-
Notifications
You must be signed in to change notification settings - Fork 1
/
script_held_out_prediction.py
443 lines (368 loc) · 20.1 KB
/
script_held_out_prediction.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
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
import numpy as np
import pandas
from scipy.interpolate import interp1d
from data_generator import generate_spikes_time_series as generate_experiment
from data_generator import make_design_matrix_hrf
from nistats.hemodynamic_models import _gamma_difference_hrf
from sklearn.gaussian_process.kernels import RBF
n_events=90
n_blank_events=15
event_spacing=6
t_r=2
hrf_length=32.
event_types=['ev1', 'ev2']
jitter_min=-1
jitter_max=1
time_offset = 20
modulation=None
n_runs = 2
hrf_peak_locations = np.array([3, 4, 5, 6, 7, 8])
paradigms, _, _, measurement_times = list(zip(*[
generate_experiment(n_events=n_events,
n_blank_events=n_blank_events,
event_spacing=event_spacing,
t_r=t_r, hrf_length=hrf_length,
event_types=event_types,
jitter_min=jitter_min,
jitter_max=jitter_max,
return_jitter=True,
time_offset=time_offset,
modulation=modulation,
seed=seed) for seed in np.arange(n_runs)]))
rng = np.random.RandomState(42)
noise_levels = np.array([0., .1, 1., 2., 5., 10.])
beta = rng.randn(len(event_types))
n_new_betas = 4
new_betas = rng.randn(len(event_types), n_new_betas) # these are to see how well the hrf spans the space
frame_times_run = np.arange(0, paradigms[0]['onset'].max() + time_offset, t_r)
n_noises = 2
noise_vectors = [rng.randn(len(frame_times_run), n_noises) for _ in range(n_runs)]
noise_vectors = [noise_vector / np.linalg.norm(noise_vector, axis=0)
for noise_vector in noise_vectors]
gammas = [1., 2., 4.]
def zero_mean(x):
return np.zeros_like(x)
from mini_algo import alternating_optimization
def get_values(simulation_peak, estimation_peak, held_out_index,
noise_level, noise_vector_list,
paradigms=paradigms, frame_times_run=frame_times_run,
beta=beta, new_betas=new_betas):
simulation_hrf = _gamma_difference_hrf(tr=1., oversampling=20,
time_length=hrf_length,
undershoot=16., delay=simulation_peak)
xs = np.linspace(0., hrf_length + 1, len(simulation_hrf), endpoint=False)
f_sim_hrf = interp1d(xs, simulation_hrf)
shifted_paradigms = [paradigm.copy()
for i, paradigm in enumerate(paradigms)
if i != held_out_index]
shifted_frame_times = []
offset = 0
# shift paradigms to concatenate them
for paradigm in shifted_paradigms:
paradigm_length = paradigm['onset'].max()
paradigm['onset'] += offset
shifted_frame_times.append(frame_times_run + offset)
offset += paradigm_length + time_offset
shifted_frame_times = np.concatenate(shifted_frame_times)
train_paradigm = pandas.concat(shifted_paradigms)
test_paradigm = paradigms[held_out_index]
train_noise = np.concatenate(
[noise for i, noise in enumerate(noise_vector_list)
if i != held_out_index])
scaled_train_noise = train_noise[:, np.newaxis] * noise_level
#test_noise = noise_vectors[held_out_index]
# design matrix dataframes
train_design_gen_df = make_design_matrix_hrf(shifted_frame_times,
train_paradigm, f_hrf=f_sim_hrf)
test_design_gen_df = make_design_matrix_hrf(frame_times_run,
test_paradigm, f_hrf=f_sim_hrf)
# design matrix without drifts
train_design_gen = train_design_gen_df[event_types].values
test_design_gen = test_design_gen_df[event_types].values
y_train_clean = train_design_gen.dot(beta)
y_train_norm = np.linalg.norm(y_train_clean) ** 2
y_train_noisy = y_train_clean[:, np.newaxis] + np.linalg.norm(y_train_clean) * scaled_train_noise
y_train_noisy_norm = np.linalg.norm(y_train_noisy, axis=0) ** 2
