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
|
import warnings
import numpy as np
cimport numpy as np
from scipy import sparse
from ..exceptions import ConvergenceWarning
cdef extern from *:
ctypedef char* const_char_p "const char*"
################################################################################
# Includes
cdef extern from "svm.h":
cdef struct svm_csr_node
cdef struct svm_csr_model
cdef struct svm_parameter
cdef struct svm_csr_problem
char *svm_csr_check_parameter(svm_csr_problem *, svm_parameter *)
svm_csr_model *svm_csr_train(svm_csr_problem *, svm_parameter *, int *) nogil
void svm_csr_free_and_destroy_model(svm_csr_model** model_ptr_ptr)
cdef extern from "libsvm_sparse_helper.c":
# this file contains methods for accessing libsvm 'hidden' fields
svm_csr_problem * csr_set_problem (char *, np.npy_intp *,
char *, np.npy_intp *, char *, char *, char *, int )
svm_csr_model *csr_set_model(svm_parameter *param, int nr_class,
char *SV_data, np.npy_intp *SV_indices_dims,
char *SV_indices, np.npy_intp *SV_intptr_dims,
char *SV_intptr,
char *sv_coef, char *rho, char *nSV,
char *probA, char *probB)
svm_parameter *set_parameter (int , int , int , double, double ,
double , double , double , double,
double, int, int, int, char *, char *, int,
int)
void copy_sv_coef (char *, svm_csr_model *)
void copy_support (char *, svm_csr_model *)
void copy_intercept (char *, svm_csr_model *, np.npy_intp *)
int copy_predict (char *, svm_csr_model *, np.npy_intp *, char *)
int csr_copy_predict_values (np.npy_intp *data_size, char *data, np.npy_intp *index_size,
char *index, np.npy_intp *intptr_size, char *size,
svm_csr_model *model, char *dec_values, int nr_class)
int csr_copy_predict (np.npy_intp *data_size, char *data, np.npy_intp *index_size,
char *index, np.npy_intp *intptr_size, char *size,
svm_csr_model *model, char *dec_values) nogil
int csr_copy_predict_proba (np.npy_intp *data_size, char *data, np.npy_intp *index_size,
char *index, np.npy_intp *intptr_size, char *size,
svm_csr_model *model, char *dec_values) nogil
int copy_predict_values(char *, svm_csr_model *, np.npy_intp *, char *, int)
int csr_copy_SV (char *values, np.npy_intp *n_indices,
char *indices, np.npy_intp *n_indptr, char *indptr,
svm_csr_model *model, int n_features)
np.npy_intp get_nonzero_SV ( svm_csr_model *)
void copy_nSV (char *, svm_csr_model *)
void copy_probA (char *, svm_csr_model *, np.npy_intp *)
void copy_probB (char *, svm_csr_model *, np.npy_intp *)
np.npy_intp get_l (svm_csr_model *)
np.npy_intp get_nr (svm_csr_model *)
int free_problem (svm_csr_problem *)
int free_model (svm_csr_model *)
int free_param (svm_parameter *)
int free_model_SV(svm_csr_model *model)
void set_verbosity(int)
np.import_array()
def libsvm_sparse_train ( int n_features,
np.ndarray[np.float64_t, ndim=1, mode='c'] values,
np.ndarray[np.int32_t, ndim=1, mode='c'] indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] Y,
int svm_type, int kernel_type, int degree, double gamma,
double coef0, double eps, double C,
np.ndarray[np.float64_t, ndim=1, mode='c'] class_weight,
np.ndarray[np.float64_t, ndim=1, mode='c'] sample_weight,
double nu, double cache_size, double p, int
shrinking, int probability, int max_iter,
int random_seed):
"""
Wrap svm_train from libsvm using a scipy.sparse.csr matrix
Work in progress.
Parameters
----------
n_features : number of features.
XXX: can we retrieve this from any other parameter ?
X : array-like, dtype=float, size=[N, D]
Y : array, dtype=float, size=[N]
target vector
...
