-
Notifications
You must be signed in to change notification settings - Fork 12
/
plot_miyawaki_reconstruction.py
338 lines (277 loc) · 8.68 KB
/
plot_miyawaki_reconstruction.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
"""
Reconstruction of visual stimuli from Miyawaki et al. 2008
==========================================================
This example reproduces the experiment presented in :footcite:t:`Miyawaki2008`.
It reconstructs 10x10 binary images from functional MRI data. Random images
are used as training set and structured images are used for reconstruction.
The code is a bit elaborate as the example uses, as the original article,
a multiscale prediction on the images seen by the subject.
For an encoding approach for the same dataset, see also
:ref:`sphx_glr_auto_examples_02_decoding_plot_miyawaki_encoding.py`
.. include:: ../../../examples/masker_note.rst
"""
# %%
import sys
import time
# %%
# First we load the Miyawaki dataset
# ----------------------------------
from nilearn import datasets
from nilearn.plotting import show
sys.stderr.write("Fetching dataset...")
t0 = time.time()
miyawaki_dataset = datasets.fetch_miyawaki2008()
# print basic information on the dataset
print(
"First functional nifti image (4D) is located "
f"at: {miyawaki_dataset.func[0]}"
)
X_random_filenames = miyawaki_dataset.func[12:]
X_figure_filenames = miyawaki_dataset.func[:12]
y_random_filenames = miyawaki_dataset.label[12:]
y_figure_filenames = miyawaki_dataset.label[:12]
y_shape = (10, 10)
sys.stderr.write(f" Done ({time.time() - t0:.2f}s).\n")
# %%
# Then we prepare and mask the data
# ---------------------------------
import numpy as np
from nilearn.maskers import MultiNiftiMasker
sys.stderr.write("Preprocessing data...")
t0 = time.time()
# Load and mask fMRI data
masker = MultiNiftiMasker(
mask_img=miyawaki_dataset.mask, detrend=True, standardize=False, n_jobs=2
)
masker.fit()
X_train = masker.transform(X_random_filenames)
X_test = masker.transform(X_figure_filenames)
y_train = [
np.reshape(
np.loadtxt(y, dtype=int, delimiter=","), (-1,) + y_shape, order="F"
)
for y in y_random_filenames
]
y_test = [
np.reshape(
np.loadtxt(y, dtype=int, delimiter=","), (-1,) + y_shape, order="F"
)
for y in y_figure_filenames
]
X_train = np.vstack([x[2:] for x in X_train])
y_train = np.vstack([y[:-2] for y in y_train]).astype(float)
X_test = np.vstack([x[2:] for x in X_test])
y_test = np.vstack([y[:-2] for y in y_test]).astype(float)
n_features = X_train.shape[1]
def flatten(list_of_2d_array):
flattened = [array.ravel() for array in list_of_2d_array]
return flattened
# Build the design matrix for multiscale computation
# Matrix is squared, y_rows == y_cols
y_cols = y_shape[1]
# Original data
design_matrix = np.eye(100)
# Example of matrix used for multiscale (sum pixels vertically)
#
# 0.5 *
#
# 1 1 0 0 0 0 0 0 0 0
# 0 1 1 0 0 0 0 0 0 0
# 0 0 1 1 0 0 0 0 0 0
# 0 0 0 1 1 0 0 0 0 0
# 0 0 0 0 1 1 0 0 0 0
# 0 0 0 0 0 1 1 0 0 0
# 0 0 0 0 0 0 1 1 0 0
# 0 0 0 0 0 0 0 1 1 0
# 0 0 0 0 0 0 0 0 1 1
height_tf = (np.eye(y_cols) + np.eye(y_cols, k=1))[: y_cols - 1] * 0.5
width_tf = height_tf.T
yt_tall = [np.dot(height_tf, m) for m in y_train]
yt_large = [np.dot(m, width_tf) for m in y_train]
yt_big = [np.dot(height_tf, np.dot(m, width_tf)) for m in y_train]
# Add it to the training set
y_train = [
np.r_[y.ravel(), t.ravel(), l.ravel(), b.ravel()]
for y, t, l, b in zip(y_train, yt_tall, yt_large, yt_big)
]
y_test = np.asarray(flatten(y_test))
y_train = np.asarray(y_train)
# Remove rest period
X_train = X_train[y_train[:, 0] != -1]
y_train = y_train[y_train[:, 0] != -1]
X_test = X_test[y_test[:, 0] != -1]
y_test = y_test[y_test[:, 0] != -1]
sys.stderr.write(f" Done ({time.time() - t0:.2f}s).\n")
# %%
# We define our prediction function
# ---------------------------------
sys.stderr.write("Training classifiers... \r")
t0 = time.time()
from sklearn.feature_selection import SelectKBest, f_classif
from sklearn.linear_model import OrthogonalMatchingPursuit as OMP
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
# Create as many OMP as voxels to predict
clfs = []
n_clfs = y_train.shape[1]
for i in range(y_train.shape[1]):
sys.stderr.write(
f"Training classifiers {int(i + 1):03}/{int(n_clfs)}... \r"
)
clf = Pipeline(
[
("selection", SelectKBest(f_classif, k=500)),
("scl", StandardScaler()),
("clf", OMP(n_nonzero_coefs=10)),
]
)
clf.fit(X_train, y_train[:, i])
clfs.append(clf)
sys.stderr.write(
f"Training classifiers {n_clfs:03d}/{n_clfs:d}... "
f"Done ({(time.time() - t0):.2f}s).\n"
)
# %%
# Here we run the prediction: the decoding itself
# -----------------------------------------------
sys.stderr.write("Calculating scores and outputs...")
