X_orig_single, |
|
|
19 |
Y_orig_single, |
|
|
20 |
n_views, |
|
|
21 |
# n_samples_per_view, |
|
|
22 |
noise_variance=0.0, |
|
|
23 |
kernel_variance=1.0, |
|
|
24 |
kernel_lengthscale=1.0, |
|
|
25 |
mean_slope=1.0, |
|
|
26 |
mean_intercept=0.0, |
|
|
27 |
): |
|
|
28 |
|
|
|
29 |
n_samples_per_view = X_orig_single.shape[0] |
|
|
30 |
n_spatial_dims = X_orig_single.shape[1] |
|
|
31 |
|
|
|
32 |
kernel = rbf_covariance |
|
|
33 |
kernel_params_true = np.array([np.log(1.0), np.log(1.0)]) |
|
|
34 |
|
|
|
35 |
X_orig = np.concatenate([X_orig_single.copy()] * n_views, axis=0) |
|
|
36 |
|
|
|
37 |
n_samples_list = [n_samples_per_view] * n_views |
|
|
38 |
cumulative_sums = np.cumsum(n_samples_list) |
|
|
39 |
cumulative_sums = np.insert(cumulative_sums, 0, 0) |
|
|
40 |
view_idx = np.array( |
|
|
41 |
[ |
|
|
42 |
np.arange(cumulative_sums[ii], cumulative_sums[ii + 1]) |
|
|
43 |
for ii in range(n_views) |
|
|
44 |
] |
|
|
45 |
) |
|
|
46 |
n = np.sum(n_samples_list) |
|
|
47 |
|
|
|
48 |
X = X_orig.copy() |
|
|
49 |
|
|
|
50 |
# Draw warped coordinates from a GP |
|
|
51 |
warp_kernel_params_true = np.array( |
|
|
52 |
[np.log(kernel_variance), np.log(kernel_lengthscale)] |
|
|
53 |
) |
|
|
54 |
|
|
|
55 |
for vv in range(n_views): |
|
|
56 |
|
|
|
57 |
for ss in range(n_spatial_dims): |
|
|
58 |
X_curr_view_warped = mvnpy.rvs( |
|
|
59 |
mean=X_orig_single[:, ss] * mean_slope + mean_intercept, |
|
|
60 |
cov=kernel(X_orig_single, X_orig_single, warp_kernel_params_true), |
|
|
61 |
) |
|
|
62 |
# import ipdb; ipdb.set_trace() |
|
|
63 |
X[ |
|
|
64 |
n_samples_per_view * vv : n_samples_per_view * (vv + 1), ss |
|
|
65 |
] = X_curr_view_warped |
|
|
66 |
|
|
|
67 |
Y = np.concatenate([Y_orig_single] * n_views, axis=0) |
|
|
68 |
Y += np.random.normal(scale=np.sqrt(noise_variance), size=Y.shape) |
|
|
69 |
|
|
|
70 |
return X, Y, n_samples_list, view_idx |
|
|
71 |
|
|
|
72 |
|
|
|
73 |
def apply_gp_warp_multimodal( |
|
|
74 |
X_orig_singles, |
|
|
75 |
Y_orig_singles, |
|
|
76 |
n_views, |
|
|
77 |
# n_samples_per_view, |
|
|
78 |
noise_variance=0.0, |
|
|
79 |
kernel_variance=1.0, |
|
|
80 |
kernel_lengthscale=1.0, |
|
|
81 |
mean_slope=1.0, |
|
|
82 |
mean_intercept=0.0, |
|
|
83 |
): |
|
|
84 |
assert len(X_orig_singles) == len(Y_orig_singles) |
|
|
85 |
|
|
|
86 |
n_modalities = len(X_orig_singles) |
|
|
87 |
|
|
|
88 |
modality_idx = np.cumsum([x.shape[0] for x in X_orig_singles]) |
|
|
89 |
modality_idx = np.insert(modality_idx, 0, 0) |
|
|
90 |
|
|
|
91 |
kernel = rbf_covariance |
|
|
92 |
kernel_params_true = np.array([np.log(1.0), np.log(1.0)]) |
|
|
93 |
|
|
|
94 |
