-
Notifications
You must be signed in to change notification settings - Fork 2
/
KS_Sampling.py
323 lines (269 loc) · 11.2 KB
/
KS_Sampling.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
import numpy as np
import numpy.ma as ma
import ctypes
import os
import os.path as path
from pathos.multiprocessing import ProcessingPool as Pool
from sklearn.metrics import pairwise_distances
prog_dir = path.dirname(path.abspath(__file__))
cpp_name = "ks_cpp"
if not os.path.isfile(prog_dir + "/" + cpp_name + ".so"):
current_dir = os.path.abspath(".")
os.chdir(path.dirname(path.abspath(__file__)))
os.system("gcc -fPIC -fopenmp -O3 -shared -o" + cpp_name + ".so " + cpp_name + ".c")
os.chdir(current_dir)
ks_cpp = np.ctypeslib.load_library(cpp_name + ".so", prog_dir)
def get_dist_unsafe(X):
"""
This implementation of distance matrix is fast but unsafe.
Numerical discrepency could occur rarely.
"""
dist = X @ X.T
t = dist.diagonal().copy()
dist *= -2
dist += t[:, None]
dist += t[None, :]
return np.sqrt(dist)
def ks_sampling(X, seed=None, n_result=None, get_dist=pairwise_distances, backend="C"):
"""
ks_sampling_general(X, seed=None, n_result=None, backend="C")
Kennard-Stone Full Sampling Program
Parameters
----------
X: np.ndarray, shape: (n_sample, n_feature)
Original data, need to be generated by user.
seed: np.ndarray or list or None, shape: (n_seed, ), optional
Initial selected seed.
If set as `None`, the program will find the two samples
which have largest distance as seed.
n_result: int or None, optional
Number of samples that should be selected.
If set as `None`, `n_sample` will be used instead, i.e.
selectet all data.
get_dist: function
A function `get_dist(X)` that will read original data, and
return distance.
Default Implementation is scikit-learn Euclidean distance.
backend: str, "Python" or "C"
Specify Kennard-Stone sampling function backend in Python
language or C language.
"""
X = np.asarray(X, dtype=float)
if n_result is None:
n_result = X.shape[0]
dist = get_dist(X)
if backend == "Python":
if seed is None or len(seed) == 0:
seed = np.unravel_index(np.argmax(dist), dist.shape)
return ks_sampling_core(dist, seed, n_result)
elif backend == "C":
return ks_sampling_core_cpp(dist, seed, n_result)
else:
raise NotImplemented("Other backends are not implemented!")
def ks_sampling_mem(X, seed=None, n_result=None, backend="C", n_proc=4, n_batch=1000):
"""
ks_sampling_mem(X, seed=None, n_result=None, backend="C", n_proc=4, n_batch=1000)
Kennard-Stone Full Sampling Program
(with limited memory)
If user have enough memory space, using `ks_sampling`
instead of `ks_sampling_mem` is strongly recommended.
This program could possibly handle very large dataset.
To make memory cost as low as possible, `n_batch` could
be set to about sqrt(X.shape[0]) manually.
However, to make efficiency as the first priority,
`n_batch` could be set to as large as possible.
NOTE! Only Euclid distance is available currently!
Parameters
----------
X: np.ndarray, shape: (n_sample, n_feature)
Original data, need to be generated by user.
seed: np.ndarray or list or None, shape: (n_seed, ), optional
Initial selected seed.
If set as `None`, the program will find the two samples
which have largest distance as seed.
n_result: int or None, optional
Number of samples that should be selected.
If set as `None`, `n_sample` will be used instead, i.e.
selectet all data.
backend: str, "Python" or "C"
Specify Kennard-Stone sampling function backend in Python
language or C language.
n_proc: int, optional
Number of Python's multiprocessing processors.
NOTE! This variable only controls Python's code.
User need to use OMP_NUM_THREADS in environment to specify
C program's paralleling behavior.
n_batch: int, optional
The dimension of distance matrix evaluation in one processor.
"""
X = np.asarray(X, dtype=float)
n_sample = X.shape[0]
if n_result is None:
n_result = X.shape[0]
# Find most distant sample indexes if no seed provided
if seed is None or len(seed) == 0:
# --- This function may have numerical instability if some data feature vector is almost the same
#
# t = np.einsum("ia, ia -> i", X, X)
#
# def get_dist_slice(sliceA, sliceB):
# distAB = t[sliceA, None] - 2 * X[sliceA] @ X[sliceB].T + t[None, sliceB]
# if sliceA == sliceB:
# np.fill_diagonal(distAB, 0)
# return np.sqrt(distAB)
def get_dist_slice(sliceA, sliceB):
return pairwise_distances(X[sliceA], X[sliceB])
def get_maxloc_slice(slice_pair):
dist_slice = get_dist_slice(slice_pair[0], slice_pair[1])
max_indexes = np.unravel_index(np.argmax(dist_slice), dist_slice.shape)
return dist_slice[max_indexes], max_indexes[0] + slice_pair[0].start, max_indexes[1] + slice_pair[1].start
p = list(np.arange(0, n_sample, n_batch)) + [n_sample]
slices = [slice(p[i], p[i+1]) for i in range(len(p) - 1)]
slice_pairs = [(slices[i], slices[j]) for i in range(len(slices)) for j in range(len(slices)) if i <= j]
with Pool(n_proc) as p:
maxloc_slice_list = p.map(get_maxloc_slice, slice_pairs)
max_indexes = maxloc_slice_list[np.argmax([v[0] for v in maxloc_slice_list])][1:]
seed = max_indexes
seed = np.asarray(seed, dtype=np.uintp)
if backend == "Python":
return ks_sampling_mem_core(X, seed, n_result)
elif backend == "C":
return ks_sampling_mem_core_cpp(X, seed, n_result)
else:
raise NotImplemented("Other backends are not implemented!")