Notes
-------------------
See sklearn.svm.predict for a complete list of parameters.
"""
cdef svm_parameter *param
cdef svm_csr_problem *problem
cdef svm_csr_model *model
cdef const_char_p error_msg
if len(sample_weight) == 0:
sample_weight = np.ones(Y.shape[0], dtype=np.float64)
else:
assert sample_weight.shape[0] == indptr.shape[0] - 1, \
"sample_weight and X have incompatible shapes: " + \
"sample_weight has %s samples while X has %s" % \
(sample_weight.shape[0], indptr.shape[0] - 1)
# we should never end up here with a precomputed kernel matrix,
# as this is always dense.
assert(kernel_type != 4)
# set libsvm problem
problem = csr_set_problem(values.data, indices.shape, indices.data,
indptr.shape, indptr.data, Y.data,
sample_weight.data, kernel_type)
cdef np.ndarray[np.int32_t, ndim=1, mode='c'] \
class_weight_label = np.arange(class_weight.shape[0], dtype=np.int32)
# set parameters
param = set_parameter(svm_type, kernel_type, degree, gamma, coef0,
nu, cache_size, C, eps, p, shrinking,
probability, <int> class_weight.shape[0],
class_weight_label.data, class_weight.data, max_iter,
random_seed)
# check parameters
if (param == NULL or problem == NULL):
raise MemoryError("Seems we've run out of memory")
error_msg = svm_csr_check_parameter(problem, param);
if error_msg:
free_problem(problem)
free_param(param)
raise ValueError(error_msg)
# call svm_train, this does the real work
cdef int fit_status = 0
with nogil:
model = svm_csr_train(problem, param, &fit_status)
cdef np.npy_intp SV_len = get_l(model)
cdef np.npy_intp n_class = get_nr(model)
# copy model.sv_coef
# we create a new array instead of resizing, otherwise
# it would not erase previous information
cdef np.ndarray sv_coef_data
sv_coef_data = np.empty((n_class-1)*SV_len, dtype=np.float64)
copy_sv_coef (sv_coef_data.data, model)
cdef np.ndarray[np.int32_t, ndim=1, mode='c'] support
support = np.empty(SV_len, dtype=np.int32)
copy_support(support.data, model)
# copy model.rho into the intercept
# the intercept is just model.rho but with sign changed
cdef np.ndarray intercept
intercept = np.empty(n_class*(n_class-1)/2, dtype=np.float64)
copy_intercept (intercept.data, model, intercept.shape)
# copy model.SV
# we erase any previous information in SV
# TODO: custom kernel
cdef np.npy_intp nonzero_SV
nonzero_SV = get_nonzero_SV (model)
cdef np.ndarray SV_data, SV_indices, SV_indptr
SV_data = np.empty(nonzero_SV, dtype=np.float64)
SV_indices = np.empty(nonzero_SV, dtype=np.int32)
SV_indptr = np.empty(<np.npy_intp>SV_len + 1, dtype=np.int32)
csr_copy_SV(SV_data.data, SV_indices.shape, SV_indices.data,
SV_indptr.shape, SV_indptr.data, model, n_features)
support_vectors_ = sparse.csr_matrix(
(SV_data, SV_indices, SV_indptr), (SV_len, n_features))
# copy model.nSV
# TODO: do only in classification
cdef np.ndarray n_class_SV
n_class_SV = np.empty(n_class, dtype=np.int32)
copy_nSV(n_class_SV.data, model)
# # copy probabilities
cdef np.ndarray probA, probB
if probability != 0:
if svm_type < 2: # SVC and NuSVC
probA = np.empty(n_class*(n_class-1)/2, dtype=np.float64)
probB = np.empty(n_class*(n_class-1)/2, dtype=np.float64)
copy_probB(probB.data, model, probB.shape)
else:
probA = np.empty(1, dtype=np.float64)
probB = np.empty(0, dtype=np.float64)
copy_probA(probA.data, model, probA.shape)
else:
probA = np.empty(0, dtype=np.float64)
probB = np.empty(0, dtype=np.float64)