#train_signal_norm[i_sim, held_out_index, :] = y_train_noisy_norm
y_test = test_design_gen.dot(beta)
y_test_new = test_design_gen.dot(new_betas)
y_test_norm = np.linalg.norm(y_test) ** 2
y_test_new_norm = np.linalg.norm(y_test_new, axis=0) ** 2
#test_signal_norm[i_sim, held_out_index, :] = y_test_norm
#new_test_signal_norm[i_sim, held_out_index, :] = y_test_new_norm
beta_hat_gen = np.linalg.pinv(train_design_gen).dot(y_train_noisy)
train_gen_resid = np.linalg.norm(y_train_noisy -
train_design_gen.dot(beta_hat_gen), axis=0) ** 2
#train_gen_train_gen[i_sim, held_out_index, :] = train_gen_resid
y_pred_gen = test_design_gen.dot(beta_hat_gen)
test_gen_resid = np.linalg.norm(y_test[:, np.newaxis] - y_pred_gen) ** 2
#train_gen_test_gen[i_sim, held_out_index, :] = test_gen_resid
#print("Generation peak {} Estimation peak {} Fold {}".format(simulation_peak, estimation_peak, held_out_index))
estimation_hrf = _gamma_difference_hrf(tr=1., oversampling=20,
time_length=hrf_length,
undershoot=16, delay=estimation_peak)
f_est_hrf = interp1d(xs, estimation_hrf)
# design matrix dataframes
train_design_est_df = make_design_matrix_hrf(shifted_frame_times,
train_paradigm, f_hrf=f_est_hrf)
test_design_est_df = make_design_matrix_hrf(frame_times_run,
test_paradigm, f_hrf=f_est_hrf)
# design matrix without drifts
train_design_est = train_design_est_df[event_types].values
test_design_est = test_design_est_df[event_types].values
beta_hat_est = np.linalg.pinv(train_design_est).dot(y_train_noisy)
train_est_resid = np.linalg.norm(y_train_noisy -
train_design_est.dot(beta_hat_est), axis=0) ** 2
#train_gen_train_est[i_sim, i_est, held_out_index, :] = train_est_resid
y_pred_est = test_design_est.dot(beta_hat_est)
test_est_resid = np.linalg.norm(y_test[:, np.newaxis] - y_pred_est, axis=0) ** 2
#train_gen_test_est[i_sim, i_est, held_out_index, :] = test_est_resid
y_test_squashed = test_design_est.dot(np.linalg.pinv(test_design_est).dot(y_test))
test_squashed_resid = np.linalg.norm(y_test - y_test_squashed, axis=0) ** 2
#test_est_test_est[i_sim, i_est, held_out_index, :] = test_squashed_resid
# now for some crazy hrf fitting
output = alternating_optimization(
train_paradigm, y_train_noisy,
hrf_length,
frame_times=shifted_frame_times,
mean=f_est_hrf,
n_alternations=10,
sigma_squared=1,
rescale_hrf=False,
optimize_kernel=True,
optimize_sigma_squared=False)
(betas, (hrf_measurement_points, hrf_measures),
residuals,
hrfs, lls, grads, looes, thetas, sigmas_squared) = output
hrf_measurement_points = np.concatenate(hrf_measurement_points)
order = np.argsort(hrf_measurement_points)
hrf_measurement_points = hrf_measurement_points[order]
hrf_measures = hrf_measures[order]
extra_points = np.array([0., -.1, hrf_length, hrf_length + 1.])
extra_values = np.zeros_like(extra_points)
hrf_func = interp1d(np.concatenate([hrf_measurement_points, extra_points]),
np.concatenate([hrf_measures, extra_points]))
fitted_train_resid = residuals[-1]
fitted_train_design_ = make_design_matrix_hrf(frame_times_run, train_paradigm,
f_hrf=hrf_func)
fitted_test_design_ = make_design_matrix_hrf(frame_times_run, test_paradigm,
f_hrf=hrf_func)
fitted_train_design = fitted_train_design_[event_types].values
fitted_test_design = fitted_test_design_[event_types].values
ftd_sparsity = np.abs(fitted_test_design).sum(axis=0) / np.sqrt((fitted_test_design ** 2).sum(axis=0))
if (ftd_sparsity < 2.).any():
#print('Spike at sim {} est {} ho {} noise {}'.format(simulation_peak, estimation_peak, held_out_index, noise_level))
#print('Removing strongest entry')
spiking = ftd_sparsity < 2.
