/
test_res.py
198 lines (174 loc) · 6.6 KB
/
test_res.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
import tempfile
import numpy as np
from datetime import datetime
import scipy
import scipy.io
import pylab as pl
from scipy import linalg, signal, sparse
from scipy.sparse import linalg as splinalg
from nipy.modalities.fmri import hemodynamic_models as hdm
import multitask as mt
import hrf_estimation as he
DIR = tempfile.mkdtemp()
fir_length = 20
canonical = hdm.glover_hrf(1., 1., fir_length).reshape((-1, 1))
# load data --
print('Loading data')
ds = np.DataSource(DIR)
X_tmp = scipy.io.mmread(ds.open('X_train.mtx')).tocsr()
Y = np.load('Y_10000.npy')
X_session = X_tmp[:X_tmp.shape[0] / 4, :X_tmp.shape[1] / 5]
for n_session, Y_session in enumerate(np.split(Y, 5)):
k = X_session.shape[0] / 7.
X_train = X_session[:5 * k]
X_test = X_session[5 * k:]
Y_train = Y_session[:5 * k]
Y_test = Y_session[5 * k:]
n_task = 5
print('Done')
print('Detrending')
from multitask.savitzky_golay import savgol_filter
Y_train = Y_train.reshape((5, 672, -1))
Y_train = Y_train[..., :n_task] - savgol_filter(
Y_train[..., :n_task], 91, 4, axis=1)
Y_train = scipy.signal.detrend(Y_train, axis=1)
Y_test = Y_test.reshape((2, 672, -1))
Y_test = Y_test[..., :n_task] - savgol_filter(
Y_test[..., :n_task], 91, 4, axis=1)
Y_test = scipy.signal.detrend(Y_test, axis=1)
norm_train = np.sqrt((Y_train * Y_train).sum(1))
norm_test = np.sqrt((Y_test * Y_test).sum(1))
for i in range(5):
Y_train[i] /= norm_train[i]
for i in range(2):
Y_test[i] /= norm_test[i]
Y_train = Y_train.reshape((3360, n_task))
Y_test = Y_test.reshape((1344, n_task))
size_u = fir_length
size_v = X_train.shape[1] / size_u
if True:
print('Precomputing initialization point')
Iu = np.kron(np.eye(size_v, size_v), canonical) # could be faster
Q = X_train.dot(Iu)
v0 = linalg.lstsq(Q, Y_train)[0]
np.save('v0.npy', v0)
else:
v0 = np.load('v0.npy')
print('Calling rank_one')
u0 = canonical
#print('GOING OBO')
#start = datetime.now()
#u0 = canonical
#out = mt.rank_one_obo(
# X_train, Y_train, fir_length, u0=u0, v0=v0,
# verbose=False, plot=False, n_jobs=1)
#print datetime.now() - start
#u_obo, v = out
print('RANK ONE CLASSIC SETTING')
u = u0
v = v0[:, 0]
for i, y_train in enumerate(Y_train.T):
pl.figure()
y_train = y_train[:, None]
# compute optimal point
u0, v0 = u, v # the one from previous voxel
u, v = he.rank_one(
X_train, y_train, fir_length, u0=u0, v0=v0,
verbose=1, method='TNC', rtol=1e-32)
uv0 = he.khatri_rao(v, u)
min_loss = .5 * (linalg.norm(y_train - X_train.dot(uv0)) ** 2)
for solver in ('TNC', 'Newton-CG', 'L-BFGS-B', 'trust-ncg', 'CG'):
loss = []
timings = []
def callback(w):
W = w.reshape((-1, 1), order='F')
u, v, c = W[:size_u], W[size_u:size_u + size_v], W[size_u + size_v:]
uv0 = he.khatri_rao(v, u)
loss.append(.5 * (linalg.norm(y_train - X_train.dot(uv0)) **
2))
timings.append((datetime.now() - start).total_seconds())