X_orig_single = np.concatenate(X_orig_singles, axis=0) |
|
|
95 |
|
|
|
96 |
X_orig_single = X_orig_single - X_orig_single.min(0) |
|
|
97 |
X_orig_single = X_orig_single / X_orig_single.max(0) |
|
|
98 |
X_orig_single *= 10 |
|
|
99 |
|
|
|
100 |
X_orig = np.concatenate([X_orig_single.copy()] * n_views, axis=0) |
|
|
101 |
|
|
|
102 |
n_samples_per_view = X_orig_single.shape[0] |
|
|
103 |
n_spatial_dims = X_orig_single.shape[1] |
|
|
104 |
|
|
|
105 |
n_samples_list = [n_samples_per_view] * n_views |
|
|
106 |
cumulative_sums = np.cumsum(n_samples_list) |
|
|
107 |
cumulative_sums = np.insert(cumulative_sums, 0, 0) |
|
|
108 |
view_idx = np.array( |
|
|
109 |
[ |
|
|
110 |
np.arange(cumulative_sums[ii], cumulative_sums[ii + 1]) |
|
|
111 |
for ii in range(n_views) |
|
|
112 |
] |
|
|
113 |
) |
|
|
114 |
n = np.sum(n_samples_list) |
|
|
115 |
|
|
|
116 |
X = X_orig.copy() |
|
|
117 |
|
|
|
118 |
# Draw warped coordinates from a GP |
|
|
119 |
warp_kernel_params_true = np.array( |
|
|
120 |
[np.log(kernel_variance), np.log(kernel_lengthscale)] |
|
|
121 |
) |
|
|
122 |
|
|
|
123 |
for vv in range(n_views): |
|
|
124 |
|
|
|
125 |
curr_idx = np.arange(n_samples_per_view * vv, n_samples_per_view * (vv + 1)) |
|
|
126 |
|
|
|
127 |
for ss in range(n_spatial_dims): |
|
|
128 |
X_curr_view_warped = mvnpy.rvs( |
|
|
129 |
mean=X_orig_single[:, ss] * mean_slope + mean_intercept, |
|
|
130 |
cov=kernel(X_orig_single, X_orig_single, warp_kernel_params_true), |
|
|
131 |
) |
|
|
132 |
# import ipdb; ipdb.set_trace() |
|
|
133 |
X[curr_idx, ss] = X_curr_view_warped |
|
|
134 |
|
|
|
135 |
view_idx = np.cumsum([n_samples_per_view * vv for vv in range(n_views + 1)]) |
|
|
136 |
|
|
|
137 |
X_warped = [] |
|
|
138 |
Y_warped = [] |
|
|
139 |
n_samples_list = [] |
|
|
140 |
for mm in range(n_modalities): |
|
|
141 |
|
|
|
142 |
curr_modality_idx = np.concatenate( |
|
|
143 |
[ |
|
|
144 |
view_idx[vv] + np.arange(modality_idx[mm], modality_idx[mm + 1]) |
|
|
145 |
for vv in range(n_views) |
|
|
146 |
] |
|
|
147 |
) |
|
|
148 |
X_warped.append(X[curr_modality_idx]) |
|
|
149 |
|
|
|
150 |
Y_full_mm = np.concatenate([Y_orig_singles[mm]] * n_views, axis=0) |
|
|
151 |
Y_full_mm += np.random.normal( |
|
|
152 |
scale=np.sqrt(noise_variance), size=Y_full_mm.shape |
|
|
153 |
) |
|
|
154 |
Y_warped.append(Y_full_mm) |
|
|
155 |
n_samples_list.append([X_orig_singles[mm].shape[0]] * n_views) |
|
|
156 |
|
|
|
157 |
return X_warped, Y_warped, n_samples_list, view_idx |
|
|
158 |
|
|
|
159 |
|
|
|
160 |
def apply_linear_warp( |
|
|
161 |
X_orig_single, |
|
|
162 |
Y_orig_single, |
|
|
163 |
n_views, |
|
|
164 |