d = fitted_test_design[:, spiking]
location = np.abs(d).argmax(0)
d[location, np.arange(len(location))] = .5 * (d[location - 1, np.arange(len(location))] +
d[location + 1, np.arange(len(location))])
fitted_test_design[:, spiking] = d
reest_betas = np.linalg.pinv(fitted_train_design).dot(y_train_noisy)
# fitted_test_pred = fitted_test_design.dot(reest_betas)
fitted_test_pred = fitted_test_design.dot(betas)
fitted_test_resid = np.linalg.norm(y_test - fitted_test_pred) ** 2
fitted_test_squashed_pred = fitted_test_design.dot(
np.linalg.pinv(fitted_test_design).dot(y_test_new))
fitted_test_squashed_resid = np.linalg.norm(
y_test_new - fitted_test_squashed_pred, axis=0) ** 2
# now do exactly the same thing again for 0 mean ...
# I know it is redundant
output = alternating_optimization(
train_paradigm, y_train_noisy,
hrf_length,
frame_times=shifted_frame_times,
mean=zero_mean,
n_alternations=10,
sigma_squared=1,
rescale_hrf=False,
optimize_kernel=True,
optimize_sigma_squared=False)
(zm_betas, (zm_hrf_measurement_points, zm_hrf_measures),
zm_residuals,
zm_hrfs, zm_lls, zm_grads, zm_looes, zm_thetas, zm_sigmas_squared) = output
zm_hrf_measurement_points = np.concatenate(zm_hrf_measurement_points)
order = np.argsort(zm_hrf_measurement_points)
zm_hrf_measurement_points = zm_hrf_measurement_points[order]
zm_hrf_measures = zm_hrf_measures[order]
extra_points = np.array([0., -.1, hrf_length, hrf_length + 1.])
extra_values = np.zeros_like(extra_points)
zm_hrf_func = interp1d(np.concatenate([zm_hrf_measurement_points, extra_points]),
np.concatenate([zm_hrf_measures, extra_points]))
zm_fitted_train_resid = zm_residuals[-1]
zm_fitted_train_design_ = make_design_matrix_hrf(frame_times_run, train_paradigm,
f_hrf=zm_hrf_func)
zm_fitted_test_design_ = make_design_matrix_hrf(frame_times_run, test_paradigm,
f_hrf=zm_hrf_func)
zm_fitted_train_design = zm_fitted_train_design_[event_types].values
zm_fitted_test_design = zm_fitted_test_design_[event_types].values
ftd_sparsity = np.abs(zm_fitted_test_design).sum(axis=0) / np.sqrt((zm_fitted_test_design ** 2).sum(axis=0))
if (ftd_sparsity < 2.).any():
#print('Spike at sim {} est {} ho {} noise {}'.format(simulation_peak, estimation_peak, held_out_index, noise_level))
#print('Removing strongest entry')
spiking = ftd_sparsity < 2.
d = zm_fitted_test_design[:, spiking]
location = np.abs(d).argmax(0)
d[location, np.arange(len(location))] = .5 * (d[location - 1, np.arange(len(location))] +
d[location + 1, np.arange(len(location))])
zm_fitted_test_design[:, spiking] = d
zm_reest_betas = np.linalg.pinv(zm_fitted_train_design).dot(y_train_noisy)
# zm_fitted_test_pred = zm_fitted_test_design.dot(zm_reest_betas)
zm_fitted_test_pred = zm_fitted_test_design.dot(zm_betas)
zm_fitted_test_resid = np.linalg.norm(y_test - zm_fitted_test_pred) ** 2
zm_fitted_test_squashed_pred = zm_fitted_test_design.dot(
np.linalg.pinv(zm_fitted_test_design).dot(y_test_new))
zm_fitted_test_squashed_resid = np.linalg.norm(
y_test_new - zm_fitted_test_squashed_pred, axis=0) ** 2
return (train_paradigm, test_paradigm,
beta, betas, zm_betas, reest_betas, zm_reest_betas,
y_train_noisy, y_test, y_train_norm, y_train_noisy_norm, y_test_norm, y_test_new_norm,
train_gen_resid, test_gen_resid, train_est_resid, test_est_resid,
test_squashed_resid, fitted_train_resid, fitted_test_resid,
fitted_test_squashed_resid, hrf_measurement_points, hrf_measures,
