/
iters.py
405 lines (343 loc) · 15.4 KB
/
iters.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
# pylint: disable=invalid-name
from __future__ import print_function
from spectral import *
import numpy as np
import scipy.spatial.distance as distance
import pylab
import spectral.io.envi as envi
import random
from scipy.spatial.distance import cdist
import gc
def generate_class_colours(n):
import colorsys
HSV_tuples = [(x*1.0/n, 0.5, 0.5) for x in range(n)]
RGB_tuples = [colorsys.hsv_to_rgb(*x) for x in HSV_tuples]
return [(int(r*255), int(g*255), int(b*255)) for (r, g, b) in RGB_tuples]
def normalized(a, order=2, axis=-1):
norms = np.atleast_1d(np.linalg.norm(a, order, axis))
norms[norms == 0] = 1
return a / np.expand_dims(norms, axis)
def show_centers(centers, title):
if title is None:
title = "Centers"
pylab.figure()
# pylab.hold(1) # default and depricated
for i in range(min(centers.shape[0], 30)):
pylab.plot(centers[i])
pylab.title(title)
pylab.show()
def kmeans_cosine(image, nclusters=10, max_iterations=20, **kwargs):
from spectral.algorithms.spymath import has_nan, NaNValueError
if has_nan(image):
raise NaNValueError('Image data contains NaN values.')
#orig_image = image
# defaults for kwargs
start_clusters = None
compare = None
iterations = None
for (key, val) in list(kwargs.items()):
if key == 'start_clusters':
start_clusters = val
elif key == 'compare':
compare = val
elif key == 'frames':
if not hasattr(val, 'append'):
raise TypeError('"frames" keyword argument must have "append"'
'attribute.')
iterations = val
else:
raise NameError('Unsupported keyword argument.')
(nrows, ncols, nbands) = image.shape
N = nrows * ncols
image = normalized(image.reshape((N, nbands)))
#image = image.reshape((N, nbands))
clusters = np.zeros((N,), int)
centers = None
if start_clusters is not None:
assert (start_clusters.shape[0] == nclusters), 'There must be \
nclusters clusters in the start_centers array.'
centers = np.array(normalized(start_clusters))
else:
# print('Initializing clusters along diagonal of N-dimensional bounding box.')
# boxMin = np.amin(image, 0)
# boxMax = np.amax(image, 0)
# delta = (boxMax - boxMin) / (nclusters - 1)
centers = np.empty((nclusters, nbands), float)
# for i in range(nclusters):
# centers[i] = boxMin + i * delta
random.seed(4)
for i in range(nclusters):
centers[i] = image[random.randrange(N)]
#show_centers(centers, "Initial centers")
#raw_input("Press Enter to continue...")
centers = centers.T
#distances = np.empty((N, nclusters), float)
old_centers = np.array(centers)
clusters = np.zeros((N,), int)
old_clusters = np.copy(clusters)
#old_n_changed = 1
itnum = 1
while (itnum <= max_iterations):
try:
if itnum % 10 == 0:
print('\rIteration %d...' % itnum, end='')
# Assign all pixels
#distances[:] = np.matmul(image, centers)
clusters[:] = np.argmax(np.matmul(image, centers), 1)
# Update cluster centers
old_centers[:] = centers
for i in range(nclusters):
inds = np.nonzero(clusters == i)[0]
if len(inds) > 0:
centers[:, i] = np.mean(image[inds], 0, float)
centers[:, i] /= np.linalg.norm(centers[:, i])
if iterations is not None:
iterations.append(clusters.reshape(nrows, ncols))
if compare and compare(old_clusters, clusters):
print('done.')
break
else:
n_changed = np.sum(clusters != old_clusters)
# print(np.abs(n_changed - old_n_changed)/(n_changed + old_n_changed))
# if np.abs(float(n_changed - old_n_changed))/(n_changed + old_n_changed) > 0.5 and n_changed != old_n_changed:
# print(centers)
# viewt = imshow(orig_image, classes=clusters.reshape(nrows, ncols))
# viewt.set_display_mode('overlay')
# viewt.class_alpha = 0.85
# old_n_changed = n_changed
# # raw_input("Press Enter to continue...")
if itnum % 10 == 0:
print('%d pixels reassigned.' % (n_changed))
if n_changed == 0:
break
old_clusters[:] = clusters
old_centers[:] = centers
itnum += 1
except KeyboardInterrupt:
print("KeyboardInterrupt: Returning clusters from previous iteration.")
return (old_clusters.reshape(nrows, ncols), old_centers.T)
print('kmeans terminated with', len(set(old_clusters.ravel())), \
'clusters after', itnum - 1, 'iterations.')
return (old_clusters.reshape(nrows, ncols), centers.T)
def kmeans_cdist(image, nclusters=10, max_iterations=20, **kwargs):
from spectral.algorithms.spymath import has_nan, NaNValueError
if has_nan(image):
raise NaNValueError('Image data contains NaN values.')