def ks_sampling_core(dist, seed, n_result):
assert(dist.shape[0] == dist.shape[1])
n_sample = dist.shape[0]
# Definition: Output Variables
result = np.zeros(n_result, dtype=int)
v_dist = np.zeros(n_result, dtype=float)
# Definition: Intermediate Variables
n_seed = len(seed)
selected = np.zeros(n_sample, dtype=bool)
min_vals = np.zeros(n_sample, dtype=float)
# --- Initialization ---
result[:n_seed] = seed # - 1
for i in seed:
selected[i] = True
if n_seed == 2:
v_dist[0] = dist[seed[0], seed[1]] # - 2
min_vals[:] = dist[seed[0]] # - 3
upper_bound = min_vals.max() # - 4
for n in seed: # - 5
np.min(np.array([min_vals, dist[n]]), axis=0, initial=upper_bound, where=np.logical_not(selected), out=min_vals)
# --- Loop argmax minimum ---
for n in range(n_seed, n_result):
sup_index = ma.array(min_vals, mask=selected).argmax() # - 1
result[n] = sup_index # - 2
v_dist[n - 1] = min_vals[sup_index] # - 3
selected[sup_index] = True # - 4 # | 5
np.min(np.array([min_vals, dist[sup_index]]), axis=0, initial=upper_bound, where=np.logical_not(selected), out=min_vals)
return result, v_dist
def ks_sampling_core_cpp(dist, seed=None, n_result=None):
"""
ks_sampling_core_cpp(dist, seed=None, n_result=None)
Kennard-Stone Sampling Program
Parameters
----------
dist: np.ndarray
shape: (n_sample, n_sample)
Distances of samples, need to be generated by user.
seed: np.ndarray or list or None, optional
shape: (n_seed, )
Initial selected seed.
If set as `None`, the C program will find the two samples
which have largest distance as seed.
n_result: int or None, optional
Number of samples that should be selected.
If set as `None`, `n_sample` will be used instead.
"""
assert(dist.shape[0] == dist.shape[1])
n_sample = dist.shape[0]
if n_result is None:
n_result = n_sample
if seed is None:
seed = np.zeros(2, dtype=np.uintp)
n_seed = 0
else:
seed = np.asarray(seed, dtype=np.uintp)
n_seed = seed.shape[0]
vdist = np.zeros(n_result, dtype=float)
result = np.zeros(n_result, dtype=np.uintp)
ks_cpp.kennard_stone(
dist.astype(float).ctypes.data_as(ctypes.c_void_p),
seed.ctypes.data_as(ctypes.c_void_p),
result.ctypes.data_as(ctypes.c_void_p),
vdist.ctypes.data_as(ctypes.c_void_p),
ctypes.c_size_t(n_sample),
ctypes.c_size_t(n_seed),
ctypes.c_size_t(n_result),
)
return result.astype(int), vdist
def ks_sampling_mem_core(X, seed, n_result):
# Definition: Output Variables
result = np.zeros(n_result, dtype=int)
v_dist = np.zeros(n_result, dtype=float)
# Definition: Intermediate Variables
n_seed = len(seed)
n_sample = X.shape[0]
# --- Initialization ---
def sliced_dist(idx):
tmp_X = X[remains] - X[idx]
return np.sqrt(np.einsum("ia, ia -> i", tmp_X, tmp_X))
remains = []
for i in range(n_sample):
if i not in seed:
remains.append(i)
result[:n_seed] = seed
if n_seed == 2:
v_dist[0] = np.linalg.norm(X[seed[0]] - X[seed[1]])
min_vals = sliced_dist(seed[0])
for n in seed:
np.min(np.array([min_vals, sliced_dist(n)]), axis=0, out=min_vals)
# --- Loop argmax minimum ---
for n in range(n_seed, n_result):
sup_index = min_vals.argmax()
result[n] = remains[sup_index]
v_dist[n - 1] = min_vals[sup_index]
remains.pop(sup_index)
min_vals[sup_index:-1] = min_vals[sup_index+1:]
min_vals = min_vals[:-1]
np.min(np.array([min_vals, sliced_dist(result[n])]), axis=0, out=min_vals)
return result, v_dist
def ks_sampling_mem_core_cpp(X, seed, n_result=None):
"""
ks_sampling_mem_core_cpp(X, seed=None, n_result=None)
Kennard-Stone Sampling Program
(with limited memory, no need of distance matrix)
Parameters
----------
X: np.ndarray, shape: (n_sample, n_feature)
Original data, need to be provided by user.
seed: np.ndarray or list or None, shape: (n_seed, )
**THIS IS NOT OPTIONAL**
Initial selected seed.
If set as `None`, the C program will find the two samples
which have largest distance as seed.
n_result: int or None, optional
Number of samples that should be selected.
If set as `None`, `n_sample` will be used instead.
"""
n_sample = X.shape[0]
n_feature = X.shape[1]
if n_result is None:
n_result = n_sample
seed = np.asarray(seed, dtype=np.uintp)
n_seed = seed.shape[0]
vdist = np.zeros(n_result, dtype=float)
result = np.zeros(n_result, dtype=np.uintp)
ks_cpp.kennard_stone_mem(
X.astype(float).ctypes.data_as(ctypes.c_void_p),
seed.ctypes.data_as(ctypes.c_void_p),
result.ctypes.data_as(ctypes.c_void_p),
vdist.ctypes.data_as(ctypes.c_void_p),
ctypes.c_size_t(n_sample),
ctypes.c_size_t(n_feature),
ctypes.c_size_t(n_seed),
ctypes.c_size_t(n_result),
)
return result.astype(int), vdist