zm_fitted_train_resid, zm_fitted_test_resid,
zm_fitted_test_squashed_resid, zm_hrf_measurement_points, zm_hrf_measures,
train_design_gen, test_design_gen, train_design_est, test_design_est,
fitted_train_design, fitted_test_design, zm_fitted_train_design, zm_fitted_test_design)
from sklearn.externals.joblib import Parallel, delayed, Memory
mem = Memory(cachedir=None)
mem_get_values = mem.cache(get_values)
from itertools import product
parameters = product(hrf_peak_locations, hrf_peak_locations,
range(n_runs), noise_levels, range(n_noises))
print('Starting Parallel for {} parameter settings'.format(len(hrf_peak_locations) ** 2 * n_runs * len(noise_levels) * n_noises))
results = Parallel(n_jobs=48)(delayed(mem_get_values)(
simulation_peak, estimation_peak, held_out_index, noise_level,
[nois[:, i] for nois in noise_vectors])
for simulation_peak, estimation_peak,
held_out_index, noise_level, i in parameters)
# def get_values(simulation_peak, estimation_peak, held_out_index,
# noise_level, noise_vector_list,
def reshaper(x):
return x.reshape(len(hrf_peak_locations), len(hrf_peak_locations),
n_runs, len(noise_levels), -1)
(beta, betas, zm_betas, reest_betas, zm_reest_betas,
y_train_noisy, y_test, y_train_norms, y_train_noisy_norms, y_test_norms, y_test_new_norms,
train_gen_resids, test_gen_resids, train_est_resids, test_est_resids,
test_squashed_resids,fitted_train_resids, fitted_test_resids,
fitted_test_squashed_resids,
hrf_measurement_points, hrf_measures, zm_fitted_train_resid, zm_fitted_test_resid,
zm_fitted_test_squashed_resid, zm_hrf_measurement_points, zm_hrf_measures,
train_design_gen, test_design_gen, train_design_est, test_design_est,
fitted_train_design, fitted_test_design, zm_fitted_train_design, zm_fitted_test_design
) = map(reshaper, map(np.array, list(zip(*results))[2:]))
train_paradigms, test_paradigms = list(zip(*results))[:2]
# train_gen_train_est = np.zeros(
# (len(hrf_peak_locations),
# len(hrf_peak_locations),
# n_runs,
# n_noises * len(noise_levels)))
# train_gen_train_gen = np.zeros(
# (len(hrf_peak_locations),
# n_runs,
# n_noises * len(noise_levels)))
# train_gen_test_est = np.zeros(
# (len(hrf_peak_locations),
# len(hrf_peak_locations),
# n_runs,
# n_noises * len(noise_levels)))
# train_gen_test_gen = np.zeros(
# (len(hrf_peak_locations),
# n_runs,
# n_noises * len(noise_levels)))
# test_est_test_est = np.zeros(
# (len(hrf_peak_locations),
# len(hrf_peak_locations),
# n_runs,
# n_new_betas))
# train_signal_norm = np.zeros(
# (len(hrf_peak_locations),
# n_runs,
# n_noises * len(noise_levels)))
# test_signal_norm = np.zeros(
# (len(hrf_peak_locations),
# n_runs,
# n_noises * len(noise_levels)))
# new_test_signal_norm = np.zeros(
# (len(hrf_peak_locations),
# n_runs,
# n_new_betas))
# for i_sim, simulation_peak in enumerate(hrf_peak_locations):
# simulation_hrf = _gamma_difference_hrf(tr=1., oversampling=20,
# time_length=hrf_length + 1,
# undershoot=16., delay=simulation_peak)
# xs = np.linspace(0., hrf_length + 1, len(simulation_hrf), endpoint=False)
# f_sim_hrf = interp1d(xs, simulation_hrf)
# for held_out_index in range(n_runs):
# shifted_paradigms = [paradigm.copy()
# for i, paradigm in enumerate(paradigms)
# if i != held_out_index]
# shifted_frame_times = []
# offset = 0
# # shift paradigms to concatenate them
# for paradigm in shifted_paradigms:
# paradigm_length = paradigm['onset'].max()
# paradigm['onset'] += offset