#orig_image = image
# defaults for kwargs
start_clusters = None
compare = None
iterations = None
for (key, val) in list(kwargs.items()):
if key == 'start_clusters':
start_clusters = val
elif key == 'compare':
compare = val
elif key == 'frames':
if not hasattr(val, 'append'):
raise TypeError('"frames" keyword argument must have "append"'
'attribute.')
iterations = val
else:
raise NameError('Unsupported keyword argument.')
(nrows, ncols, nbands) = image.shape
N = nrows * ncols
image = image.reshape((N, nbands))
#image = image.reshape((N, nbands))
clusters = np.zeros((N,), int)
centers = None
if start_clusters is not None:
assert (start_clusters.shape[0] == nclusters), 'There must be \
nclusters clusters in the start_centers array.'
centers = np.array(start_clusters)
else:
# print('Initializing clusters along diagonal of N-dimensional bounding box.')
# boxMin = np.amin(image, 0)
# boxMax = np.amax(image, 0)
# delta = (boxMax - boxMin) / (nclusters - 1)
centers = np.empty((nclusters, nbands), float)
# for i in range(nclusters):
# centers[i] = boxMin + i * delta
random.seed(4)
for i in range(nclusters):
centers[i] = image[random.randrange(N)]
#show_centers(centers, "Initial centers")
#raw_input("Press Enter to continue...")
centers = centers.T
#distances = np.empty((N, nclusters), float)
old_centers = np.array(centers)
clusters = np.zeros((N,), int)
old_clusters = np.copy(clusters)
#old_n_changed = 1
itnum = 1
while (itnum <= max_iterations):
try:
if itnum % 10 == 0:
print('Iteration %d...' % itnum)
# Assign all pixels
#distances[:] = np.matmul(image, centers)
clusters[:] = np.argmin(cdist(image, centers.T, metric='sqeuclidean'), axis=1)
# Update cluster centers
old_centers[:] = centers
for i in range(nclusters):
inds = np.nonzero(clusters == i)[0]
if len(inds) > 0:
centers[:, i] = np.mean(image[inds], 0, float)
if iterations is not None:
iterations.append(clusters.reshape(nrows, ncols))
if compare and compare(old_clusters, clusters):
print('done.')
break
else:
n_changed = np.sum(clusters != old_clusters)
if itnum % 10 == 0:
print('%d pixels reassigned.' % (n_changed))
if n_changed == 0:
break
old_clusters[:] = clusters
old_centers[:] = centers
itnum += 1
except KeyboardInterrupt:
print("KeyboardInterrupt: Returning clusters from previous iteration.")
return (old_clusters.reshape(nrows, ncols), old_centers.T)
print('kmeans terminated with', len(set(old_clusters.ravel())), \
'clusters after', itnum - 1, 'iterations.')
return (old_clusters.reshape(nrows, ncols), centers.T)
def find_related_clusters(image, min_correlation, **kwargs):
from spectral.algorithms.spymath import has_nan, NaNValueError
if has_nan(image):
raise NaNValueError('Image data contains NaN values.')
start_centers = None
for (key, val) in list(kwargs.items()):
if key == 'start_centers':
start_centers = normalized(val)
else:
raise NameError('Unsupported keyword argument.')
(nrows, ncols, nbands) = image.shape
N = nrows * ncols
image = normalized(image.reshape((N, nbands)))
clusters = np.zeros((N,), int) - 1
MAX_CENTERS = 65536
centers = np.zeros((nbands, MAX_CENTERS))
num_centers = 0
if start_centers is not None:
centers[:, :start_centers.shape[0]] = start_centers.T
num_centers = start_centers.shape[0]
else:
centers[:, 0] = image[0]
num_centers = 1
percentage = 0.0
max_exceeded_warning_printed = False
for i in range(N):
match_index = np.argmax(np.matmul(image[i], centers[:, :num_centers]))
if np.dot(image[i], centers[:, match_index]) < min_correlation:
if num_centers < MAX_CENTERS:
clusters[i] = num_centers
centers[:, num_centers] = image[i]
num_centers += 1
else:
if not max_exceeded_warning_printed:
print('Exceeded max number of centers, pixels will be assigned to best existing match. Try with lower coefficient.')
max_exceeded_warning_printed = True
clusters[i] = match_index
else:
clusters[i] = match_index
if float(i)/N >= percentage + 0.2:
percentage = float(i)/N
print('\r%d%% completed' % int(percentage * 100), end='\r')
return (clusters.reshape(nrows, ncols), centers[:, :num_centers].T.copy())
#img = open_image('92AV3C.lan')
# gt = open_image('92AV3GT.GIS').read_band(0)
img = envi.open('f080611t01p00r07rdn_c_sc01_ort_img.hdr')
print(img)
data = img[400:1000, 200:, :]
data[data <= 0] = 1
print(data.dtype)
view = imshow(data, (29, 20, 12), title="Image")
#raw_input("Press Enter to continue...")
