-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcnn_raw.py
171 lines (150 loc) · 6.49 KB
/
cnn_raw.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
#%%
import h5py
import matplotlib.pyplot as plt
import numpy as np
import time
from collections import namedtuple
def col2im(im_col, height, width, channels):
assert im_col.shape == (channels, height * width)
out = np.zeros((height, width, channels))
for c in range(channels):
im_ch = im_col[c, :]
assert im_ch.shape == (height * width,)
out[:, :, c] = np.reshape(im_ch, (height, width))
return out
def im2col(im, fl_height, fl_width, stride):
in_height, in_width, in_channels = im.shape
out_height = (in_height - fl_height) // stride + 1
out_width = (in_width - fl_width) // stride + 1
out = np.zeros((fl_height * fl_width * in_channels, out_height * out_width))
for h in range(out_height):
for w in range(out_width):
h_s = h * stride
w_s = w * stride
patch = im[h_s : h_s + fl_height, w_s : w_s + fl_width, ...]
assert patch.shape == (fl_height, fl_width, in_channels)
out[:, h * out_width + w] = np.reshape(patch, -1)
return out
def conv2d_forward(x, filters, bias, stride, pad):
batches, in_height, in_width, in_channels = x.shape
fl_height, fl_width, in_channels, out_channels = filters.shape
out_width = (in_width - fl_width + 2 * pad) // stride + 1
out_height = (in_height - fl_height + 2 * pad) // stride + 1
out = np.zeros((batches, out_height, out_width, out_channels))
weights = filters.reshape(-1, filters.shape[-1]).T
assert weights.shape == (out_channels, fl_height * fl_width * in_channels)
for i in range(batches):
im = x[i]
assert im.shape == (in_height, in_width, in_channels)
im_pad = np.pad(im, ((pad, pad), (pad, pad), (0, 0)), 'constant')
assert im_pad.shape == (in_height + 2 * pad, in_width + 2 * pad, in_channels)
im_col = im2col(im_pad, fl_height, fl_width, stride)
assert im_col.shape == (
fl_height * fl_width * in_channels,
out_height * out_width,
)
z = np.dot(weights, im_col) + bias.reshape(out_channels, 1)
assert z.shape == (out_channels, out_height * out_width)
out[i, :, :, :] = col2im(z, out_height, out_width, out_channels)
cache = (x, filters, bias, stride, pad)
return out, cache
def conv2d_backward(dout, cache):
x, filters, bias, stride, pad = cache
batches, out_height, out_width, out_channels = dout.shape
fl_height, fl_width, in_channels, out_channels = filters.shape
dx = np.zeros(x.shape)
dw = np.zeros(filters.shape)
db = np.zeros(bias.shape)
x_pad = np.pad(x, ((0, 0), (pad, pad), (pad, pad), (0, 0)), 'constant')
dx_pad = np.zeros(x_pad.shape)
for i in range(batches):
for h in range(out_height):
for w in range(out_width):
for c in range(out_channels):
h_s = h * stride
w_s = w * stride
patch = x_pad[i, h_s : h_s + fl_height, w_s : w_s + fl_width, :]
assert patch.shape == (fl_height, fl_width, in_channels)
weights = filters[:, :, :, c]
assert weights.shape == (fl_height, fl_width, in_channels)
dx_pad[i, h_s : h_s + fl_height, w_s : w_s + fl_width, :] += (
weights * dout[i, h, w, c]
)
dw[:, :, :, c] += patch * dout[i, h, w, c]
db[:, :, :, c] += dout[i, h, w, c]
dx[i, :, :, :] = dx_pad[i, pad:-pad, pad:-pad, :]
assert dx.shape == x.shape
return dx, dw, db
def pool_forward(x, window, stride, mode='max'):
batches, in_height, in_width, in_channels = x.shape
out_height = (in_height - window[0]) // stride + 1
out_width = (in_width - window[1]) // stride + 1
out = np.zeros((batches, out_height, out_width, in_channels))
for i in range(batches):
for h in range(out_height):
for w in range(out_width):
for c in range(in_channels):
h_s = h * stride
w_s = w * stride
patch = x[i, h_s : h_s + window[0], w_s : w_s + window[1], c]
assert patch.shape == (window[0], window[1])
if mode == 'max':
out[i, h, w, c] = np.max(patch)
elif mode == 'avg':
out[i, h, w, c] = np.average(patch)
cache = (x, window, stride, mode)
return out, cache
def pool_backward(dout, cache):
x, window, stride, _ = cache
batches, out_height, out_width, channels = dout.shape
dx = np.zeros(x.shape)
for i in range(batches):
for h in range(out_height):
for w in range(out_width):
for c in range(channels):
h_s = h * stride
w_s = w * stride
patch = x[i, h_s : h_s + window[0], w_s : w_s + window[1], c]
assert patch.shape == (window[0], window[1])
mask = patch == np.max(patch)
dx[i, h_s : h_s + window[0], w_s : w_s + window[1], c] += (
mask * dout[i, h, w, c]
)
assert dx.shape == x.shape
return dx
def load_dataset(file_name, prefix):
model = h5py.File(file_name, 'r')
X = np.array(model[prefix + '_x'][:], dtype=np.float)
# X = X.reshape((X.shape[0], -1)).T
X = X / 255
Y = np.array(model[prefix + '_y'][:], dtype=np.int)
# Y = Y.reshape((1, Y.shape[0]))
return (X, Y)
def main():
(train_X, train_Y) = load_dataset('data/images_train.h5', 'train_set')
(test_X, test_Y) = load_dataset('data/images_test.h5', 'test_set')
print('{} X{} Y{}'.format('train', train_X.shape, train_Y.shape))
print('{} X{} Y{}'.format('test', test_X.shape, test_Y.shape))
np.random.seed(1)
A_prev = np.random.randn(10, 4, 4, 3)
W = np.random.randn(2, 2, 3, 8)
b = np.random.randn(1, 1, 1, 8)
Z, cache = conv2d_forward(A_prev, W, b, 2, 2)
print(Z.shape)
print("Z's mean =", np.mean(Z))
print("Z[3,2,1] =", Z[3, 2, 1])
np.random.seed(1)
dA, dW, db = conv2d_backward(Z, cache)
print("dA_mean =", np.mean(dA))
print("dW_mean =", np.mean(dW))
print("db_mean =", np.mean(db))
np.random.seed(1)
A_prev = np.random.randn(5, 5, 3, 2)
A, cache = pool_forward(A_prev, [2, 2], 1)
dA = np.random.randn(5, 4, 2, 2)
dA_prev = pool_backward(dA, cache)
print("mode = max")
print('mean of dA = ', np.mean(dA))
print('dA_prev[1,1] = ', dA_prev[1, 1])
print()
main()