# shifted_frame_times.append(frame_times_run + offset)
# offset += paradigm_length + time_offset
# shifted_frame_times = np.concatenate(shifted_frame_times)
# train_paradigm = pandas.concat(shifted_paradigms)
# test_paradigm = paradigms[held_out_index]
# train_noise = np.concatenate(
# [noise for i, noise in enumerate(noise_vectors)
# if i != held_out_index], axis=0)
# scaled_train_noise = (train_noise[:, np.newaxis] *
# noise_levels[np.newaxis, :, np.newaxis]
# ).reshape(train_noise.shape[0], -1)
# #test_noise = noise_vectors[held_out_index]
# # design matrix dataframes
# train_design_gen_df = make_design_matrix_hrf(shifted_frame_times,
# train_paradigm, f_hrf=f_sim_hrf)
# test_design_gen_df = make_design_matrix_hrf(frame_times_run,
# test_paradigm, f_hrf=f_sim_hrf)
# # design matrix without drifts
# train_design_gen = train_design_gen_df[event_types].values
# test_design_gen = test_design_gen_df[event_types].values
# y_train_clean = train_design_gen.dot(beta)
# y_train_norm = np.linalg.norm(y_train_clean) ** 2
# y_train_noisy = y_train_clean[:, np.newaxis] + np.linalg.norm(y_train_clean) * scaled_train_noise
# y_train_noisy_norm = np.linalg.norm(y_train_noisy, axis=0) ** 2
# train_signal_norm[i_sim, held_out_index, :] = y_train_noisy_norm
# y_test = test_design_gen.dot(beta)
# y_test_new = test_design_gen.dot(new_betas)
# y_test_norm = np.linalg.norm(y_test) ** 2
# y_test_new_norm = np.linalg.norm(y_test_new, axis=0) ** 2
# test_signal_norm[i_sim, held_out_index, :] = y_test_norm
# new_test_signal_norm[i_sim, held_out_index, :] = y_test_new_norm
# beta_hat_gen = np.linalg.pinv(train_design_gen).dot(y_train_noisy)
# train_gen_resid = np.linalg.norm(y_train_noisy -
# train_design_gen.dot(beta_hat_gen), axis=0) ** 2
# train_gen_train_gen[i_sim, held_out_index, :] = train_gen_resid
# y_pred_gen = test_design_gen.dot(beta_hat_gen)
# test_gen_resid = np.linalg.norm(y_test[:, np.newaxis] - y_pred_gen) ** 2
# train_gen_test_gen[i_sim, held_out_index, :] = test_gen_resid
# for i_est, estimation_peak in enumerate(hrf_peak_locations):
# #print("Generation peak {} Estimation peak {} Fold {}".format(simulation_peak, estimation_peak, held_out_index))
# estimation_hrf = _gamma_difference_hrf(tr=1., oversampling=20,
# time_length=hrf_length + 1,
# undershoot=16, delay=estimation_peak)
# f_est_hrf = interp1d(xs, estimation_hrf)
# # design matrix dataframes
# train_design_est_df = make_design_matrix_hrf(shifted_frame_times,
# train_paradigm, f_hrf=f_est_hrf)
# test_design_est_df = make_design_matrix_hrf(frame_times_run,
# test_paradigm, f_hrf=f_est_hrf)
# # design matrix without drifts
# train_design_est = train_design_est_df[event_types].values
# test_design_est = test_design_est_df[event_types].values
# beta_hat_est = np.linalg.pinv(train_design_est).dot(y_train_noisy)
# train_est_resid = np.linalg.norm(y_train_noisy -
# train_design_est.dot(beta_hat_est), axis=0) ** 2
# train_gen_train_est[i_sim, i_est, held_out_index, :] = train_est_resid
# y_pred_est = test_design_est.dot(beta_hat_est)
# test_est_resid = np.linalg.norm(y_test[:, np.newaxis] - y_pred_est, axis=0) ** 2
# train_gen_test_est[i_sim, i_est, held_out_index, :] = test_est_resid
# y_test_squashed = test_design_est.dot(np.linalg.pinv(test_design_est).dot(y_test))
# test_squashed_resid = np.linalg.norm(y_test - y_test_squashed, axis=0) ** 2
# test_est_test_est[i_sim, i_est, held_out_index, :] = test_squashed_resid