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
|
# cython: infer_types=True
# cython: cdivision=True
cimport cython
from libc.stdint cimport int32_t
from libc.string cimport memset, memcpy
from cymem.cymem cimport Pool
ctypedef float weight_t
DEF USE_BLAS = False
DEF EPS = 1e-5
IF USE_BLAS:
cimport blis.cy
cdef extern from "math.h" nogil:
weight_t exp(weight_t x)
weight_t sqrt(weight_t x)
cdef class Matrix:
cdef readonly Pool mem
cdef weight_t* data
cdef readonly int32_t nr_row
cdef readonly int32_t nr_col
cdef class Vec:
@staticmethod
cdef inline int arg_max(const weight_t* scores, const int n_classes) nogil:
if n_classes == 2:
return 0 if scores[0] > scores[1] else 1
cdef int i
cdef int best = 0
cdef weight_t mode = scores[0]
for i in range(1, n_classes):
if scores[i] > mode:
mode = scores[i]
best = i
return best
@staticmethod
cdef inline weight_t max(const weight_t* x, int32_t nr) nogil:
if nr == 0:
return 0
cdef int i
cdef weight_t mode = x[0]
for i in range(1, nr):
if x[i] > mode:
mode = x[i]
return mode
@staticmethod
cdef inline weight_t sum(const weight_t* vec, int32_t nr) nogil:
cdef int i
cdef weight_t total = 0
for i in range(nr):
total += vec[i]
return total
@staticmethod
cdef inline weight_t norm(const weight_t* vec, int32_t nr) nogil:
cdef weight_t total = 0
for i in range(nr):
total += vec[i] ** 2
return sqrt(total)
@staticmethod
cdef inline void add(weight_t* output, const weight_t* x,
weight_t inc, int32_t nr) nogil:
memcpy(output, x, sizeof(output[0]) * nr)
Vec.add_i(output, inc, nr)
@staticmethod
cdef inline void add_i(weight_t* vec, weight_t inc, int32_t nr) nogil:
cdef int i
for i in range(nr):
vec[i] += inc
@staticmethod
cdef inline void mul(weight_t* output, const weight_t* vec, weight_t scal,
int32_t nr) nogil:
memcpy(output, vec, sizeof(output[0]) * nr)
Vec.mul_i(output, scal, nr)
@staticmethod
cdef inline void mul_i(weight_t* vec, weight_t scal, int32_t nr) nogil:
cdef int i
IF USE_BLAS:
blis.cy.scalv(BLIS_NO_CONJUGATE, nr, scal, vec, 1)
ELSE:
for i in range(nr):
vec[i] *= scal
@staticmethod
cdef inline void pow(weight_t* output, const weight_t* vec, weight_t scal,
int32_t nr) nogil:
memcpy(output, vec, sizeof(output[0]) * nr)
Vec.pow_i(output, scal, nr)
@staticmethod
cdef inline void pow_i(weight_t* vec, const weight_t scal, int32_t nr) nogil:
cdef int i
for i in range(nr):
vec[i] **= scal
@staticmethod
@cython.cdivision(True)
cdef inline void div(weight_t* output, const weight_t* vec, weight_t scal,
int32_t nr) nogil:
memcpy(output, vec, sizeof(output[0]) * nr)
Vec.div_i(output, scal, nr)
@staticmethod
@cython.cdivision(True)
cdef inline void div_i(weight_t* vec, const weight_t scal, int32_t nr) nogil:
cdef int i
for i in range(nr):
vec[i] /= scal
@staticmethod
cdef inline void exp(weight_t* output, const weight_t* vec, int32_t nr) nogil:
memcpy(output, vec, sizeof(output[0]) * nr)
Vec.exp_i(output, nr)
@staticmethod
cdef inline void exp_i(weight_t* vec, int32_t nr) nogil:
cdef int i
for i in range(nr):
vec[i] = exp(vec[i])
@staticmethod
cdef inline void reciprocal_i(weight_t* vec, int32_t nr) nogil:
cdef int i
for i in range(nr):
vec[i] = 1.0 / vec[i]
@staticmethod
cdef inline weight_t mean(const weight_t* X, int32_t nr_dim) nogil:
cdef weight_t mean = 0.