svm_csr_free_and_destroy_model (&model)
free_problem(problem)
free_param(param)
return (support, support_vectors_, sv_coef_data, intercept, n_class_SV,
probA, probB, fit_status)
def libsvm_sparse_predict (np.ndarray[np.float64_t, ndim=1, mode='c'] T_data,
np.ndarray[np.int32_t, ndim=1, mode='c'] T_indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] T_indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] SV_data,
np.ndarray[np.int32_t, ndim=1, mode='c'] SV_indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] SV_indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] sv_coef,
np.ndarray[np.float64_t, ndim=1, mode='c']
intercept, int svm_type, int kernel_type, int
degree, double gamma, double coef0, double
eps, double C,
np.ndarray[np.float64_t, ndim=1] class_weight,
double nu, double p, int
shrinking, int probability,
np.ndarray[np.int32_t, ndim=1, mode='c'] nSV,
np.ndarray[np.float64_t, ndim=1, mode='c'] probA,
np.ndarray[np.float64_t, ndim=1, mode='c'] probB):
"""
Predict values T given a model.
For speed, all real work is done at the C level in function
copy_predict (libsvm_helper.c).
We have to reconstruct model and parameters to make sure we stay
in sync with the python object.
See sklearn.svm.predict for a complete list of parameters.
Parameters
----------
X : array-like, dtype=float
Y : array
target vector
Returns
-------
dec_values : array
predicted values.
"""
cdef np.ndarray[np.float64_t, ndim=1, mode='c'] dec_values
cdef svm_parameter *param
cdef svm_csr_model *model
cdef np.ndarray[np.int32_t, ndim=1, mode='c'] \
class_weight_label = np.arange(class_weight.shape[0], dtype=np.int32)
cdef int rv
param = set_parameter(svm_type, kernel_type, degree, gamma,
coef0, nu,
100., # cache size has no effect on predict
C, eps, p, shrinking,
probability, <int> class_weight.shape[0], class_weight_label.data,
class_weight.data, -1,
-1) # random seed has no effect on predict either
model = csr_set_model(param, <int> nSV.shape[0], SV_data.data,
SV_indices.shape, SV_indices.data,
SV_indptr.shape, SV_indptr.data,
sv_coef.data, intercept.data,
nSV.data, probA.data, probB.data)
#TODO: use check_model
dec_values = np.empty(T_indptr.shape[0]-1)
with nogil:
rv = csr_copy_predict(T_data.shape, T_data.data,
T_indices.shape, T_indices.data,
T_indptr.shape, T_indptr.data,
model, dec_values.data)
if rv < 0:
raise MemoryError("We've run out of memory")
# free model and param
free_model_SV(model)
free_model(model)
free_param(param)
return dec_values
def libsvm_sparse_predict_proba(
np.ndarray[np.float64_t, ndim=1, mode='c'] T_data,
np.ndarray[np.int32_t, ndim=1, mode='c'] T_indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] T_indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] SV_data,
np.ndarray[np.int32_t, ndim=1, mode='c'] SV_indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] SV_indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] sv_coef,
np.ndarray[np.float64_t, ndim=1, mode='c']
intercept, int svm_type, int kernel_type, int
degree, double gamma, double coef0, double
eps, double C,
np.ndarray[np.float64_t, ndim=1] class_weight,
double nu, double p, int shrinking, int probability,
np.ndarray[np.int32_t, ndim=1, mode='c'] nSV,
np.ndarray[np.float64_t, ndim=1, mode='c'] probA,
np.ndarray[np.float64_t, ndim=1, mode='c'] probB):
"""
Predict values T given a model.