start = datetime.now()
u0 = canonical
out = he.rank_one(
X_train, y_train, fir_length, u0=u0, v0=v0,
verbose=1, callback=callback, method=solver, rtol=1e-12)
print datetime.now() - start
u, v = out
pl.plot(timings, loss - min_loss, label=solver, lw=4)
#pl.scatter(timings, loss, marker='x', color='black', lw=.5)
pl.legend(loc='lower left')
#pl.xlim((0, .15))
#
pl.yscale('log')
#pl.ylim((1e-6, 1))
pl.axis('tight')
pl.ylabel(r'$f(x_k) - f(x^{*})$', fontsize='x-large')
pl.xlabel('Time (in seconds)', fontsize='x-large')
pl.savefig('bench_r1.png', transparent=True)
pl.show()
import ipdb; ipdb.set_trace()
residuals_rank_one = []
for i in range(n_task):
Iu = sparse.kron(sparse.eye(size_v, size_v), u[:, i][:, None])
# could be faster
Q = X_test.dot(Iu)
v0 = splinalg.lsqr(Q, Y_test[:, i])[0]
res = linalg.norm(Y_test[:, i] - Q.dot(v0))
residuals_rank_one.append(res)
print('Done %s out of %s' % (i, n_task))
print('RESIDUALS RANK ONE: %s' % np.sum(residuals_rank_one))
residuals_rank_one_obo = []
for i in range(n_task):
Iu = sparse.kron(sparse.eye(size_v, size_v), u_obo[:, i][:, None])
# could be faster
Q = X_test.dot(Iu)
v0 = splinalg.lsqr(Q, Y_test[:, i])[0]
res = linalg.norm(Y_test[:, i] - Q.dot(v0))
residuals_rank_one_obo.append(res)
print('Done %s out of %s' % (i, n_task))
print('RESIDUALS RANK ONE OBO: %s' % np.sum(residuals_rank_one_obo))
residuals_canonical = []
Iu = np.kron(np.eye(size_v, size_v), canonical)
# could be faster
Q = X_test.dot(Iu)
v_tmp = linalg.lstsq(Q, Y_test)[0]
for i in range(n_task):
res = linalg.norm(Y_test[:, i] - Q.dot(v_tmp[:, i]))
residuals_canonical.append(res)
print('Done %s out of %s' % (i, n_task))
print('RESIDUALS CANONICAL: %s' % np.sum(residuals_canonical))
pl.figure()
pl.title('GLM rank one vs GLM canonical, session %s' % n_session)
pl.scatter(residuals_rank_one, residuals_canonical)
xx = np.linspace(min(pl.ylim()[0], pl.xlim()[0]),
max(pl.ylim()[1], pl.xlim()[1]))
pl.plot(xx, xx)
pl.xlabel('MSE glm rank one')
pl.ylabel('MSE glm Canonical')
pl.axis('tight')
pl.savefig('mse_r1_vs_can_%03d.png' % n_session)
# pl.show()
pl.figure()
pl.title('GLM rank one OBO vs GLM canonical, session %s' % n_session)
pl.scatter(residuals_rank_one_obo, residuals_canonical)
xx = np.linspace(min(pl.ylim()[0], pl.xlim()[0]),
max(pl.ylim()[1], pl.xlim()[1]))
pl.plot(xx, xx)
pl.xlabel('MSE glm rank one OBO')
pl.ylabel('MSE glm Canonical')
pl.axis('equal')
pl.axis('tight')
pl.savefig('mse_obo_vs_can_%03d.png' % n_session)
# pl.show()
pl.figure()
pl.title('GLM rank one OBO vs GLM rank one, session %s' % n_session)
pl.scatter(residuals_rank_one_obo, residuals_rank_one)
xx = np.linspace(min(pl.ylim()[0], pl.xlim()[0]),
max(pl.ylim()[1], pl.xlim()[1]))
pl.plot(xx, xx)
pl.xlabel('MSE glm rank one OBO')
pl.ylabel('MSE glm rank one')
pl.axis('equal')
pl.axis('tight')
pl.savefig('mse_obo_vs_r1_%03d.png' % n_session)
# pl.show()