for x in X[:nr_dim]:
mean += x
return mean / nr_dim
@staticmethod
cdef inline weight_t variance(const weight_t* X, int32_t nr_dim) nogil:
# See https://www.johndcook.com/blog/standard_deviation/
cdef double m = X[0]
cdef double v = 0.
for i in range(1, nr_dim):
diff = X[i]-m
m += diff / (i+1)
v += diff * (X[i] - m)
return v / nr_dim
cdef class VecVec:
@staticmethod
cdef inline void add(weight_t* output,
const weight_t* x,
const weight_t* y,
weight_t scale,
int32_t nr) nogil:
memcpy(output, x, sizeof(output[0]) * nr)
VecVec.add_i(output, y, scale, nr)
@staticmethod
cdef inline void add_i(weight_t* x,
const weight_t* y,
weight_t scale,
int32_t nr) nogil:
cdef int i
IF USE_BLAS:
blis.cy.axpyv(BLIS_NO_CONJUGATE, nr, scale, y, 1, x, 1)
ELSE:
for i in range(nr):
x[i] += y[i] * scale
@staticmethod
cdef inline void batch_add_i(weight_t* x,
const weight_t* y,
weight_t scale,
int32_t nr, int32_t nr_batch) nogil:
# For fixed x, matrix of y
cdef int i, _
for _ in range(nr_batch):
VecVec.add_i(x,
y, scale, nr)
y += nr
@staticmethod
cdef inline void add_pow(weight_t* output,
const weight_t* x, const weight_t* y, weight_t power, int32_t nr) nogil:
memcpy(output, x, sizeof(output[0]) * nr)
VecVec.add_pow_i(output, y, power, nr)
@staticmethod
cdef inline void add_pow_i(weight_t* x,
const weight_t* y, weight_t power, int32_t nr) nogil:
cdef int i
for i in range(nr):
x[i] += y[i] ** power
@staticmethod
cdef inline void mul(weight_t* output,
const weight_t* x, const weight_t* y, int32_t nr) nogil:
memcpy(output, x, sizeof(output[0]) * nr)
VecVec.mul_i(output, y, nr)
@staticmethod
cdef inline void mul_i(weight_t* x,
const weight_t* y, int32_t nr) nogil:
cdef int i
for i in range(nr):
x[i] *= y[i]
@staticmethod
cdef inline weight_t dot(
const weight_t* x, const weight_t* y, int32_t nr) nogil:
cdef int i
cdef weight_t total = 0
for i in range(nr):
total += x[i] * y[i]
return total
@staticmethod
cdef inline int arg_max_if_true(
const weight_t* scores, const int* is_valid, const int n_classes) nogil:
cdef int i
cdef int best = -1
for i in range(n_classes):
if is_valid[i] and (best == -1 or scores[i] > scores[best]):
best = i
return best
@staticmethod
cdef inline int arg_max_if_zero(
const weight_t* scores, const weight_t* costs, const int n_classes) nogil:
cdef int i
cdef int best = -1
for i in range(n_classes):
if costs[i] == 0 and (best == -1 or scores[i] > scores[best]):
best = i
return best
cdef class Mat:
@staticmethod
cdef inline void mean_row(weight_t* Ex,
const weight_t* mat, int32_t nr_row, int32_t nr_col) nogil:
memset(Ex, 0, sizeof(Ex[0]) * nr_col)
for i in range(nr_row):
VecVec.add_i(Ex, &mat[i * nr_col], 1.0, nr_col)
Vec.mul_i(Ex, 1.0 / nr_row, nr_col)
@staticmethod
cdef inline void var_row(weight_t* Vx,
const weight_t* mat, const weight_t* Ex,
int32_t nr_row, int32_t nr_col, weight_t eps) nogil:
# From https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
if nr_row == 0 or nr_col == 0:
return
cdef weight_t sum_, sum2
for i in range(nr_col):
sum_ = 0.0
sum2 = 0.0
for j in range(nr_row):
x = mat[j * nr_col + i]
sum2 += (x - Ex[i]) ** 2
sum_ += x - Ex[i]
Vx[i] = (sum2 - sum_**2 / nr_row) / nr_row
Vx[i] += eps
|