linear_slope_variance=0.1, |
|
|
165 |
linear_intercept_variance=0.1, |
|
|
166 |
noise_variance=0.01, |
|
|
167 |
rotation=True, |
|
|
168 |
): |
|
|
169 |
|
|
|
170 |
n_samples_per_view = X_orig_single.shape[0] |
|
|
171 |
n_spatial_dims = X_orig_single.shape[1] |
|
|
172 |
|
|
|
173 |
kernel = rbf_covariance |
|
|
174 |
kernel_params_true = np.array([np.log(1.0), np.log(1.0)]) |
|
|
175 |
|
|
|
176 |
X_orig = np.concatenate([X_orig_single.copy()] * n_views, axis=0) |
|
|
177 |
|
|
|
178 |
n_samples_list = [n_samples_per_view] * n_views |
|
|
179 |
cumulative_sums = np.cumsum(n_samples_list) |
|
|
180 |
cumulative_sums = np.insert(cumulative_sums, 0, 0) |
|
|
181 |
view_idx = np.array( |
|
|
182 |
[ |
|
|
183 |
np.arange(cumulative_sums[ii], cumulative_sums[ii + 1]) |
|
|
184 |
for ii in range(n_views) |
|
|
185 |
] |
|
|
186 |
) |
|
|
187 |
n = np.sum(n_samples_list) |
|
|
188 |
|
|
|
189 |
X = X_orig.copy() |
|
|
190 |
|
|
|
191 |
for vv in range(n_views): |
|
|
192 |
|
|
|
193 |
# curr_slopes = np.eye(n_spatial_dims) + np.random.normal( |
|
|
194 |
# loc=0, |
|
|
195 |
# scale=np.sqrt(linear_slope_variance), |
|
|
196 |
# size=(n_spatial_dims, n_spatial_dims), |
|
|
197 |
# ) |
|
|
198 |
# import ipdb; ipdb.set_trace() |
|
|
199 |
# curr_slopes /= np.linalg.norm(curr_slopes, ord=2, axis=0) |
|
|
200 |
# curr_slopes = np.linalg.svd(curr_slopes)[0] |
|
|
201 |
|
|
|
202 |
# curr_slopes = np.random.normal( |
|
|
203 |
# loc=1, |
|
|
204 |
# scale=np.sqrt(linear_slope_variance), |
|
|
205 |
# size=n_spatial_dims, |
|
|
206 |
# ) |
|
|
207 |
|
|
|
208 |
# curr_intercepts = np.random.normal( |
|
|
209 |
# loc=0, scale=np.sqrt(linear_intercept_variance), size=n_spatial_dims |
|
|
210 |
# ) |
|
|
211 |
|
|
|
212 |
curr_slopes = np.random.uniform( |
|
|
213 |
low=1 - linear_slope_variance, |
|
|
214 |
high=1 + linear_slope_variance, |
|
|
215 |
size=n_spatial_dims, |
|
|
216 |
) |
|
|
217 |
|
|
|
218 |
curr_intercepts = np.random.uniform( |
|
|
219 |
low=linear_intercept_variance, |
|
|
220 |
high=linear_intercept_variance, |
|
|
221 |
size=n_spatial_dims, |
|
|
222 |
) |
|
|
223 |
# print(curr_slopes, curr_intercepts) |
|
|
224 |
|
|
|
225 |
X_curr_view_warped = X_orig_single * curr_slopes + curr_intercepts |
|
|
226 |
X[ |
|
|
227 |
n_samples_per_view * vv : n_samples_per_view * (vv + 1), : |
|
|
228 |
] = X_curr_view_warped |
|
|
229 |
|
|
|
230 |
Y = np.concatenate([Y_orig_single] * n_views, axis=0) |
|
|
231 |
Y += np.random.normal(scale=np.sqrt(noise_variance), size=Y.shape) |
|
|
232 |
|
|
|
233 |
return X, Y, n_samples_list, view_idx |
|
|
234 |
|
|
|
235 |
|
|
|
236 |
def apply_polar_warp( |
|