#print(distance.cosine(data[50,15,:], data[100, 1, :]))
nclusters = 500
# cosine distance based clustering
#(class_map, centers) = kmeans_cosine(data, nclusters=nclusters, max_iterations=300)
# L2 distance based clustering
#(class_map, centers) = kmeans_L2(data, nclusters=nclusters, max_iterations=500)
#(class_map, centers) = kmeans_cdist(data, nclusters=nclusters, max_iterations=900)
(class_map, centers) = find_related_clusters(data, 0.99)
#(class_map, centers) = find_related_clusters(data, 0.99, start_centers=centers)
# compute image of modules
data_double = data.astype(float)
modules = np.sqrt(np.einsum('ijk,ijk->ij', data_double, data_double))
scale_map = np.empty_like(modules)
image_shape = data.shape
reconstruction = np.empty(image_shape)
diffs = np.empty(image_shape)
abs_diffs = np.empty(image_shape)
# rel_diffs = np.empty(image_shape)
accumulated_iter_count = 0
for iter in [1, 3, 6, 10, 30, 60, 100]:
gc.collect()
print('>>>>>>>> starting iteration ' + str(accumulated_iter_count + 1) + " and batch of " + str(iter) + " iterations :")
accumulated_iter_count += iter
(class_map, centers) = kmeans_cosine(data, nclusters=centers.shape[0], max_iterations=iter, start_clusters=centers)
print('Centers\' shape: ', centers.shape)
class_colours = generate_class_colours(centers.shape[0])
view = imshow(classes=class_map, colors=class_colours, title="Classes after " + str(accumulated_iter_count) + " iterations")
# compute modules of centers
center_modules = np.sqrt(np.einsum('ij,ij->i', centers, centers))
# compute scale factor map
for i in range(scale_map.shape[0]):
for j in range(scale_map.shape[1]):
scale_map[i, j] = modules[i, j] / center_modules[class_map[i, j]]
# =====================================================================
# =====================================================================
# we can now save image_shape, scale_map, class_map and centers as they contain lossy compressed image
# we can make loseless compression if we calculate differences and then compress those
# diff can also tell as about how much information is lost with lossy compression
# to calculate diff, we first need to calculate reconstruction
for i in range(image_shape[0]):
for j in range(image_shape[1]):
reconstruction[i, j] = centers[class_map[i, j]] * scale_map[i, j]
diffs[:] = data - reconstruction
abs_diffs[:] = np.abs(diffs)
#rel_diffs[:] = abs_diffs / data
diffs_norm = np.linalg.norm(diffs)
print('Norm: ' + str(diffs_norm))
as1darray = abs_diffs.reshape((diffs.shape[0] * diffs.shape[1] * diffs.shape[2],))
print('Max: ' + str(np.max(as1darray)))
print('Min: ' + str(np.min(as1darray)))
print('Avg: ' + str(np.average(as1darray))
#index_of_max = np.argmax(as1darray)
#band_of_max = index_of_max % diffs.shape[2]
#index_of_max /= diffs.shape[2]
#j_of_max = index_of_max % diffs.shape[1]
#i_of_max = index_of_max / diffs.shape[1]
(i_of_max, j_of_max, band_of_max) = np.unravel_index(np.argmax(abs_diffs), abs_diffs.shape)
print('Coordinates of max: ', i_of_max, j_of_max, band_of_max)
print('Check the difference: ', data[i_of_max, j_of_max, band_of_max] - reconstruction[i_of_max, j_of_max, band_of_max])
# print(data[i_of_max, j_of_max] - reconstruction[i_of_max, j_of_max])
# pylab.figure()
# # pylab.hold(1) # default and depricated
# pylab.plot(data[i_of_max, j_of_max], label = 'original')
# pylab.plot(reconstruction[i_of_max, j_of_max], label = 'reconstruction')
# pylab.title("Original and reconstruction of pixel with greatest error")
# pylab.legend()
# pylab.show()
# (i_of_max, j_of_max, band_of_max) = np.unravel_index(np.argmax(rel_diffs), rel_diffs.shape)
# print('Coordinates of rel max: ', i_of_max, j_of_max, band_of_max)
# print('Values: ', data[i_of_max, j_of_max, band_of_max], reconstruction[i_of_max, j_of_max, band_of_max])
# pylab.figure()
# # pylab.hold(1) # default and depricated
# pylab.plot(data[i_of_max, j_of_max], label = 'original')
# pylab.plot(reconstruction[i_of_max, j_of_max], label = 'reconstruction')
# pylab.title("Original and reconstruction of pixel with greatest relative error")
# pylab.legend()
# pylab.show()
# rel_diffs_norm = np.linalg.norm(rel_diffs)
# print('Rel. Norm: ' + str(rel_diffs_norm))
# rel_as1darray = rel_diffs.reshape((rel_diffs.shape[0] * rel_diffs.shape[1] * rel_diffs.shape[2],))
# print('Rel. Max: ' + str(np.max(rel_as1darray))
# print('Rel. Min: ' + str(np.min(rel_as1darray))
# print('Rel. Avg: ' + str(np.average(rel_as1darray))
print('============================================ ' + str(accumulated_iter_count) + " iters completed ===================================")
gc.collect()
raw_input("Press Enter to continue...")