t0 = time.time()
y_pred = [clf.predict(X_test) for clf in clfs]
y_pred = np.asarray(y_pred).T
# We need to the multi scale reconstruction
def split_multi_scale(y, y_shape):
"""Split data into 4 original multi_scale images"""
yw, yh = y_shape
# Index of original image
split_index = [yw * yh]
# Index of large image
split_index.append(split_index[-1] + (yw - 1) * yh)
# Index of tall image
split_index.append(split_index[-1] + yw * (yh - 1))
# Index of big image
split_index.append(split_index[-1] + (yw - 1) * (yh - 1))
# We split according to computed indices
y_preds = np.split(y, split_index, axis=1)
# y_pred is the original image
y_pred = y_preds[0]
# y_pred_tall is the image with 1x2 patch application. We have to make
# some calculus to get it back in original shape
height_tf_i = (np.eye(y_cols) + np.eye(y_cols, k=-1))[
:, : y_cols - 1
] * 0.5
height_tf_i.flat[0] = 1
height_tf_i.flat[-1] = 1
y_pred_tall = [
np.dot(height_tf_i, np.reshape(m, (yw - 1, yh))).flatten()
for m in y_preds[1]
]
y_pred_tall = np.asarray(y_pred_tall)
# y_pred_large is the image with 2x1 patch application. We have to make
# some calculus to get it back in original shape
width_tf_i = (np.eye(y_cols) + np.eye(y_cols, k=1))[: y_cols - 1] * 0.5
width_tf_i.flat[0] = 1
width_tf_i.flat[-1] = 1
y_pred_large = [
np.dot(np.reshape(m, (yw, yh - 1)), width_tf_i).flatten()
for m in y_preds[2]
]
y_pred_large = np.asarray(y_pred_large)
# y_pred_big is the image with 2x2 patch application. We use previous
# matrices to get it back in original shape
y_pred_big = [
np.dot(np.reshape(m, (yw - 1, yh - 1)), width_tf_i) for m in y_preds[3]
]
y_pred_big = [
np.dot(height_tf_i, np.reshape(m, (yw - 1, yh))).flatten()
for m in y_pred_big
]
y_pred_big = np.asarray(y_pred_big)
return (y_pred, y_pred_tall, y_pred_large, y_pred_big)
y_pred, y_pred_tall, y_pred_large, y_pred_big = split_multi_scale(
y_pred, y_shape
)
y_pred = (
0.25 * y_pred
+ 0.25 * y_pred_tall
+ 0.25 * y_pred_large
+ 0.25 * y_pred_big
)
sys.stderr.write(f" Done ({time.time() - t0:.2f}s).\n")
# %%
# Let us quantify our prediction error
# ------------------------------------
from sklearn.metrics import (
accuracy_score,
f1_score,
precision_score,
recall_score,
)
print("Scores")
print("------")
print(
" - Accuracy (percent): %f"
% np.mean(
[accuracy_score(y_test[:, i], y_pred[:, i] > 0.5) for i in range(100)]
)
)
print(
" - Precision: %f"
% np.mean(
[precision_score(y_test[:, i], y_pred[:, i] > 0.5) for i in range(100)]
)
)
print(
" - Recall: %f"
% np.mean(
[
recall_score(y_test[:, i], y_pred[:, i] > 0.5, zero_division=0)
for i in range(100)
]
)
)
print(
" - F1-score: %f"
% np.mean([f1_score(y_test[:, i], y_pred[:, i] > 0.5) for i in range(100)])
)
# %%
# And finally, we plot six reconstructed images, to compare with
# ground truth
from pathlib import Path
from matplotlib import pyplot as plt
output_dir = Path.cwd() / "results" / "plot_miyawaki_reconstruction"
output_dir.mkdir(exist_ok=True, parents=True)
print(f"Output will be saved to: {output_dir}")
for i in range(6):
j = 10 * i
fig = plt.figure()
sp1 = plt.subplot(131)
sp1.axis("off")
plt.title("Stimulus")
sp2 = plt.subplot(132)
sp2.axis("off")
plt.title("Reconstruction")
sp3 = plt.subplot(133)
sp3.axis("off")
plt.title("Binarized")
sp1.imshow(
np.reshape(y_test[j], (10, 10)),
cmap=plt.cm.gray,
interpolation="nearest",
),
sp2.imshow(
np.reshape(y_pred[j], (10, 10)),
cmap=plt.cm.gray,
interpolation="nearest",
),
sp3.imshow(
np.reshape(y_pred[j] > 0.5, (10, 10)),
cmap=plt.cm.gray,
interpolation="nearest",
)
plt.savefig(output_dir / f"miyawaki2008_reconstruction_{int(i)}.png")
show()
# %%
# References
# ----------
#
# .. footbibliography::