|
237 |
X_orig_single, |
|
|
238 |
Y_orig_single, |
|
|
239 |
n_views, |
|
|
240 |
linear_slope_variance=0.1, |
|
|
241 |
linear_intercept_variance=0.1, |
|
|
242 |
noise_variance=0.01, |
|
|
243 |
rotation=True, |
|
|
244 |
): |
|
|
245 |
|
|
|
246 |
n_samples_per_view = X_orig_single.shape[0] |
|
|
247 |
n_spatial_dims = X_orig_single.shape[1] |
|
|
248 |
|
|
|
249 |
kernel = rbf_covariance |
|
|
250 |
kernel_params_true = np.array([np.log(1.0), np.log(1.0)]) |
|
|
251 |
|
|
|
252 |
X_orig = np.concatenate([X_orig_single.copy()] * n_views, axis=0) |
|
|
253 |
|
|
|
254 |
n_samples_list = [n_samples_per_view] * n_views |
|
|
255 |
cumulative_sums = np.cumsum(n_samples_list) |
|
|
256 |
cumulative_sums = np.insert(cumulative_sums, 0, 0) |
|
|
257 |
view_idx = np.array( |
|
|
258 |
[ |
|
|
259 |
np.arange(cumulative_sums[ii], cumulative_sums[ii + 1]) |
|
|
260 |
for ii in range(n_views) |
|
|
261 |
] |
|
|
262 |
) |
|
|
263 |
n = np.sum(n_samples_list) |
|
|
264 |
|
|
|
265 |
X = X_orig.copy() |
|
|
266 |
|
|
|
267 |
# no_distortion_B = np.array([ |
|
|
268 |
# [0, np.pi * 0.5], |
|
|
269 |
# [0, 0] |
|
|
270 |
# ]) |
|
|
271 |
|
|
|
272 |
for vv in range(n_views): |
|
|
273 |
|
|
|
274 |
# B = np.random.normal( |
|
|
275 |
# loc=0, |
|
|
276 |
# scale=np.sqrt(linear_slope_variance), |
|
|
277 |
# size=(n_spatial_dims, n_spatial_dims), |
|
|
278 |
# ) |
|
|
279 |
B = np.random.uniform( |
|
|
280 |
low=-linear_slope_variance, |
|
|
281 |
high=linear_slope_variance, |
|
|
282 |
size=(n_spatial_dims, n_spatial_dims), |
|
|
283 |
) |
|
|
284 |
polar_params = X_orig_single @ B |
|
|
285 |
r, theta = polar_params[:, 0], polar_params[:, 1] |
|
|
286 |
X_curr_view_warped = np.array( |
|
|
287 |
[ |
|
|
288 |
X_orig_single[:, 0] + r * np.cos(theta), |
|
|
289 |
X_orig_single[:, 1] + r * np.sin(theta), |
|
|
290 |
] |
|
|
291 |
).T |
|
|
292 |
# import ipdb; ipdb.set_trace() |
|
|
293 |
# additive_warp = B[:, 0] * X_orig_single * np.vstack([np.cos(X_orig_single[:, 0] * B[0, 1]), np.sin(X_orig_single[:, 1] * B[1, 1])]).T |
|
|
294 |
# X_curr_view_warped = X_orig_single + additive_warp |
|
|
295 |
|
|
|
296 |
# X_curr_view_warped = X_orig_single @ curr_slopes + curr_intercepts |
|
|
297 |
X[ |
|
|
298 |
n_samples_per_view * vv : n_samples_per_view * (vv + 1), : |
|
|
299 |
] = X_curr_view_warped |
|
|
300 |
|
|
|
301 |
Y = np.concatenate([Y_orig_single] * n_views, axis=0) |
|
|
302 |
Y += np.random.normal(scale=np.sqrt(noise_variance), size=Y.shape) |
|
|
303 |
|
|
|
304 |
return X, Y, n_samples_list, view_idx |
|
|
305 |
|
|
|
306 |
|
|
|
307 |
if __name__ == "__main__": |
|
|
308 |
|
|
|
309 |
pass |