-
Notifications
You must be signed in to change notification settings - Fork 2
/
convUtil.pyx
executable file
·167 lines (143 loc) · 6.09 KB
/
convUtil.pyx
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
# dot_cython.pyx
import numpy as np
cimport numpy as np
cimport cython
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)
cdef np.ndarray[np.double_t, ndim=2]_img2col(np.ndarray[np.double_t, ndim=4] X, int pad, int stride, int f):
cdef np.ndarray[np.double_t, ndim = 2] Z
cdef np.ndarray[np.double_t, ndim = 4] X_pad
cdef int ff, m, n_H_prev, n_W_prev, n_C_prev,i,h,w,c,row,vert_start,horiz_start,t,hh,ww,cc
ff = f * f
m, n_H_prev, n_W_prev, n_C_prev= X.shape[0],X.shape[1],X.shape[2],X.shape[3]
n_H = int((n_H_prev - f + 2 * pad) / stride) + 1
n_W = int((n_W_prev - f + 2 * pad) / stride) + 1
Z = np.zeros((m * n_H * n_W, f * f * n_C_prev), dtype=np.float64)
X_pad = np.pad(X, ((0, 0), (pad, pad), (pad, pad), (0, 0)), 'constant', constant_values=0)
row = -1
for i in range(m):
for h in range(n_H):
for w in range(n_W):
row += 1
vert_start = h * stride
horiz_start = w * stride
col=0
for hh in range(f):
for ww in range(f):
for cc in range(n_C_prev):
Z[row, col] = X_pad[i, vert_start + hh, horiz_start + ww, cc]
col += 1
# for col in range(f * f * n_C_prev):
# t = col // n_C_prev
# hh = t // f
# ww = t % f
# cc = col % n_C_prev
#
# Z[row, col] = X_pad[i, vert_start + hh, horiz_start + ww, cc]
return Z
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)
cdef np.ndarray[np.double_t, ndim=4] _col2img(np.ndarray[np.double_t, ndim=2] X, tuple output_size, int pad, int stride, int f):
cdef np.ndarray[np.double_t, ndim = 4] Z
cdef int ff, m, n_H_prev, n_W_prev, n_C_prev,i,h,w,c,row,vert_start,horiz_start,t,hh,ww,cc
ff = f * f
n_H_prev, n_W_prev, n_C_prev = output_size[0],output_size[1],output_size[2]
n_X_H_prev, n_X_W_prev = X.shape[0],X.shape[1]
n_H = int((n_H_prev - f + 2 * pad) / stride) + 1
n_W = int((n_W_prev - f + 2 * pad) / stride) + 1
n_C = n_X_W_prev // ff
m = n_X_H_prev // (n_H * n_W)
Z = np.zeros((m, n_H_prev + 2*pad, n_W_prev + 2*pad, n_C_prev), dtype=np.float64)
row = -1
for i in range(m):
for h in range(n_H):
for w in range(n_W):
row += 1
vert_start = h * stride
horiz_start = w * stride
col=0
for hh in range(f):
for ww in range(f):
for cc in range(n_C_prev):
Z[i, vert_start + hh, horiz_start + ww, cc] += X[row, col]
col += 1
# for col in range(f * f * n_C_prev):
# t = col // n_C_prev
# hh = t // f
# ww = t % f
# cc = col % n_C_prev
# Z[i, vert_start + hh, horiz_start + ww, cc] += X[row, col]
if pad > 0:
return Z[:, pad:-pad, pad:-pad, :]
else:
return Z
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)
cdef np.ndarray[np.double_t, ndim=2] _img2col_HW(np.ndarray[np.double_t, ndim=4] X,int stride, int f):
cdef np.ndarray[np.double_t, ndim=2] Z
cdef int m,n_H_prev,n_W_prev,n_C_prev,n_H,n_W
cdef i, h, w, c, horiz_start, vert_start, row, col, hh, ww
m, n_H_prev, n_W_prev, n_C_prev= X.shape[0],X.shape[1],X.shape[2],X.shape[3]
n_H = int(1 + (n_H_prev - f) / stride)
n_W = int(1 + (n_W_prev - f) / stride)
Z = np.zeros((m * n_H * n_W * n_C_prev, f * f), dtype=np.float64)
row = -1
for i in range(m):
for h in range(n_H):
for w in range(n_W):
vert_start = h * stride
horiz_start = w * stride
for c in range(n_C_prev):
row += 1
col = 0
for hh in range(f):
for ww in range(f):
Z[row, col] = X[i, vert_start + hh, horiz_start + ww, c]
col += 1
# for col in range(f * f):
# hh = col // f
# ww = col % f
# Z[row, col] = X[i, vert_start + hh, horiz_start + ww, c]
return Z
@cython.boundscheck(False)
@cython.wraparound(False)
@cython.nonecheck(False)
cdef np.ndarray[np.double_t, ndim = 4] _col2img_HW(np.ndarray[np.double_t, ndim = 2] X, tuple output_size, int stride, int f):
cdef np.ndarray[np.double_t, ndim = 4] Z
cdef int m, n_H_prev, n_W_prev, n_C_prev, n_H, n_W
cdef i, h, w, c, horiz_start, vert_start, row, col, hh, ww
n_H_prev, n_W_prev, n_C_prev = output_size[0],output_size[1],output_size[2]
n_X_H_prev, n_X_W_prev = X.shape[0],X.shape[1]
n_H = int(1 + (n_H_prev - f) / stride)
n_W = int(1 + (n_W_prev - f) / stride)
m = n_X_H_prev // (n_H * n_W * n_C_prev)
Z = np.zeros((m, n_H_prev, n_W_prev, n_C_prev), dtype=np.float64)
row = -1
for i in range(m):
for h in range(n_H):
for w in range(n_W):
vert_start = h * stride
horiz_start = w * stride
for c in range(n_C_prev):
row += 1
col=0
for hh in range(f):
for ww in range(f):
Z[i, vert_start + hh, horiz_start + ww, c] += X[row, col]
col += 1
# for col in range(f * f):
# hh = col // f
# ww = col % f
# Z[i, vert_start + hh, horiz_start + ww, c] += X[row, col]
return Z
def img2col(X, pad, stride, f):
return _img2col(X,pad,stride,f)
def col2img(X, output_size, pad, stride, f):
return _col2img(X, output_size, pad, stride, f)
def img2col_HW(X,stride, f):
return _img2col_HW(X,stride,f)
def col2img_HW(X, output_size, stride, f):
return _col2img_HW(X,output_size,stride,f)