"""
cdef np.ndarray[np.float64_t, ndim=2, mode='c'] dec_values
cdef svm_parameter *param
cdef svm_csr_model *model
cdef np.ndarray[np.int32_t, ndim=1, mode='c'] \
class_weight_label = np.arange(class_weight.shape[0], dtype=np.int32)
param = set_parameter(svm_type, kernel_type, degree, gamma,
coef0, nu,
100., # cache size has no effect on predict
C, eps, p, shrinking,
probability, <int> class_weight.shape[0], class_weight_label.data,
class_weight.data, -1,
-1) # random seed has no effect on predict either
model = csr_set_model(param, <int> nSV.shape[0], SV_data.data,
SV_indices.shape, SV_indices.data,
SV_indptr.shape, SV_indptr.data,
sv_coef.data, intercept.data,
nSV.data, probA.data, probB.data)
#TODO: use check_model
cdef np.npy_intp n_class = get_nr(model)
cdef int rv
dec_values = np.empty((T_indptr.shape[0]-1, n_class), dtype=np.float64)
with nogil:
rv = csr_copy_predict_proba(T_data.shape, T_data.data,
T_indices.shape, T_indices.data,
T_indptr.shape, T_indptr.data,
model, dec_values.data)
if rv < 0:
raise MemoryError("We've run out of memory")
# free model and param
free_model_SV(model)
free_model(model)
free_param(param)
return dec_values
def libsvm_sparse_decision_function(
np.ndarray[np.float64_t, ndim=1, mode='c'] T_data,
np.ndarray[np.int32_t, ndim=1, mode='c'] T_indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] T_indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] SV_data,
np.ndarray[np.int32_t, ndim=1, mode='c'] SV_indices,
np.ndarray[np.int32_t, ndim=1, mode='c'] SV_indptr,
np.ndarray[np.float64_t, ndim=1, mode='c'] sv_coef,
np.ndarray[np.float64_t, ndim=1, mode='c']
intercept, int svm_type, int kernel_type, int
degree, double gamma, double coef0, double
eps, double C,
np.ndarray[np.float64_t, ndim=1] class_weight,
double nu, double p, int shrinking, int probability,
np.ndarray[np.int32_t, ndim=1, mode='c'] nSV,
np.ndarray[np.float64_t, ndim=1, mode='c'] probA,
np.ndarray[np.float64_t, ndim=1, mode='c'] probB):
"""
Predict margin (libsvm name for this is predict_values)
We have to reconstruct model and parameters to make sure we stay
in sync with the python object.
"""
cdef np.ndarray[np.float64_t, ndim=2, mode='c'] dec_values
cdef svm_parameter *param
cdef np.npy_intp n_class
cdef svm_csr_model *model
cdef np.ndarray[np.int32_t, ndim=1, mode='c'] \
class_weight_label = np.arange(class_weight.shape[0], dtype=np.int32)
param = set_parameter(svm_type, kernel_type, degree, gamma,
coef0, nu,
100., # cache size has no effect on predict
C, eps, p, shrinking,
probability, <int> class_weight.shape[0],
class_weight_label.data, class_weight.data, -1, -1)
model = csr_set_model(param, <int> nSV.shape[0], SV_data.data,
SV_indices.shape, SV_indices.data,
SV_indptr.shape, SV_indptr.data,
sv_coef.data, intercept.data,
nSV.data, probA.data, probB.data)
if svm_type > 1:
n_class = 1
else:
n_class = get_nr(model)
n_class = n_class * (n_class - 1) / 2
dec_values = np.empty((T_indptr.shape[0] - 1, n_class), dtype=np.float64)
if csr_copy_predict_values(T_data.shape, T_data.data,
T_indices.shape, T_indices.data,
T_indptr.shape, T_indptr.data,
model, dec_values.data, n_class) < 0:
raise MemoryError("We've run out of memory")
# free model and param
free_model_SV(model)
free_model(model)
free_param(param)
return dec_values
def set_verbosity_wrap(int verbosity):
"""
Control verbosity of libsvm library
"""
set_verbosity(verbosity)
|
